PGMAllBth.h@ 18291

Last change on this file since 18291 was 18192, checked in by vboxsync, 16 years ago
nc
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 188.8 KB

Line
1	/* $Id: PGMAllBth.h 18192 2009-03-24 14:40:03Z 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 uErr=%x GCPhys=%RGp fPageShw=%RX64 fPageGst=%RX64\n", pvFault, (uint32_t)uErr, 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	# if defined(IN_RC)
998	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
999	PGMDynLockHCPage(pVM, (uint8_t *)pPdeDst);
1000	# endif
1001
1002	/*
1003	* Get the guest PD entry and calc big page.
1004	*/
1005	# if PGM_GST_TYPE == PGM_TYPE_32BIT
1006	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
1007	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
1008	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
1009	# else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
1010	unsigned iPDSrc;
1011	# if PGM_GST_TYPE == PGM_TYPE_PAE
1012	X86PDPE PdpeSrc;
1013	PX86PDPAE pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
1014	# else /* AMD64 */
1015	PX86PML4E pPml4eSrc;
1016	X86PDPE PdpeSrc;
1017	PX86PDPAE pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
1018	# endif
1019	GSTPDE PdeSrc;
1020
1021	if (pPDSrc)
1022	PdeSrc = pPDSrc->a[iPDSrc];
1023	else
1024	PdeSrc.u = 0;
1025	# endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
1026
1027	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1028	const bool fIsBigPage = PdeSrc.b.u1Size;
1029	# else
1030	const bool fIsBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1031	# endif
1032
1033	# ifdef IN_RING3
1034	/*
1035	* If a CR3 Sync is pending we may ignore the invalidate page operation
1036	* depending on the kind of sync and if it's a global page or not.
1037	* This doesn't make sense in GC/R0 so we'll skip it entirely there.
1038	*/
1039	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
1040	if ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3)
1041	\|\| ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3_NON_GLOBAL)
1042	&& fIsBigPage
1043	&& PdeSrc.b.u1Global
1044	)
1045	)
1046	# else
1047	if (VM_FF_ISPENDING(pVM, VM_FF_PGM_SYNC_CR3 \| VM_FF_PGM_SYNC_CR3_NON_GLOBAL) )
1048	# endif
1049	{
1050	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePageSkipped));
1051	return VINF_SUCCESS;
1052	}
1053	# endif /* IN_RING3 */
1054
1055	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1056	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1057
1058	/* Fetch the pgm pool shadow descriptor. */
1059	PPGMPOOLPAGE pShwPdpt = pgmPoolGetPageByHCPhys(pVM, pPml4eDst->u & X86_PML4E_PG_MASK);
1060	Assert(pShwPdpt);
1061
1062	/* Fetch the pgm pool shadow descriptor. */
1063	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & SHW_PDPE_PG_MASK);
1064	Assert(pShwPde);
1065
1066	Assert(pPml4eDst->n.u1Present && (pPml4eDst->u & SHW_PDPT_MASK));
1067	RTGCPHYS GCPhysPdpt = pPml4eSrc->u & X86_PML4E_PG_MASK;
1068
1069	if ( !pPml4eSrc->n.u1Present
1070	\|\| pShwPdpt->GCPhys != GCPhysPdpt)
1071	{
1072	LogFlow(("InvalidatePage: Out-of-sync PML4E (P/GCPhys) at %RGv GCPhys=%RGp vs %RGp Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1073	GCPtrPage, pShwPdpt->GCPhys, GCPhysPdpt, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1074	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1075	pPml4eDst->u = 0;
1076	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1077	PGM_INVL_GUEST_TLBS();
1078	return VINF_SUCCESS;
1079	}
1080	if ( pPml4eSrc->n.u1User != pPml4eDst->n.u1User
1081	\|\| (!pPml4eSrc->n.u1Write && pPml4eDst->n.u1Write))
1082	{
1083	/*
1084	* Mark not present so we can resync the PML4E when it's used.
1085	*/
1086	LogFlow(("InvalidatePage: Out-of-sync PML4E at %RGv Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1087	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1088	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1089	pPml4eDst->u = 0;
1090	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1091	PGM_INVL_GUEST_TLBS();
1092	}
1093	else if (!pPml4eSrc->n.u1Accessed)
1094	{
1095	/*
1096	* Mark not present so we can set the accessed bit.
1097	*/
1098	LogFlow(("InvalidatePage: Out-of-sync PML4E (A) at %RGv Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
1099	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
1100	pgmPoolFreeByPage(pPool, pShwPdpt, pVM->pgm.s.CTX_SUFF(pShwPageCR3)->idx, iPml4);
1101	pPml4eDst->u = 0;
1102	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1103	PGM_INVL_GUEST_TLBS();
1104	}
1105
1106	/* Check if the PDPT entry has changed. */
1107	Assert(pPdpeDst->n.u1Present && pPdpeDst->u & SHW_PDPT_MASK);
1108	RTGCPHYS GCPhysPd = PdpeSrc.u & GST_PDPE_PG_MASK;
1109	if ( !PdpeSrc.n.u1Present
1110	\|\| pShwPde->GCPhys != GCPhysPd)
1111	{
1112	LogFlow(("InvalidatePage: Out-of-sync PDPE (P/GCPhys) at %RGv GCPhys=%RGp vs %RGp PdpeSrc=%RX64 PdpeDst=%RX64\n",
1113	GCPtrPage, pShwPde->GCPhys, GCPhysPd, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1114	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1115	pPdpeDst->u = 0;
1116	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1117	PGM_INVL_GUEST_TLBS();
1118	return VINF_SUCCESS;
1119	}
1120	if ( PdpeSrc.lm.u1User != pPdpeDst->lm.u1User
1121	\|\| (!PdpeSrc.lm.u1Write && pPdpeDst->lm.u1Write))
1122	{
1123	/*
1124	* Mark not present so we can resync the PDPTE when it's used.
1125	*/
1126	LogFlow(("InvalidatePage: Out-of-sync PDPE at %RGv PdpeSrc=%RX64 PdpeDst=%RX64\n",
1127	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1128	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1129	pPdpeDst->u = 0;
1130	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1131	PGM_INVL_GUEST_TLBS();
1132	}
1133	else if (!PdpeSrc.lm.u1Accessed)
1134	{
1135	/*
1136	* Mark not present so we can set the accessed bit.
1137	*/
1138	LogFlow(("InvalidatePage: Out-of-sync PDPE (A) at %RGv PdpeSrc=%RX64 PdpeDst=%RX64\n",
1139	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1140	pgmPoolFreeByPage(pPool, pShwPde, pShwPdpt->idx, iPdpt);
1141	pPdpeDst->u = 0;
1142	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1143	PGM_INVL_GUEST_TLBS();
1144	}
1145	# endif /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
1146
1147
1148	/*
1149	* Deal with the Guest PDE.
1150	*/
1151	rc = VINF_SUCCESS;
1152	if (PdeSrc.n.u1Present)
1153	{
1154	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
1155	{
1156	/*
1157	* Conflict - Let SyncPT deal with it to avoid duplicate code.
1158	*/
1159	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1160	Assert(PGMGetGuestMode(pVM) <= PGMMODE_PAE);
1161	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
1162	}
1163	else if ( PdeSrc.n.u1User != PdeDst.n.u1User
1164	\|\| (!PdeSrc.n.u1Write && PdeDst.n.u1Write))
1165	{
1166	/*
1167	* Mark not present so we can resync the PDE when it's used.
1168	*/
1169	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1170	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1171	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1172	pPdeDst->u = 0;
1173	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1174	PGM_INVL_GUEST_TLBS();
1175	}
1176	else if (!PdeSrc.n.u1Accessed)
1177	{
1178	/*
1179	* Mark not present so we can set the accessed bit.
1180	*/
1181	LogFlow(("InvalidatePage: Out-of-sync (A) at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1182	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1183	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1184	pPdeDst->u = 0;
1185	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNAs));
1186	PGM_INVL_GUEST_TLBS();
1187	}
1188	else if (!fIsBigPage)
1189	{
1190	/*
1191	* 4KB - page.
1192	*/
1193	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1194	RTGCPHYS GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1195	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1196	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1197	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1198	# endif
1199	if (pShwPage->GCPhys == GCPhys)
1200	{
1201	# if 0 /* likely cause of a major performance regression; must be SyncPageWorkerTrackDeref then */
1202	const unsigned iPTEDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1203	PSHWPT pPT = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1204	if (pPT->a[iPTEDst].n.u1Present)
1205	{
1206	# ifdef PGMPOOL_WITH_USER_TRACKING
1207	/* This is very unlikely with caching/monitoring enabled. */
1208	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPT->a[iPTEDst].u & SHW_PTE_PG_MASK);
1209	# endif
1210	pPT->a[iPTEDst].u = 0;
1211	}
1212	# else /* Syncing it here isn't 100% safe and it's probably not worth spending time syncing it. */
1213	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
1214	if (RT_SUCCESS(rc))
1215	rc = VINF_SUCCESS;
1216	# endif
1217	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4KBPages));
1218	PGM_INVL_PG(GCPtrPage);
1219	}
1220	else
1221	{
1222	/*
1223	* The page table address changed.
1224	*/
1225	LogFlow(("InvalidatePage: Out-of-sync at %RGp PdeSrc=%RX64 PdeDst=%RX64 ShwGCPhys=%RGp iPDDst=%#x\n",
1226	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, iPDDst));
1227	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1228	pPdeDst->u = 0;
1229	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDOutOfSync));
1230	PGM_INVL_GUEST_TLBS();
1231	}
1232	}
1233	else
1234	{
1235	/*
1236	* 2/4MB - page.
1237	*/
1238	/* Before freeing the page, check if anything really changed. */
1239	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1240	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1241	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1242	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1243	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1244	# endif
1245	if ( pShwPage->GCPhys == GCPhys
1246	&& pShwPage->enmKind == BTH_PGMPOOLKIND_PT_FOR_BIG)
1247	{
1248	/* ASSUMES a the given bits are identical for 4M and normal PDEs */
1249	/** @todo PAT */
1250	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1251	== (PdeDst.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1252	&& ( PdeSrc.b.u1Dirty /** @todo rainy day: What about read-only 4M pages? not very common, but still... */
1253	\|\| (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)))
1254	{
1255	LogFlow(("Skipping flush for big page containing %RGv (PD=%X .u=%RX64)-> nothing has changed!\n", GCPtrPage, iPDSrc, PdeSrc.u));
1256	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPagesSkip));
1257	# if defined(IN_RC)
1258	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
1259	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
1260	# endif
1261	return VINF_SUCCESS;
1262	}
1263	}
1264
1265	/*
1266	* Ok, the page table is present and it's been changed in the guest.
1267	* If we're in host context, we'll just mark it as not present taking the lazy approach.
1268	* We could do this for some flushes in GC too, but we need an algorithm for
1269	* deciding which 4MB pages containing code likely to be executed very soon.
1270	*/
1271	LogFlow(("InvalidatePage: Out-of-sync PD at %RGp PdeSrc=%RX64 PdeDst=%RX64\n",
1272	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1273	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1274	pPdeDst->u = 0;
1275	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePage4MBPages));
1276	PGM_INVL_BIG_PG(GCPtrPage);
1277	}
1278	}
1279	else
1280	{
1281	/*
1282	* Page directory is not present, mark shadow PDE not present.
1283	*/
1284	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
1285	{
1286	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, pShwPde->idx, iPDDst);
1287	pPdeDst->u = 0;
1288	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDNPs));
1289	PGM_INVL_PG(GCPtrPage);
1290	}
1291	else
1292	{
1293	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1294	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,InvalidatePagePDMappings));
1295	}
1296	}
1297	# if defined(IN_RC)
1298	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
1299	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
1300	# endif
1301	return rc;
1302
1303	#else /* guest real and protected mode */
1304	/* There's no such thing as InvalidatePage when paging is disabled, so just ignore. */
1305	return VINF_SUCCESS;
1306	#endif
1307	}
1308
1309
1310	#ifdef PGMPOOL_WITH_USER_TRACKING
1311	/**
1312	* Update the tracking of shadowed pages.
1313	*
1314	* @param pVM The VM handle.
1315	* @param pShwPage The shadow page.
1316	* @param HCPhys The physical page we is being dereferenced.
1317	*/
1318	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys)
1319	{
1320	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1321	STAM_PROFILE_START(&pVM->pgm.s.StatTrackDeref, a);
1322	LogFlow(("SyncPageWorkerTrackDeref: Damn HCPhys=%RHp pShwPage->idx=%#x!!!\n", HCPhys, pShwPage->idx));
1323
1324	/** @todo If this turns out to be a bottle neck (very likely) two things can be done:
1325	* 1. have a medium sized HCPhys -> GCPhys TLB (hash?)
1326	* 2. write protect all shadowed pages. I.e. implement caching.
1327	*/
1328	/*
1329	* Find the guest address.
1330	*/
1331	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
1332	pRam;
1333	pRam = pRam->CTX_SUFF(pNext))
1334	{
1335	unsigned iPage = pRam->cb >> PAGE_SHIFT;
1336	while (iPage-- > 0)
1337	{
1338	if (PGM_PAGE_GET_HCPHYS(&pRam->aPages[iPage]) == HCPhys)
1339	{
1340	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1341	pgmTrackDerefGCPhys(pPool, pShwPage, &pRam->aPages[iPage]);
1342	pShwPage->cPresent--;
1343	pPool->cPresent--;
1344	STAM_PROFILE_STOP(&pVM->pgm.s.StatTrackDeref, a);
1345	return;
1346	}
1347	}
1348	}
1349
1350	for (;;)
1351	AssertReleaseMsgFailed(("HCPhys=%RHp wasn't found!\n", HCPhys));
1352	# else /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1353	pShwPage->cPresent--;
1354	pVM->pgm.s.CTX_SUFF(pPool)->cPresent--;
1355	# endif /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1356	}
1357
1358
1359	/**
1360	* Update the tracking of shadowed pages.
1361	*
1362	* @param pVM The VM handle.
1363	* @param pShwPage The shadow page.
1364	* @param u16 The top 16-bit of the pPage->HCPhys.
1365	* @param pPage Pointer to the guest page. this will be modified.
1366	* @param iPTDst The index into the shadow table.
1367	*/
1368	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackAddref)(PVM pVM, PPGMPOOLPAGE pShwPage, uint16_t u16, PPGMPAGE pPage, const unsigned iPTDst)
1369	{
1370	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1371	/*
1372	* Just deal with the simple first time here.
1373	*/
1374	if (!u16)
1375	{
1376	STAM_COUNTER_INC(&pVM->pgm.s.StatTrackVirgin);
1377	u16 = PGMPOOL_TD_MAKE(1, pShwPage->idx);
1378	}
1379	else
1380	u16 = pgmPoolTrackPhysExtAddref(pVM, u16, pShwPage->idx);
1381
1382	/* write back */
1383	Log2(("SyncPageWorkerTrackAddRef: u16=%#x->%#x iPTDst=%#x\n", u16, PGM_PAGE_GET_TRACKING(pPage), iPTDst));
1384	PGM_PAGE_SET_TRACKING(pPage, u16);
1385
1386	# endif /* PGMPOOL_WITH_GCPHYS_TRACKING */
1387
1388	/* update statistics. */
1389	pVM->pgm.s.CTX_SUFF(pPool)->cPresent++;
1390	pShwPage->cPresent++;
1391	if (pShwPage->iFirstPresent > iPTDst)
1392	pShwPage->iFirstPresent = iPTDst;
1393	}
1394	#endif /* PGMPOOL_WITH_USER_TRACKING */
1395
1396
1397	/**
1398	* Creates a 4K shadow page for a guest page.
1399	*
1400	* For 4M pages the caller must convert the PDE4M to a PTE, this includes adjusting the
1401	* physical address. The PdeSrc argument only the flags are used. No page structured
1402	* will be mapped in this function.
1403	*
1404	* @param pVM VM handle.
1405	* @param pPteDst Destination page table entry.
1406	* @param PdeSrc Source page directory entry (i.e. Guest OS page directory entry).
1407	* Can safely assume that only the flags are being used.
1408	* @param PteSrc Source page table entry (i.e. Guest OS page table entry).
1409	* @param pShwPage Pointer to the shadow page.
1410	* @param iPTDst The index into the shadow table.
1411	*
1412	* @remark Not used for 2/4MB pages!
1413	*/
1414	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorker)(PVM pVM, PSHWPTE pPteDst, GSTPDE PdeSrc, GSTPTE PteSrc, PPGMPOOLPAGE pShwPage, unsigned iPTDst)
1415	{
1416	if (PteSrc.n.u1Present)
1417	{
1418	/*
1419	* Find the ram range.
1420	*/
1421	PPGMPAGE pPage;
1422	int rc = pgmPhysGetPageEx(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK, &pPage);
1423	if (RT_SUCCESS(rc))
1424	{
1425	#ifdef VBOX_WITH_NEW_PHYS_CODE
1426	# ifndef VBOX_WITH_NEW_LAZY_PAGE_ALLOC
1427	/* Try make the page writable if necessary. */
1428	if ( PteSrc.n.u1Write
1429	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
1430	&& PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1431	{
1432	rc = pgmPhysPageMakeWritableUnlocked(pVM, pPage, PteSrc.u & GST_PTE_PG_MASK);
1433	AssertRC(rc);
1434	}
1435	# endif
1436	#endif
1437
1438	/** @todo investiage PWT, PCD and PAT. */
1439	/*
1440	* Make page table entry.
1441	*/
1442	SHWPTE PteDst;
1443	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1444	{
1445	/** @todo r=bird: Are we actually handling dirty and access bits for pages with access handlers correctly? No. */
1446	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1447	{
1448	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1449	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage);
1450	PteDst.n.u1Present = 1;
1451	PteDst.n.u1Execute = 1;
1452	PteDst.n.u1IgnorePAT = 1;
1453	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1454	/* PteDst.n.u1Write = 0 && PteDst.n.u1Size = 0 */
1455	#else
1456	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))
1457	\| PGM_PAGE_GET_HCPHYS(pPage);
1458	#endif
1459	}
1460	else
1461	{
1462	LogFlow(("SyncPageWorker: monitored page (%RHp) -> mark not present\n", PGM_PAGE_GET_HCPHYS(pPage)));
1463	PteDst.u = 0;
1464	}
1465	/** @todo count these two kinds. */
1466	}
1467	else
1468	{
1469	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1470	/*
1471	* If the page or page directory entry is not marked accessed,
1472	* we mark the page not present.
1473	*/
1474	if (!PteSrc.n.u1Accessed \|\| !PdeSrc.n.u1Accessed)
1475	{
1476	LogFlow(("SyncPageWorker: page and or page directory not accessed -> mark not present\n"));
1477	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,AccessedPage));
1478	PteDst.u = 0;
1479	}
1480	else
1481	/*
1482	* If the page is not flagged as dirty and is writable, then make it read-only, so we can set the dirty bit
1483	* when the page is modified.
1484	*/
1485	if (!PteSrc.n.u1Dirty && (PdeSrc.n.u1Write & PteSrc.n.u1Write))
1486	{
1487	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPage));
1488	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))
1489	\| PGM_PAGE_GET_HCPHYS(pPage)
1490	\| PGM_PTFLAGS_TRACK_DIRTY;
1491	}
1492	else
1493	#endif
1494	{
1495	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageSkipped));
1496	#if PGM_SHW_TYPE == PGM_TYPE_EPT
1497	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage);
1498	PteDst.n.u1Present = 1;
1499	PteDst.n.u1Write = 1;
1500	PteDst.n.u1Execute = 1;
1501	PteDst.n.u1IgnorePAT = 1;
1502	PteDst.n.u3EMT = VMX_EPT_MEMTYPE_WB;
1503	/* PteDst.n.u1Size = 0 */
1504	#else
1505	PteDst.u = (PteSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1506	\| PGM_PAGE_GET_HCPHYS(pPage);
1507	#endif
1508	}
1509	}
1510
1511	#ifdef VBOX_WITH_NEW_PHYS_CODE
1512	/*
1513	* Make sure only allocated pages are mapped writable.
1514	*/
1515	if ( PteDst.n.u1Write
1516	&& PteDst.n.u1Present
1517	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED)
1518	{
1519	PteDst.n.u1Write = 0; /** @todo this isn't quite working yet. */
1520	Log3(("SyncPageWorker: write-protecting %RGp pPage=%R[pgmpage]at iPTDst=%d\n", (RTGCPHYS)(PteSrc.u & X86_PTE_PAE_PG_MASK), pPage, iPTDst));
1521	}
1522	#endif
1523
1524	#ifdef PGMPOOL_WITH_USER_TRACKING
1525	/*
1526	* Keep user track up to date.
1527	*/
1528	if (PteDst.n.u1Present)
1529	{
1530	if (!pPteDst->n.u1Present)
1531	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, PGM_PAGE_GET_TRACKING(pPage), pPage, iPTDst);
1532	else if ((pPteDst->u & SHW_PTE_PG_MASK) != (PteDst.u & SHW_PTE_PG_MASK))
1533	{
1534	Log2(("SyncPageWorker: deref! *pPteDst=%RX64 PteDst=%RX64\n", (uint64_t)pPteDst->u, (uint64_t)PteDst.u));
1535	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1536	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, PGM_PAGE_GET_TRACKING(pPage), pPage, iPTDst);
1537	}
1538	}
1539	else if (pPteDst->n.u1Present)
1540	{
1541	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1542	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1543	}
1544	#endif /* PGMPOOL_WITH_USER_TRACKING */
1545
1546	/*
1547	* Update statistics and commit the entry.
1548	*/
1549	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
1550	if (!PteSrc.n.u1Global)
1551	pShwPage->fSeenNonGlobal = true;
1552	#endif
1553	*pPteDst = PteDst;
1554	}
1555	/* else MMIO or invalid page, we must handle them manually in the #PF handler. */
1556	/** @todo count these. */
1557	}
1558	else
1559	{
1560	/*
1561	* Page not-present.
1562	*/
1563	LogFlow(("SyncPageWorker: page not present in Pte\n"));
1564	#ifdef PGMPOOL_WITH_USER_TRACKING
1565	/* Keep user track up to date. */
1566	if (pPteDst->n.u1Present)
1567	{
1568	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1569	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1570	}
1571	#endif /* PGMPOOL_WITH_USER_TRACKING */
1572	pPteDst->u = 0;
1573	/** @todo count these. */
1574	}
1575	}
1576
1577
1578	/**
1579	* Syncs a guest OS page.
1580	*
1581	* There are no conflicts at this point, neither is there any need for
1582	* page table allocations.
1583	*
1584	* @returns VBox status code.
1585	* @returns VINF_PGM_SYNCPAGE_MODIFIED_PDE if it modifies the PDE in any way.
1586	* @param pVM VM handle.
1587	* @param PdeSrc Page directory entry of the guest.
1588	* @param GCPtrPage Guest context page address.
1589	* @param cPages Number of pages to sync (PGM_SYNC_N_PAGES) (default=1).
1590	* @param uErr Fault error (X86_TRAP_PF_*).
1591	*/
1592	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCPTR GCPtrPage, unsigned cPages, unsigned uErr)
1593	{
1594	LogFlow(("SyncPage: GCPtrPage=%RGv cPages=%u uErr=%#x\n", GCPtrPage, cPages, uErr));
1595
1596	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
1597	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
1598	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
1599	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1600	&& PGM_SHW_TYPE != PGM_TYPE_EPT
1601
1602	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1603	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1604	# endif
1605
1606	/*
1607	* Assert preconditions.
1608	*/
1609	Assert(PdeSrc.n.u1Present);
1610	Assert(cPages);
1611	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPagePD[(GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK]);
1612
1613	/*
1614	* Get the shadow PDE, find the shadow page table in the pool.
1615	*/
1616	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1617	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1618	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
1619
1620	/* Fetch the pgm pool shadow descriptor. */
1621	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
1622	Assert(pShwPde);
1623
1624	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1625	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1626	PPGMPOOLPAGE pShwPde;
1627	PX86PDPAE pPDDst;
1628
1629	/* Fetch the pgm pool shadow descriptor. */
1630	int rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
1631	AssertRCSuccessReturn(rc, rc);
1632	Assert(pShwPde);
1633
1634	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
1635	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
1636
1637	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1638	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
1639	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
1640	PX86PDPAE pPDDst;
1641	PX86PDPT pPdptDst;
1642
1643	int rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
1644	AssertRCSuccessReturn(rc, rc);
1645	Assert(pPDDst && pPdptDst);
1646	PX86PDEPAE pPdeDst = &pPDDst->a[iPDDst];
1647	# endif
1648
1649	SHWPDE PdeDst = *pPdeDst;
1650	AssertMsg(PdeDst.n.u1Present, ("%p=%llx\n", pPdeDst, (uint64_t)PdeDst.u));
1651	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1652
1653	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1654	/* Fetch the pgm pool shadow descriptor. */
1655	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
1656	Assert(pShwPde);
1657	# endif
1658
1659	# if defined(IN_RC)
1660	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
1661	PGMDynLockHCPage(pVM, (uint8_t *)pPdeDst);
1662	# endif
1663
1664	/*
1665	* Check that the page is present and that the shadow PDE isn't out of sync.
1666	*/
1667	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1668	const bool fBigPage = PdeSrc.b.u1Size;
1669	# else
1670	const bool fBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1671	# endif
1672	RTGCPHYS GCPhys;
1673	if (!fBigPage)
1674	{
1675	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1676	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1677	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1678	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1679	# endif
1680	}
1681	else
1682	{
1683	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
1684	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1685	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1686	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1687	# endif
1688	}
1689	if ( pShwPage->GCPhys == GCPhys
1690	&& PdeSrc.n.u1Present
1691	&& (PdeSrc.n.u1User == PdeDst.n.u1User)
1692	&& (PdeSrc.n.u1Write == PdeDst.n.u1Write \|\| !PdeDst.n.u1Write)
1693	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
1694	&& (!fNoExecuteBitValid \|\| PdeSrc.n.u1NoExecute == PdeDst.n.u1NoExecute)
1695	# endif
1696	)
1697	{
1698	/*
1699	* Check that the PDE is marked accessed already.
1700	* Since we set the accessed bit before getting here on a #PF, this
1701	* check is only meant for dealing with non-#PF'ing paths.
1702	*/
1703	if (PdeSrc.n.u1Accessed)
1704	{
1705	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1706	if (!fBigPage)
1707	{
1708	/*
1709	* 4KB Page - Map the guest page table.
1710	*/
1711	PGSTPT pPTSrc;
1712	int rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
1713	if (RT_SUCCESS(rc))
1714	{
1715	# ifdef PGM_SYNC_N_PAGES
1716	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1717	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1718	{
1719	/*
1720	* This code path is currently only taken when the caller is PGMTrap0eHandler
1721	* for non-present pages!
1722	*
1723	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1724	* deal with locality.
1725	*/
1726	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1727	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1728	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1729	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
1730	# else
1731	const unsigned offPTSrc = 0;
1732	# endif
1733	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1734	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1735	iPTDst = 0;
1736	else
1737	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1738	for (; iPTDst < iPTDstEnd; iPTDst++)
1739	{
1740	if (!pPTDst->a[iPTDst].n.u1Present)
1741	{
1742	GSTPTE PteSrc = pPTSrc->a[offPTSrc + iPTDst];
1743	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(GST_PT_MASK << GST_PT_SHIFT)) \| ((offPTSrc + iPTDst) << PAGE_SHIFT);
1744	NOREF(GCPtrCurPage);
1745	#ifndef IN_RING0
1746	/*
1747	* Assuming kernel code will be marked as supervisor - and not as user level
1748	* and executed using a conforming code selector - And marked as readonly.
1749	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
1750	*/
1751	PPGMPAGE pPage;
1752	if ( ((PdeSrc.u & PteSrc.u) & (X86_PTE_RW \| X86_PTE_US))
1753	\|\| iPTDst == ((GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK) /* always sync GCPtrPage */
1754	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)GCPtrCurPage)
1755	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
1756	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1757	)
1758	#endif /* else: CSAM not active */
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} PteDst=%08llx%s\n",
1761	GCPtrCurPage, PteSrc.n.u1Present,
1762	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1763	PteSrc.n.u1User & PdeSrc.n.u1User,
1764	(uint64_t)PteSrc.u,
1765	(uint64_t)pPTDst->a[iPTDst].u,
1766	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1767	}
1768	}
1769	}
1770	else
1771	# endif /* PGM_SYNC_N_PAGES */
1772	{
1773	const unsigned iPTSrc = (GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK;
1774	GSTPTE PteSrc = pPTSrc->a[iPTSrc];
1775	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1776	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1777	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}%s\n",
1778	GCPtrPage, PteSrc.n.u1Present,
1779	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1780	PteSrc.n.u1User & PdeSrc.n.u1User,
1781	(uint64_t)PteSrc.u,
1782	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1783	}
1784	}
1785	else /* MMIO or invalid page: emulated in #PF handler. */
1786	{
1787	LogFlow(("PGM_GCPHYS_2_PTR %RGp failed with %Rrc\n", GCPhys, rc));
1788	Assert(!pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK].n.u1Present);
1789	}
1790	}
1791	else
1792	{
1793	/*
1794	* 4/2MB page - lazy syncing shadow 4K pages.
1795	* (There are many causes of getting here, it's no longer only CSAM.)
1796	*/
1797	/* Calculate the GC physical address of this 4KB shadow page. */
1798	RTGCPHYS GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc) \| (GCPtrPage & GST_BIG_PAGE_OFFSET_MASK);
1799	/* Find ram range. */
1800	PPGMPAGE pPage;
1801	int rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
1802	if (RT_SUCCESS(rc))
1803	{
1804	# ifdef VBOX_WITH_NEW_PHYS_CODE
1805	# ifndef VBOX_WITH_NEW_LAZY_PAGE_ALLOC
1806	/* Try make the page writable if necessary. */
1807	if ( PdeSrc.n.u1Write
1808	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
1809	&& PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1810	{
1811	rc = pgmPhysPageMakeWritableUnlocked(pVM, pPage, GCPhys);
1812	AssertRC(rc);
1813	}
1814	# endif
1815	# endif
1816
1817	/*
1818	* Make shadow PTE entry.
1819	*/
1820	SHWPTE PteDst;
1821	PteDst.u = (PdeSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1822	\| PGM_PAGE_GET_HCPHYS(pPage);
1823	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1824	{
1825	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1826	PteDst.n.u1Write = 0;
1827	else
1828	PteDst.u = 0;
1829	}
1830	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1831	# ifdef PGMPOOL_WITH_USER_TRACKING
1832	if (PteDst.n.u1Present && !pPTDst->a[iPTDst].n.u1Present)
1833	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, PGM_PAGE_GET_TRACKING(pPage), pPage, iPTDst);
1834	# endif
1835	# ifdef VBOX_WITH_NEW_PHYS_CODE
1836	/* Make sure only allocated pages are mapped writable. */
1837	if ( PteDst.n.u1Write
1838	&& PteDst.n.u1Present
1839	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED)
1840	{
1841	PteDst.n.u1Write = 0; /** @todo this isn't quite working yet... */
1842	Log3(("SyncPage: write-protecting %RGp pPage=%R[pgmpage] at %RGv\n", GCPhys, pPage, GCPtrPage));
1843	}
1844	# endif
1845
1846	pPTDst->a[iPTDst] = PteDst;
1847
1848
1849	/*
1850	* If the page is not flagged as dirty and is writable, then make it read-only
1851	* at PD level, so we can set the dirty bit when the page is modified.
1852	*
1853	* ASSUMES that page access handlers are implemented on page table entry level.
1854	* Thus we will first catch the dirty access and set PDE.D and restart. If
1855	* there is an access handler, we'll trap again and let it work on the problem.
1856	*/
1857	/** @todo r=bird: figure out why we need this here, SyncPT should've taken care of this already.
1858	* As for invlpg, it simply frees the whole shadow PT.
1859	* ...It's possibly because the guest clears it and the guest doesn't really tell us... */
1860	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
1861	{
1862	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
1863	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
1864	PdeDst.n.u1Write = 0;
1865	}
1866	else
1867	{
1868	PdeDst.au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
1869	PdeDst.n.u1Write = PdeSrc.n.u1Write;
1870	}
1871	*pPdeDst = PdeDst;
1872	Log2(("SyncPage: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} GCPhys=%RGp%s\n",
1873	GCPtrPage, PdeSrc.n.u1Present, PdeSrc.n.u1Write, PdeSrc.n.u1User, (uint64_t)PdeSrc.u, GCPhys,
1874	PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1875	}
1876	else
1877	LogFlow(("PGM_GCPHYS_2_PTR %RGp (big) failed with %Rrc\n", GCPhys, rc));
1878	}
1879	# if defined(IN_RC)
1880	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
1881	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
1882	# endif
1883	return VINF_SUCCESS;
1884	}
1885	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDNAs));
1886	}
1887	else
1888	{
1889	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPagePDOutOfSync));
1890	Log2(("SyncPage: Out-Of-Sync PDE at %RGp PdeSrc=%RX64 PdeDst=%RX64 (GCPhys %RGp vs %RGp)\n",
1891	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, GCPhys));
1892	}
1893
1894	/*
1895	* Mark the PDE not present. Restart the instruction and let #PF call SyncPT.
1896	* Yea, I'm lazy.
1897	*/
1898	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1899	pgmPoolFreeByPage(pPool, pShwPage, pShwPde->idx, iPDDst);
1900
1901	pPdeDst->u = 0;
1902
1903	# if defined(IN_RC)
1904	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
1905	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
1906	# endif
1907	PGM_INVL_GUEST_TLBS();
1908	return VINF_PGM_SYNCPAGE_MODIFIED_PDE;
1909
1910	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
1911	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
1912	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
1913	&& !defined(IN_RC)
1914
1915	# ifdef PGM_SYNC_N_PAGES
1916	/*
1917	* Get the shadow PDE, find the shadow page table in the pool.
1918	*/
1919	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1920	X86PDE PdeDst = pgmShwGet32BitPDE(&pVM->pgm.s, GCPtrPage);
1921
1922	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
1923	X86PDEPAE PdeDst = pgmShwGetPaePDE(&pVM->pgm.s, GCPtrPage);
1924
1925	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
1926	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1927	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64; NOREF(iPdpt);
1928	PX86PDPAE pPDDst;
1929	X86PDEPAE PdeDst;
1930	PX86PDPT pPdptDst;
1931
1932	int rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
1933	AssertRCSuccessReturn(rc, rc);
1934	Assert(pPDDst && pPdptDst);
1935	PdeDst = pPDDst->a[iPDDst];
1936	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
1937	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
1938	PEPTPD pPDDst;
1939	EPTPDE PdeDst;
1940
1941	int rc = pgmShwGetEPTPDPtr(pVM, GCPtrPage, NULL, &pPDDst);
1942	if (rc != VINF_SUCCESS)
1943	{
1944	AssertRC(rc);
1945	return rc;
1946	}
1947	Assert(pPDDst);
1948	PdeDst = pPDDst->a[iPDDst];
1949	# endif
1950	AssertMsg(PdeDst.n.u1Present, ("%#llx\n", (uint64_t)PdeDst.u));
1951	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1952	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1953
1954	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1955	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1956	{
1957	/*
1958	* This code path is currently only taken when the caller is PGMTrap0eHandler
1959	* for non-present pages!
1960	*
1961	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1962	* deal with locality.
1963	*/
1964	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1965	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
1966	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1967	iPTDst = 0;
1968	else
1969	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1970	for (; iPTDst < iPTDstEnd; iPTDst++)
1971	{
1972	if (!pPTDst->a[iPTDst].n.u1Present)
1973	{
1974	GSTPTE PteSrc;
1975
1976	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
1977
1978	/* Fake the page table entry */
1979	PteSrc.u = GCPtrCurPage;
1980	PteSrc.n.u1Present = 1;
1981	PteSrc.n.u1Dirty = 1;
1982	PteSrc.n.u1Accessed = 1;
1983	PteSrc.n.u1Write = 1;
1984	PteSrc.n.u1User = 1;
1985
1986	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1987
1988	Log2(("SyncPage: 4K+ %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1989	GCPtrCurPage, PteSrc.n.u1Present,
1990	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1991	PteSrc.n.u1User & PdeSrc.n.u1User,
1992	(uint64_t)PteSrc.u,
1993	(uint64_t)pPTDst->a[iPTDst].u,
1994	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1995	}
1996	else
1997	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));
1998	}
1999	}
2000	else
2001	# endif /* PGM_SYNC_N_PAGES */
2002	{
2003	GSTPTE PteSrc;
2004	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
2005	RTGCPTR GCPtrCurPage = (GCPtrPage & ~(RTGCPTR)(SHW_PT_MASK << SHW_PT_SHIFT)) \| (iPTDst << PAGE_SHIFT);
2006
2007	/* Fake the page table entry */
2008	PteSrc.u = GCPtrCurPage;
2009	PteSrc.n.u1Present = 1;
2010	PteSrc.n.u1Dirty = 1;
2011	PteSrc.n.u1Accessed = 1;
2012	PteSrc.n.u1Write = 1;
2013	PteSrc.n.u1User = 1;
2014	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
2015
2016	Log2(("SyncPage: 4K %RGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}PteDst=%08llx%s\n",
2017	GCPtrPage, PteSrc.n.u1Present,
2018	PteSrc.n.u1Write & PdeSrc.n.u1Write,
2019	PteSrc.n.u1User & PdeSrc.n.u1User,
2020	(uint64_t)PteSrc.u,
2021	(uint64_t)pPTDst->a[iPTDst].u,
2022	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2023	}
2024	return VINF_SUCCESS;
2025
2026	#else
2027	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
2028	return VERR_INTERNAL_ERROR;
2029	#endif
2030	}
2031
2032
2033	#if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2034	/**
2035	* Investigate page fault and handle write protection page faults caused by
2036	* dirty bit tracking.
2037	*
2038	* @returns VBox status code.
2039	* @param pVM VM handle.
2040	* @param uErr Page fault error code.
2041	* @param pPdeDst Shadow page directory entry.
2042	* @param pPdeSrc Guest page directory entry.
2043	* @param GCPtrPage Guest context page address.
2044	*/
2045	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCPTR GCPtrPage)
2046	{
2047	bool fWriteProtect = !!(CPUMGetGuestCR0(pVM) & X86_CR0_WP);
2048	bool fUserLevelFault = !!(uErr & X86_TRAP_PF_US);
2049	bool fWriteFault = !!(uErr & X86_TRAP_PF_RW);
2050	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2051	bool fBigPagesSupported = true;
2052	# else
2053	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2054	# endif
2055	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2056	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
2057	# endif
2058	unsigned uPageFaultLevel;
2059	int rc;
2060
2061	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2062	LogFlow(("CheckPageFault: GCPtrPage=%RGv uErr=%#x PdeSrc=%08x\n", GCPtrPage, uErr, pPdeSrc->u));
2063
2064	# if PGM_GST_TYPE == PGM_TYPE_PAE \
2065	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
2066
2067	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2068	PX86PML4E pPml4eSrc;
2069	PX86PDPE pPdpeSrc;
2070
2071	pPdpeSrc = pgmGstGetLongModePDPTPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc);
2072	Assert(pPml4eSrc);
2073
2074	/*
2075	* Real page fault? (PML4E level)
2076	*/
2077	if ( (uErr & X86_TRAP_PF_RSVD)
2078	\|\| !pPml4eSrc->n.u1Present
2079	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPml4eSrc->n.u1NoExecute)
2080	\|\| (fWriteFault && !pPml4eSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2081	\|\| (fUserLevelFault && !pPml4eSrc->n.u1User)
2082	)
2083	{
2084	uPageFaultLevel = 0;
2085	goto l_UpperLevelPageFault;
2086	}
2087	Assert(pPdpeSrc);
2088
2089	# else /* PAE */
2090	PX86PDPE pPdpeSrc = pgmGstGetPaePDPEPtr(&pVM->pgm.s, GCPtrPage);
2091	# endif /* PAE */
2092
2093	/*
2094	* Real page fault? (PDPE level)
2095	*/
2096	if ( (uErr & X86_TRAP_PF_RSVD)
2097	\|\| !pPdpeSrc->n.u1Present
2098	# if PGM_GST_TYPE == PGM_TYPE_AMD64 /* NX, r/w, u/s bits in the PDPE are long mode only */
2099	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdpeSrc->lm.u1NoExecute)
2100	\|\| (fWriteFault && !pPdpeSrc->lm.u1Write && (fUserLevelFault \|\| fWriteProtect))
2101	\|\| (fUserLevelFault && !pPdpeSrc->lm.u1User)
2102	# endif
2103	)
2104	{
2105	uPageFaultLevel = 1;
2106	goto l_UpperLevelPageFault;
2107	}
2108	# endif
2109
2110	/*
2111	* Real page fault? (PDE level)
2112	*/
2113	if ( (uErr & X86_TRAP_PF_RSVD)
2114	\|\| !pPdeSrc->n.u1Present
2115	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2116	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdeSrc->n.u1NoExecute)
2117	# endif
2118	\|\| (fWriteFault && !pPdeSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2119	\|\| (fUserLevelFault && !pPdeSrc->n.u1User) )
2120	{
2121	uPageFaultLevel = 2;
2122	goto l_UpperLevelPageFault;
2123	}
2124
2125	/*
2126	* First check the easy case where the page directory has been marked read-only to track
2127	* the dirty bit of an emulated BIG page
2128	*/
2129	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2130	{
2131	/* Mark guest page directory as accessed */
2132	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2133	pPml4eSrc->n.u1Accessed = 1;
2134	pPdpeSrc->lm.u1Accessed = 1;
2135	# endif
2136	pPdeSrc->b.u1Accessed = 1;
2137
2138	/*
2139	* Only write protection page faults are relevant here.
2140	*/
2141	if (fWriteFault)
2142	{
2143	/* Mark guest page directory as dirty (BIG page only). */
2144	pPdeSrc->b.u1Dirty = 1;
2145
2146	if (pPdeDst->n.u1Present && (pPdeDst->u & PGM_PDFLAGS_TRACK_DIRTY))
2147	{
2148	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2149
2150	Assert(pPdeSrc->b.u1Write);
2151
2152	pPdeDst->n.u1Write = 1;
2153	pPdeDst->n.u1Accessed = 1;
2154	pPdeDst->au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
2155	PGM_INVL_BIG_PG(GCPtrPage);
2156	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2157	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2158	}
2159	}
2160	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2161	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2162	}
2163	/* else: 4KB page table */
2164
2165	/*
2166	* Map the guest page table.
2167	*/
2168	PGSTPT pPTSrc;
2169	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2170	if (RT_SUCCESS(rc))
2171	{
2172	/*
2173	* Real page fault?
2174	*/
2175	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2176	const GSTPTE PteSrc = *pPteSrc;
2177	if ( !PteSrc.n.u1Present
2178	# if PGM_WITH_NX(PGM_GST_TYPE, PGM_SHW_TYPE)
2179	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && PteSrc.n.u1NoExecute)
2180	# endif
2181	\|\| (fWriteFault && !PteSrc.n.u1Write && (fUserLevelFault \|\| fWriteProtect))
2182	\|\| (fUserLevelFault && !PteSrc.n.u1User)
2183	)
2184	{
2185	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2186	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2187	LogFlow(("CheckPageFault: real page fault at %RGv PteSrc.u=%08x (2)\n", GCPtrPage, PteSrc.u));
2188
2189	/* Check the present bit as the shadow tables can cause different error codes by being out of sync.
2190	* See the 2nd case above as well.
2191	*/
2192	if (pPdeSrc->n.u1Present && pPteSrc->n.u1Present)
2193	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2194
2195	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2196	return VINF_EM_RAW_GUEST_TRAP;
2197	}
2198	LogFlow(("CheckPageFault: page fault at %RGv PteSrc.u=%08x\n", GCPtrPage, PteSrc.u));
2199
2200	/*
2201	* Set the accessed bits in the page directory and the page table.
2202	*/
2203	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2204	pPml4eSrc->n.u1Accessed = 1;
2205	pPdpeSrc->lm.u1Accessed = 1;
2206	# endif
2207	pPdeSrc->n.u1Accessed = 1;
2208	pPteSrc->n.u1Accessed = 1;
2209
2210	/*
2211	* Only write protection page faults are relevant here.
2212	*/
2213	if (fWriteFault)
2214	{
2215	/* Write access, so mark guest entry as dirty. */
2216	# ifdef VBOX_WITH_STATISTICS
2217	if (!pPteSrc->n.u1Dirty)
2218	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtiedPage));
2219	else
2220	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageAlreadyDirty));
2221	# endif
2222
2223	pPteSrc->n.u1Dirty = 1;
2224
2225	if (pPdeDst->n.u1Present)
2226	{
2227	#ifndef IN_RING0
2228	/* Bail out here as pgmPoolGetPageByHCPhys will return NULL and we'll crash below.
2229	* Our individual shadow handlers will provide more information and force a fatal exit.
2230	*/
2231	if (MMHyperIsInsideArea(pVM, (RTGCPTR)GCPtrPage))
2232	{
2233	LogRel(("CheckPageFault: write to hypervisor region %RGv\n", GCPtrPage));
2234	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2235	return VINF_SUCCESS;
2236	}
2237	#endif
2238	/*
2239	* Map shadow page table.
2240	*/
2241	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2242	if (pShwPage)
2243	{
2244	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2245	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2246	if ( pPteDst->n.u1Present /** @todo Optimize accessed bit emulation? */
2247	&& (pPteDst->u & PGM_PTFLAGS_TRACK_DIRTY))
2248	{
2249	LogFlow(("DIRTY page trap addr=%RGv\n", GCPtrPage));
2250	# ifdef VBOX_STRICT
2251	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, pPteSrc->u & GST_PTE_PG_MASK);
2252	if (pPage)
2253	AssertMsg(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage),
2254	("Unexpected dirty bit tracking on monitored page %RGv (phys %RGp)!!!!!!\n", GCPtrPage, pPteSrc->u & X86_PTE_PAE_PG_MASK));
2255	# endif
2256	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageTrap));
2257
2258	Assert(pPteSrc->n.u1Write);
2259
2260	pPteDst->n.u1Write = 1;
2261	pPteDst->n.u1Dirty = 1;
2262	pPteDst->n.u1Accessed = 1;
2263	pPteDst->au32[0] &= ~PGM_PTFLAGS_TRACK_DIRTY;
2264	PGM_INVL_PG(GCPtrPage);
2265
2266	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2267	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
2268	}
2269	}
2270	else
2271	AssertMsgFailed(("pgmPoolGetPageByHCPhys %RGp failed!\n", pPdeDst->u & SHW_PDE_PG_MASK));
2272	}
2273	}
2274	/** @todo Optimize accessed bit emulation? */
2275	# ifdef VBOX_STRICT
2276	/*
2277	* Sanity check.
2278	*/
2279	else if ( !pPteSrc->n.u1Dirty
2280	&& (pPdeSrc->n.u1Write & pPteSrc->n.u1Write)
2281	&& pPdeDst->n.u1Present)
2282	{
2283	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2284	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2285	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2286	if ( pPteDst->n.u1Present
2287	&& pPteDst->n.u1Write)
2288	LogFlow(("Writable present page %RGv not marked for dirty bit tracking!!!\n", GCPtrPage));
2289	}
2290	# endif /* VBOX_STRICT */
2291	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2292	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2293	}
2294	AssertRC(rc);
2295	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2296	return rc;
2297
2298
2299	l_UpperLevelPageFault:
2300	/*
2301	* Pagefault detected while checking the PML4E, PDPE or PDE.
2302	* Single exit handler to get rid of duplicate code paths.
2303	*/
2304	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyTrackRealPF));
2305	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyBitTracking), a);
2306	Log(("CheckPageFault: real page fault at %RGv (%d)\n", GCPtrPage, uPageFaultLevel));
2307
2308	if (
2309	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2310	pPml4eSrc->n.u1Present &&
2311	# endif
2312	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
2313	pPdpeSrc->n.u1Present &&
2314	# endif
2315	pPdeSrc->n.u1Present)
2316	{
2317	/* Check the present bit as the shadow tables can cause different error codes by being out of sync. */
2318	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2319	{
2320	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2321	}
2322	else
2323	{
2324	/*
2325	* Map the guest page table.
2326	*/
2327	PGSTPT pPTSrc;
2328	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2329	if (RT_SUCCESS(rc))
2330	{
2331	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2332	const GSTPTE PteSrc = *pPteSrc;
2333	if (pPteSrc->n.u1Present)
2334	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2335	}
2336	AssertRC(rc);
2337	}
2338	}
2339	return VINF_EM_RAW_GUEST_TRAP;
2340	}
2341	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
2342
2343
2344	/**
2345	* Sync a shadow page table.
2346	*
2347	* The shadow page table is not present. This includes the case where
2348	* there is a conflict with a mapping.
2349	*
2350	* @returns VBox status code.
2351	* @param pVM VM handle.
2352	* @param iPD Page directory index.
2353	* @param pPDSrc Source page directory (i.e. Guest OS page directory).
2354	* Assume this is a temporary mapping.
2355	* @param GCPtrPage GC Pointer of the page that caused the fault
2356	*/
2357	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPDSrc, PGSTPD pPDSrc, RTGCPTR GCPtrPage)
2358	{
2359	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2360	STAM_COUNTER_INC(&pVM->pgm.s.StatSyncPtPD[iPDSrc]);
2361	LogFlow(("SyncPT: GCPtrPage=%RGv\n", GCPtrPage));
2362
2363	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
2364	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
2365	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64) \
2366	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2367	&& PGM_SHW_TYPE != PGM_TYPE_EPT
2368
2369	int rc = VINF_SUCCESS;
2370
2371	/*
2372	* Validate input a little bit.
2373	*/
2374	AssertMsg(iPDSrc == ((GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK), ("iPDSrc=%x GCPtrPage=%RGv\n", iPDSrc, GCPtrPage));
2375	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2376	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
2377	PSHWPDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2378
2379	/* Fetch the pgm pool shadow descriptor. */
2380	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
2381	Assert(pShwPde);
2382
2383	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2384	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2385	PPGMPOOLPAGE pShwPde;
2386	PX86PDPAE pPDDst;
2387	PSHWPDE pPdeDst;
2388
2389	/* Fetch the pgm pool shadow descriptor. */
2390	rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
2391	AssertRCSuccessReturn(rc, rc);
2392	Assert(pShwPde);
2393
2394	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
2395	pPdeDst = &pPDDst->a[iPDDst];
2396
2397	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2398	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2399	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2400	PX86PDPAE pPDDst;
2401	PX86PDPT pPdptDst;
2402	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
2403	AssertRCSuccessReturn(rc, rc);
2404	Assert(pPDDst);
2405	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2406	# endif
2407	SHWPDE PdeDst = *pPdeDst;
2408
2409	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2410	/* Fetch the pgm pool shadow descriptor. */
2411	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
2412	Assert(pShwPde);
2413	# endif
2414
2415	# ifndef PGM_WITHOUT_MAPPINGS
2416	/*
2417	* Check for conflicts.
2418	* GC: In case of a conflict we'll go to Ring-3 and do a full SyncCR3.
2419	* HC: Simply resolve the conflict.
2420	*/
2421	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
2422	{
2423	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
2424	# ifndef IN_RING3
2425	Log(("SyncPT: Conflict at %RGv\n", GCPtrPage));
2426	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2427	return VERR_ADDRESS_CONFLICT;
2428	# else
2429	PPGMMAPPING pMapping = pgmGetMapping(pVM, (RTGCPTR)GCPtrPage);
2430	Assert(pMapping);
2431	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2432	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2433	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2434	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2435	# else
2436	AssertFailed(); /* can't happen for amd64 */
2437	# endif
2438	if (RT_FAILURE(rc))
2439	{
2440	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2441	return rc;
2442	}
2443	PdeDst = *pPdeDst;
2444	# endif
2445	}
2446	# else /* PGM_WITHOUT_MAPPINGS */
2447	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
2448	# endif /* PGM_WITHOUT_MAPPINGS */
2449	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2450
2451	# if defined(IN_RC)
2452	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
2453	PGMDynLockHCPage(pVM, (uint8_t *)pPdeDst);
2454	# endif
2455
2456	/*
2457	* Sync page directory entry.
2458	*/
2459	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2460	if (PdeSrc.n.u1Present)
2461	{
2462	/*
2463	* Allocate & map the page table.
2464	*/
2465	PSHWPT pPTDst;
2466	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2467	const bool fPageTable = !PdeSrc.b.u1Size;
2468	# else
2469	const bool fPageTable = !PdeSrc.b.u1Size \|\| !(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2470	# endif
2471	PPGMPOOLPAGE pShwPage;
2472	RTGCPHYS GCPhys;
2473	if (fPageTable)
2474	{
2475	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
2476	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2477	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2478	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE / 2);
2479	# endif
2480	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2481	}
2482	else
2483	{
2484	GCPhys = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
2485	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2486	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
2487	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
2488	# endif
2489	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, pShwPde->idx, iPDDst, &pShwPage);
2490	}
2491	if (rc == VINF_SUCCESS)
2492	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2493	else if (rc == VINF_PGM_CACHED_PAGE)
2494	{
2495	/*
2496	* The PT was cached, just hook it up.
2497	*/
2498	if (fPageTable)
2499	PdeDst.u = pShwPage->Core.Key
2500	\| (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));
2501	else
2502	{
2503	PdeDst.u = pShwPage->Core.Key
2504	\| (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));
2505	/* (see explanation and assumptions further down.) */
2506	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2507	{
2508	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2509	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2510	PdeDst.b.u1Write = 0;
2511	}
2512	}
2513	*pPdeDst = PdeDst;
2514	# if defined(IN_RC)
2515	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2516	# endif
2517	return VINF_SUCCESS;
2518	}
2519	else if (rc == VERR_PGM_POOL_FLUSHED)
2520	{
2521	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3);
2522	# if defined(IN_RC)
2523	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2524	# endif
2525	return VINF_PGM_SYNC_CR3;
2526	}
2527	else
2528	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2529	PdeDst.u &= X86_PDE_AVL_MASK;
2530	PdeDst.u \|= pShwPage->Core.Key;
2531
2532	/*
2533	* Page directory has been accessed (this is a fault situation, remember).
2534	*/
2535	pPDSrc->a[iPDSrc].n.u1Accessed = 1;
2536	if (fPageTable)
2537	{
2538	/*
2539	* Page table - 4KB.
2540	*
2541	* Sync all or just a few entries depending on PGM_SYNC_N_PAGES.
2542	*/
2543	Log2(("SyncPT: 4K %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx}\n",
2544	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u));
2545	PGSTPT pPTSrc;
2546	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
2547	if (RT_SUCCESS(rc))
2548	{
2549	/*
2550	* Start by syncing the page directory entry so CSAM's TLB trick works.
2551	*/
2552	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| X86_PDE_AVL_MASK))
2553	\| (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));
2554	*pPdeDst = PdeDst;
2555	# if defined(IN_RC)
2556	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2557	# endif
2558
2559	/*
2560	* Directory/page user or supervisor privilege: (same goes for read/write)
2561	*
2562	* Directory Page Combined
2563	* U/S U/S U/S
2564	* 0 0 0
2565	* 0 1 0
2566	* 1 0 0
2567	* 1 1 1
2568	*
2569	* Simple AND operation. Table listed for completeness.
2570	*
2571	*/
2572	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4K));
2573	# ifdef PGM_SYNC_N_PAGES
2574	unsigned iPTBase = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
2575	unsigned iPTDst = iPTBase;
2576	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, RT_ELEMENTS(pPTDst->a));
2577	if (iPTDst <= PGM_SYNC_NR_PAGES / 2)
2578	iPTDst = 0;
2579	else
2580	iPTDst -= PGM_SYNC_NR_PAGES / 2;
2581	# else /* !PGM_SYNC_N_PAGES */
2582	unsigned iPTDst = 0;
2583	const unsigned iPTDstEnd = RT_ELEMENTS(pPTDst->a);
2584	# endif /* !PGM_SYNC_N_PAGES */
2585	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2586	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2587	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
2588	# else
2589	const unsigned offPTSrc = 0;
2590	# endif
2591	for (; iPTDst < iPTDstEnd; iPTDst++)
2592	{
2593	const unsigned iPTSrc = iPTDst + offPTSrc;
2594	const GSTPTE PteSrc = pPTSrc->a[iPTSrc];
2595
2596	if (PteSrc.n.u1Present) /* we've already cleared it above */
2597	{
2598	# ifndef IN_RING0
2599	/*
2600	* Assuming kernel code will be marked as supervisor - and not as user level
2601	* and executed using a conforming code selector - And marked as readonly.
2602	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
2603	*/
2604	PPGMPAGE pPage;
2605	if ( ((PdeSrc.u & pPTSrc->a[iPTSrc].u) & (X86_PTE_RW \| X86_PTE_US))
2606	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)))
2607	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
2608	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2609	)
2610	# endif
2611	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
2612	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",
2613	(RTGCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)),
2614	PteSrc.n.u1Present,
2615	PteSrc.n.u1Write & PdeSrc.n.u1Write,
2616	PteSrc.n.u1User & PdeSrc.n.u1User,
2617	(uint64_t)PteSrc.u,
2618	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : "", pPTDst->a[iPTDst].u, iPTSrc, PdeSrc.au32[0],
2619	(RTGCPHYS)((PdeSrc.u & GST_PDE_PG_MASK) + iPTSrc*sizeof(PteSrc)) ));
2620	}
2621	} /* for PTEs */
2622	}
2623	}
2624	else
2625	{
2626	/*
2627	* Big page - 2/4MB.
2628	*
2629	* We'll walk the ram range list in parallel and optimize lookups.
2630	* We will only sync on shadow page table at a time.
2631	*/
2632	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT4M));
2633
2634	/**
2635	* @todo It might be more efficient to sync only a part of the 4MB page (similar to what we do for 4kb PDs).
2636	*/
2637
2638	/*
2639	* Start by syncing the page directory entry.
2640	*/
2641	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| (X86_PDE_AVL_MASK & ~PGM_PDFLAGS_TRACK_DIRTY)))
2642	\| (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));
2643
2644	/*
2645	* If the page is not flagged as dirty and is writable, then make it read-only
2646	* at PD level, so we can set the dirty bit when the page is modified.
2647	*
2648	* ASSUMES that page access handlers are implemented on page table entry level.
2649	* Thus we will first catch the dirty access and set PDE.D and restart. If
2650	* there is an access handler, we'll trap again and let it work on the problem.
2651	*/
2652	/** @todo move the above stuff to a section in the PGM documentation. */
2653	Assert(!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY));
2654	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2655	{
2656	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,DirtyPageBig));
2657	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2658	PdeDst.b.u1Write = 0;
2659	}
2660	*pPdeDst = PdeDst;
2661	# if defined(IN_RC)
2662	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
2663	# endif
2664
2665	/*
2666	* Fill the shadow page table.
2667	*/
2668	/* Get address and flags from the source PDE. */
2669	SHWPTE PteDstBase;
2670	PteDstBase.u = PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT);
2671
2672	/* Loop thru the entries in the shadow PT. */
2673	const RTGCPTR GCPtr = (GCPtrPage >> SHW_PD_SHIFT) << SHW_PD_SHIFT; NOREF(GCPtr);
2674	Log2(("SyncPT: BIG %RGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} Shw=%RGv GCPhys=%RGp %s\n",
2675	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u, GCPtr,
2676	GCPhys, PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2677	PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRanges);
2678	unsigned iPTDst = 0;
2679	while (iPTDst < RT_ELEMENTS(pPTDst->a))
2680	{
2681	/* Advance ram range list. */
2682	while (pRam && GCPhys > pRam->GCPhysLast)
2683	pRam = pRam->CTX_SUFF(pNext);
2684	if (pRam && GCPhys >= pRam->GCPhys)
2685	{
2686	unsigned iHCPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2687	do
2688	{
2689	/* Make shadow PTE. */
2690	PPGMPAGE pPage = &pRam->aPages[iHCPage];
2691	SHWPTE PteDst;
2692
2693	# ifdef VBOX_WITH_NEW_PHYS_CODE
2694	# ifndef VBOX_WITH_NEW_LAZY_PAGE_ALLOC
2695	/* Try make the page writable if necessary. */
2696	if ( PteDstBase.n.u1Write
2697	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
2698	&& PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
2699	{
2700	rc = pgmPhysPageMakeWritableUnlocked(pVM, pPage, GCPhys);
2701	AssertRCReturn(rc, rc);
2702	}
2703	# endif
2704	# else /* !VBOX_WITH_NEW_PHYS_CODE */
2705	/* Make sure the RAM has already been allocated. */
2706	if (pRam->fFlags & MM_RAM_FLAGS_DYNAMIC_ALLOC) /** @todo PAGE FLAGS */
2707	{
2708	if (RT_UNLIKELY(!PGM_PAGE_GET_HCPHYS(pPage)))
2709	{
2710	# ifdef IN_RING3
2711	int rc = pgmr3PhysGrowRange(pVM, GCPhys);
2712	# else
2713	int rc = CTXALLMID(VMM, CallHost)(pVM, VMMCALLHOST_PGM_RAM_GROW_RANGE, GCPhys);
2714	# endif
2715	if (rc != VINF_SUCCESS)
2716	return rc;
2717	}
2718	}
2719	# endif /* !VBOX_WITH_NEW_PHYS_CODE */
2720
2721	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2722	{
2723	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
2724	{
2725	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2726	PteDst.n.u1Write = 0;
2727	}
2728	else
2729	PteDst.u = 0;
2730	}
2731	# ifndef IN_RING0
2732	/*
2733	* Assuming kernel code will be marked as supervisor and not as user level and executed
2734	* using a conforming code selector. Don't check for readonly, as that implies the whole
2735	* 4MB can be code or readonly data. Linux enables write access for its large pages.
2736	*/
2737	else if ( !PdeSrc.n.u1User
2738	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT))))
2739	PteDst.u = 0;
2740	# endif
2741	else
2742	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2743
2744	# ifdef VBOX_WITH_NEW_PHYS_CODE
2745	/* Only map writable pages writable. */
2746	if ( PteDst.n.u1Write
2747	&& PteDst.n.u1Present
2748	&& PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED)
2749	{
2750	PteDst.n.u1Write = 0; /** @todo this isn't quite working yet... */
2751	Log3(("SyncPT: write-protecting %RGp pPage=%R[pgmpage] at %RGv\n", GCPhys, pPage, (RTGCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT))));
2752	}
2753	# endif
2754
2755	# ifdef PGMPOOL_WITH_USER_TRACKING
2756	if (PteDst.n.u1Present)
2757	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, PGM_PAGE_GET_TRACKING(pPage), pPage, iPTDst);
2758	# endif
2759	/* commit it */
2760	pPTDst->a[iPTDst] = PteDst;
2761	Log4(("SyncPT: BIG %RGv PteDst:{P=%d RW=%d U=%d raw=%08llx}%s\n",
2762	(RTGCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT)), PteDst.n.u1Present, PteDst.n.u1Write, PteDst.n.u1User, (uint64_t)PteDst.u,
2763	PteDst.u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2764
2765	/* advance */
2766	GCPhys += PAGE_SIZE;
2767	iHCPage++;
2768	iPTDst++;
2769	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2770	&& GCPhys <= pRam->GCPhysLast);
2771	}
2772	else if (pRam)
2773	{
2774	Log(("Invalid pages at %RGp\n", GCPhys));
2775	do
2776	{
2777	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2778	GCPhys += PAGE_SIZE;
2779	iPTDst++;
2780	} while ( iPTDst < RT_ELEMENTS(pPTDst->a)
2781	&& GCPhys < pRam->GCPhys);
2782	}
2783	else
2784	{
2785	Log(("Invalid pages at %RGp (2)\n", GCPhys));
2786	for ( ; iPTDst < RT_ELEMENTS(pPTDst->a); iPTDst++)
2787	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2788	}
2789	} /* while more PTEs */
2790	} /* 4KB / 4MB */
2791	}
2792	else
2793	AssertRelease(!PdeDst.n.u1Present);
2794
2795	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2796	if (RT_FAILURE(rc))
2797	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPTFailed));
2798	return rc;
2799
2800	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
2801	&& PGM_SHW_TYPE != PGM_TYPE_NESTED \
2802	&& (PGM_SHW_TYPE != PGM_TYPE_EPT \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
2803	&& !defined(IN_RC)
2804
2805	/*
2806	* Validate input a little bit.
2807	*/
2808	int rc = VINF_SUCCESS;
2809	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2810	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2811	PSHWPDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
2812
2813	/* Fetch the pgm pool shadow descriptor. */
2814	PPGMPOOLPAGE pShwPde = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
2815	Assert(pShwPde);
2816
2817	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2818	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2819	PPGMPOOLPAGE pShwPde;
2820	PX86PDPAE pPDDst;
2821	PSHWPDE pPdeDst;
2822
2823	/* Fetch the pgm pool shadow descriptor. */
2824	rc = pgmShwGetPaePoolPagePD(&pVM->pgm.s, GCPtrPage, &pShwPde);
2825	AssertRCSuccessReturn(rc, rc);
2826	Assert(pShwPde);
2827
2828	pPDDst = (PX86PDPAE)PGMPOOL_PAGE_2_PTR_BY_PGM(&pVM->pgm.s, pShwPde);
2829	pPdeDst = &pPDDst->a[iPDDst];
2830
2831	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2832	const unsigned iPdpt = (GCPtrPage >> X86_PDPT_SHIFT) & X86_PDPT_MASK_AMD64;
2833	const unsigned iPDDst = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
2834	PX86PDPAE pPDDst;
2835	PX86PDPT pPdptDst;
2836	rc = pgmShwGetLongModePDPtr(pVM, GCPtrPage, NULL, &pPdptDst, &pPDDst);
2837	AssertRCSuccessReturn(rc, rc);
2838	Assert(pPDDst);
2839	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2840
2841	/* Fetch the pgm pool shadow descriptor. */
2842	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & X86_PDPE_PG_MASK);
2843	Assert(pShwPde);
2844
2845	# elif PGM_SHW_TYPE == PGM_TYPE_EPT
2846	const unsigned iPdpt = (GCPtrPage >> EPT_PDPT_SHIFT) & EPT_PDPT_MASK;
2847	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2848	PEPTPD pPDDst;
2849	PEPTPDPT pPdptDst;
2850
2851	rc = pgmShwGetEPTPDPtr(pVM, GCPtrPage, &pPdptDst, &pPDDst);
2852	if (rc != VINF_SUCCESS)
2853	{
2854	AssertRC(rc);
2855	return rc;
2856	}
2857	Assert(pPDDst);
2858	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2859
2860	/* Fetch the pgm pool shadow descriptor. */
2861	PPGMPOOLPAGE pShwPde = pgmPoolGetPageByHCPhys(pVM, pPdptDst->a[iPdpt].u & EPT_PDPTE_PG_MASK);
2862	Assert(pShwPde);
2863	# endif
2864	SHWPDE PdeDst = *pPdeDst;
2865
2866	Assert(!(PdeDst.u & PGM_PDFLAGS_MAPPING));
2867	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2868
2869	GSTPDE PdeSrc;
2870	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2871	PdeSrc.n.u1Present = 1;
2872	PdeSrc.n.u1Write = 1;
2873	PdeSrc.n.u1Accessed = 1;
2874	PdeSrc.n.u1User = 1;
2875
2876	/*
2877	* Allocate & map the page table.
2878	*/
2879	PSHWPT pPTDst;
2880	PPGMPOOLPAGE pShwPage;
2881	RTGCPHYS GCPhys;
2882
2883	/* Virtual address = physical address */
2884	GCPhys = GCPtrPage & X86_PAGE_4K_BASE_MASK;
2885	rc = pgmPoolAlloc(pVM, GCPhys & ~(RT_BIT_64(SHW_PD_SHIFT) - 1), BTH_PGMPOOLKIND_PT_FOR_PT, pShwPde->idx, iPDDst, &pShwPage);
2886
2887	if ( rc == VINF_SUCCESS
2888	\|\| rc == VINF_PGM_CACHED_PAGE)
2889	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2890	else
2891	AssertMsgFailedReturn(("rc=%Rrc\n", rc), VERR_INTERNAL_ERROR);
2892
2893	PdeDst.u &= X86_PDE_AVL_MASK;
2894	PdeDst.u \|= pShwPage->Core.Key;
2895	PdeDst.n.u1Present = 1;
2896	PdeDst.n.u1Write = 1;
2897	# if PGM_SHW_TYPE == PGM_TYPE_EPT
2898	PdeDst.n.u1Execute = 1;
2899	# else
2900	PdeDst.n.u1User = 1;
2901	PdeDst.n.u1Accessed = 1;
2902	# endif
2903	*pPdeDst = PdeDst;
2904
2905	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, PGM_SYNC_NR_PAGES, 0 /* page not present */);
2906	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2907	return rc;
2908
2909	#else
2910	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_SHW_TYPE, PGM_GST_TYPE));
2911	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncPT), a);
2912	return VERR_INTERNAL_ERROR;
2913	#endif
2914	}
2915
2916
2917
2918	/**
2919	* Prefetch a page/set of pages.
2920	*
2921	* Typically used to sync commonly used pages before entering raw mode
2922	* after a CR3 reload.
2923	*
2924	* @returns VBox status code.
2925	* @param pVM VM handle.
2926	* @param GCPtrPage Page to invalidate.
2927	*/
2928	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCPTR GCPtrPage)
2929	{
2930	#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) \
2931	&& PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
2932	/*
2933	* Check that all Guest levels thru the PDE are present, getting the
2934	* PD and PDE in the processes.
2935	*/
2936	int rc = VINF_SUCCESS;
2937	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
2938	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2939	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
2940	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
2941	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2942	unsigned iPDSrc;
2943	X86PDPE PdpeSrc;
2944	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
2945	if (!pPDSrc)
2946	return VINF_SUCCESS; /* not present */
2947	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
2948	unsigned iPDSrc;
2949	PX86PML4E pPml4eSrc;
2950	X86PDPE PdpeSrc;
2951	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
2952	if (!pPDSrc)
2953	return VINF_SUCCESS; /* not present */
2954	# endif
2955	const GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2956	# else
2957	PGSTPD pPDSrc = NULL;
2958	const unsigned iPDSrc = 0;
2959	GSTPDE PdeSrc;
2960
2961	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2962	PdeSrc.n.u1Present = 1;
2963	PdeSrc.n.u1Write = 1;
2964	PdeSrc.n.u1Accessed = 1;
2965	PdeSrc.n.u1User = 1;
2966	# endif
2967
2968	if (PdeSrc.n.u1Present && PdeSrc.n.u1Accessed)
2969	{
2970	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2971	const X86PDE PdeDst = pgmShwGet32BitPDE(&pVM->pgm.s, GCPtrPage);
2972	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
2973	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2974	PX86PDPAE pPDDst;
2975	X86PDEPAE PdeDst;
2976	# if PGM_GST_TYPE != PGM_TYPE_PAE
2977	X86PDPE PdpeSrc;
2978
2979	/* Fake PDPT entry; access control handled on the page table level, so allow everything. */
2980	PdpeSrc.u = X86_PDPE_P; /* rw/us are reserved for PAE pdpte's; accessed bit causes invalid VT-x guest state errors */
2981	# endif
2982	int rc = pgmShwSyncPaePDPtr(pVM, GCPtrPage, &PdpeSrc, &pPDDst);
2983	if (rc != VINF_SUCCESS)
2984	{
2985	AssertRC(rc);
2986	return rc;
2987	}
2988	Assert(pPDDst);
2989	PdeDst = pPDDst->a[iPDDst];
2990
2991	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
2992	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
2993	PX86PDPAE pPDDst;
2994	X86PDEPAE PdeDst;
2995
2996	# if PGM_GST_TYPE == PGM_TYPE_PROT
2997	/* AMD-V nested paging */
2998	X86PML4E Pml4eSrc;
2999	X86PDPE PdpeSrc;
3000	PX86PML4E pPml4eSrc = &Pml4eSrc;
3001
3002	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
3003	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_NX \| X86_PML4E_A;
3004	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_NX \| X86_PDPE_A;
3005	# endif
3006
3007	int rc = pgmShwSyncLongModePDPtr(pVM, GCPtrPage, pPml4eSrc, &PdpeSrc, &pPDDst);
3008	if (rc != VINF_SUCCESS)
3009	{
3010	AssertRC(rc);
3011	return rc;
3012	}
3013	Assert(pPDDst);
3014	PdeDst = pPDDst->a[iPDDst];
3015	# endif
3016	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
3017	{
3018	if (!PdeDst.n.u1Present)
3019	/** r=bird: This guy will set the A bit on the PDE, probably harmless. */
3020	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
3021	else
3022	{
3023	/** @note We used to sync PGM_SYNC_NR_PAGES pages, which triggered assertions in CSAM, because
3024	* R/W attributes of nearby pages were reset. Not sure how that could happen. Anyway, it
3025	* makes no sense to prefetch more than one page.
3026	*/
3027	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
3028	if (RT_SUCCESS(rc))
3029	rc = VINF_SUCCESS;
3030	}
3031	}
3032	}
3033	return rc;
3034
3035	#elif PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3036	return VINF_SUCCESS; /* ignore */
3037	#endif
3038	}
3039
3040
3041
3042
3043	/**
3044	* Syncs a page during a PGMVerifyAccess() call.
3045	*
3046	* @returns VBox status code (informational included).
3047	* @param GCPtrPage The address of the page to sync.
3048	* @param fPage The effective guest page flags.
3049	* @param uErr The trap error code.
3050	*/
3051	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCPTR GCPtrPage, unsigned fPage, unsigned uErr)
3052	{
3053	LogFlow(("VerifyAccessSyncPage: GCPtrPage=%RGv fPage=%#x uErr=%#x\n", GCPtrPage, fPage, uErr));
3054
3055	Assert(!HWACCMIsNestedPagingActive(pVM));
3056	#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) \
3057	&& PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
3058
3059	# ifndef IN_RING0
3060	if (!(fPage & X86_PTE_US))
3061	{
3062	/*
3063	* Mark this page as safe.
3064	*/
3065	/** @todo not correct for pages that contain both code and data!! */
3066	Log(("CSAMMarkPage %RGv; scanned=%d\n", GCPtrPage, true));
3067	CSAMMarkPage(pVM, (RTRCPTR)GCPtrPage, true);
3068	}
3069	# endif
3070
3071	/*
3072	* Get guest PD and index.
3073	*/
3074	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
3075	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3076	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
3077	PGSTPD pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3078	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3079	unsigned iPDSrc;
3080	X86PDPE PdpeSrc;
3081	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc, &PdpeSrc);
3082
3083	if (pPDSrc)
3084	{
3085	Log(("PGMVerifyAccess: access violation for %RGv due to non-present PDPTR\n", GCPtrPage));
3086	return VINF_EM_RAW_GUEST_TRAP;
3087	}
3088	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
3089	unsigned iPDSrc;
3090	PX86PML4E pPml4eSrc;
3091	X86PDPE PdpeSrc;
3092	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3093	if (!pPDSrc)
3094	{
3095	Log(("PGMVerifyAccess: access violation for %RGv due to non-present PDPTR\n", GCPtrPage));
3096	return VINF_EM_RAW_GUEST_TRAP;
3097	}
3098	# endif
3099	# else
3100	PGSTPD pPDSrc = NULL;
3101	const unsigned iPDSrc = 0;
3102	# endif
3103	int rc = VINF_SUCCESS;
3104
3105	/*
3106	* First check if the shadow pd is present.
3107	*/
3108	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3109	PX86PDE pPdeDst = pgmShwGet32BitPDEPtr(&pVM->pgm.s, GCPtrPage);
3110	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3111	PX86PDEPAE pPdeDst;
3112	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
3113	PX86PDPAE pPDDst;
3114	# if PGM_GST_TYPE != PGM_TYPE_PAE
3115	X86PDPE PdpeSrc;
3116
3117	/* Fake PDPT entry; access control handled on the page table level, so allow everything. */
3118	PdpeSrc.u = X86_PDPE_P; /* rw/us are reserved for PAE pdpte's; accessed bit causes invalid VT-x guest state errors */
3119	# endif
3120	rc = pgmShwSyncPaePDPtr(pVM, GCPtrPage, &PdpeSrc, &pPDDst);
3121	if (rc != VINF_SUCCESS)
3122	{
3123	AssertRC(rc);
3124	return rc;
3125	}
3126	Assert(pPDDst);
3127	pPdeDst = &pPDDst->a[iPDDst];
3128
3129	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3130	const unsigned iPDDst = ((GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK);
3131	PX86PDPAE pPDDst;
3132	PX86PDEPAE pPdeDst;
3133
3134	# if PGM_GST_TYPE == PGM_TYPE_PROT
3135	/* AMD-V nested paging */
3136	X86PML4E Pml4eSrc;
3137	X86PDPE PdpeSrc;
3138	PX86PML4E pPml4eSrc = &Pml4eSrc;
3139
3140	/* Fake PML4 & PDPT entry; access control handled on the page table level, so allow everything. */
3141	Pml4eSrc.u = X86_PML4E_P \| X86_PML4E_RW \| X86_PML4E_US \| X86_PML4E_NX \| X86_PML4E_A;
3142	PdpeSrc.u = X86_PDPE_P \| X86_PDPE_RW \| X86_PDPE_US \| X86_PDPE_NX \| X86_PDPE_A;
3143	# endif
3144
3145	rc = pgmShwSyncLongModePDPtr(pVM, GCPtrPage, pPml4eSrc, &PdpeSrc, &pPDDst);
3146	if (rc != VINF_SUCCESS)
3147	{
3148	AssertRC(rc);
3149	return rc;
3150	}
3151	Assert(pPDDst);
3152	pPdeDst = &pPDDst->a[iPDDst];
3153	# endif
3154
3155	if (!pPdeDst->n.u1Present)
3156	{
3157	# if defined(IN_RC)
3158	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
3159	PGMDynLockHCPage(pVM, (uint8_t *)pPdeDst);
3160	# endif
3161	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
3162	# if defined(IN_RC)
3163	/* Make sure the dynamic pPdeDst mapping will not be reused during this function. */
3164	PGMDynUnlockHCPage(pVM, (uint8_t *)pPdeDst);
3165	# endif
3166	AssertRC(rc);
3167	if (rc != VINF_SUCCESS)
3168	return rc;
3169	}
3170
3171	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
3172	/* Check for dirty bit fault */
3173	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, pPdeDst, &pPDSrc->a[iPDSrc], GCPtrPage);
3174	if (rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT)
3175	Log(("PGMVerifyAccess: success (dirty)\n"));
3176	else
3177	{
3178	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
3179	#else
3180	{
3181	GSTPDE PdeSrc;
3182	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
3183	PdeSrc.n.u1Present = 1;
3184	PdeSrc.n.u1Write = 1;
3185	PdeSrc.n.u1Accessed = 1;
3186	PdeSrc.n.u1User = 1;
3187
3188	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE) */
3189	Assert(rc != VINF_EM_RAW_GUEST_TRAP);
3190	if (uErr & X86_TRAP_PF_US)
3191	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncUser));
3192	else /* supervisor */
3193	STAM_COUNTER_INC(&pVM->pgm.s.CTX_MID_Z(Stat,PageOutOfSyncSupervisor));
3194
3195	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
3196	if (RT_SUCCESS(rc))
3197	{
3198	/* Page was successfully synced */
3199	Log2(("PGMVerifyAccess: success (sync)\n"));
3200	rc = VINF_SUCCESS;
3201	}
3202	else
3203	{
3204	Log(("PGMVerifyAccess: access violation for %RGv rc=%d\n", GCPtrPage, rc));
3205	return VINF_EM_RAW_GUEST_TRAP;
3206	}
3207	}
3208	return rc;
3209
3210	#else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
3211
3212	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
3213	return VERR_INTERNAL_ERROR;
3214	#endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
3215	}
3216
3217
3218	#if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3219	# if PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_AMD64
3220	/**
3221	* Figures out which kind of shadow page this guest PDE warrants.
3222	*
3223	* @returns Shadow page kind.
3224	* @param pPdeSrc The guest PDE in question.
3225	* @param cr4 The current guest cr4 value.
3226	*/
3227	DECLINLINE(PGMPOOLKIND) PGM_BTH_NAME(CalcPageKind)(const GSTPDE *pPdeSrc, uint32_t cr4)
3228	{
3229	# if PMG_GST_TYPE == PGM_TYPE_AMD64
3230	if (!pPdeSrc->n.u1Size)
3231	# else
3232	if (!pPdeSrc->n.u1Size \|\| !(cr4 & X86_CR4_PSE))
3233	# endif
3234	return BTH_PGMPOOLKIND_PT_FOR_PT;
3235	//switch (pPdeSrc->u & (X86_PDE4M_RW \| X86_PDE4M_US /\| X86_PDE4M_PAE_NX/))
3236	//{
3237	// case 0:
3238	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RO;
3239	// case X86_PDE4M_RW:
3240	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW;
3241	// case X86_PDE4M_US:
3242	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US;
3243	// case X86_PDE4M_RW \| X86_PDE4M_US:
3244	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US;
3245	# if 0
3246	// case X86_PDE4M_PAE_NX:
3247	// return BTH_PGMPOOLKIND_PT_FOR_BIG_NX;
3248	// case X86_PDE4M_RW \| X86_PDE4M_PAE_NX:
3249	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_NX;
3250	// case X86_PDE4M_US \| X86_PDE4M_PAE_NX:
3251	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US_NX;
3252	// case X86_PDE4M_RW \| X86_PDE4M_US \| X86_PDE4M_PAE_NX:
3253	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US_NX;
3254	# endif
3255	return BTH_PGMPOOLKIND_PT_FOR_BIG;
3256	//}
3257	}
3258	# endif
3259	#endif
3260
3261	#undef MY_STAM_COUNTER_INC
3262	#define MY_STAM_COUNTER_INC(a) do { } while (0)
3263
3264
3265	/**
3266	* Syncs the paging hierarchy starting at CR3.
3267	*
3268	* @returns VBox status code, no specials.
3269	* @param pVM The virtual machine.
3270	* @param cr0 Guest context CR0 register
3271	* @param cr3 Guest context CR3 register
3272	* @param cr4 Guest context CR4 register
3273	* @param fGlobal Including global page directories or not
3274	*/
3275	PGM_BTH_DECL(int, SyncCR3)(PVM pVM, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal)
3276	{
3277	if (VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
3278	fGlobal = true; /* Change this CR3 reload to be a global one. */
3279
3280	LogFlow(("SyncCR3 %d\n", fGlobal));
3281
3282	#if PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT
3283	/*
3284	* Update page access handlers.
3285	* The virtual are always flushed, while the physical are only on demand.
3286	* WARNING: We are incorrectly not doing global flushing on Virtual Handler updates. We'll
3287	* have to look into that later because it will have a bad influence on the performance.
3288	* @note SvL: There's no need for that. Just invalidate the virtual range(s).
3289	* bird: Yes, but that won't work for aliases.
3290	*/
3291	/** @todo this MUST go away. See #1557. */
3292	STAM_PROFILE_START(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Handlers), h);
3293	PGM_GST_NAME(HandlerVirtualUpdate)(pVM, cr4);
3294	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_MID_Z(Stat,SyncCR3Handlers), h);
3295	#endif
3296
3297	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3298	/*
3299	* Nested / EPT - almost no work.
3300	*/
3301	/** @todo check if this is really necessary; the call does it as well... */
3302	HWACCMFlushTLB(pVM);
3303	return VINF_SUCCESS;
3304
3305	#elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3306	/*
3307	* AMD64 (Shw & Gst) - No need to check all paging levels; we zero
3308	* out the shadow parts when the guest modifies its tables.
3309	*/
3310	return VINF_SUCCESS;
3311
3312	#else /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT && PGM_SHW_TYPE != PGM_TYPE_AMD64 */
3313
3314	# ifdef PGM_WITHOUT_MAPPINGS
3315	Assert(pVM->pgm.s.fMappingsFixed);
3316	return VINF_SUCCESS;
3317	# else
3318	/* Nothing to do when mappings are fixed. */
3319	if (pVM->pgm.s.fMappingsFixed)
3320	return VINF_SUCCESS;
3321
3322	int rc = PGMMapResolveConflicts(pVM);
3323	Assert(rc == VINF_SUCCESS \|\| rc == VINF_PGM_SYNC_CR3);
3324	if (rc == VINF_PGM_SYNC_CR3)
3325	{
3326	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3327	return VINF_PGM_SYNC_CR3;
3328	}
3329	# endif
3330	return VINF_SUCCESS;
3331	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT && PGM_SHW_TYPE != PGM_TYPE_AMD64 */
3332	}
3333
3334
3335
3336
3337	#ifdef VBOX_STRICT
3338	#ifdef IN_RC
3339	# undef AssertMsgFailed
3340	# define AssertMsgFailed Log
3341	#endif
3342	#ifdef IN_RING3
3343	# include <VBox/dbgf.h>
3344
3345	/**
3346	* Dumps a page table hierarchy use only physical addresses and cr4/lm flags.
3347	*
3348	* @returns VBox status code (VINF_SUCCESS).
3349	* @param pVM The VM handle.
3350	* @param cr3 The root of the hierarchy.
3351	* @param crr The cr4, only PAE and PSE is currently used.
3352	* @param fLongMode Set if long mode, false if not long mode.
3353	* @param cMaxDepth Number of levels to dump.
3354	* @param pHlp Pointer to the output functions.
3355	*/
3356	__BEGIN_DECLS
3357	VMMR3DECL(int) PGMR3DumpHierarchyHC(PVM pVM, uint32_t cr3, uint32_t cr4, bool fLongMode, unsigned cMaxDepth, PCDBGFINFOHLP pHlp);
3358	__END_DECLS
3359
3360	#endif
3361
3362	/**
3363	* Checks that the shadow page table is in sync with the guest one.
3364	*
3365	* @returns The number of errors.
3366	* @param pVM The virtual machine.
3367	* @param cr3 Guest context CR3 register
3368	* @param cr4 Guest context CR4 register
3369	* @param GCPtr Where to start. Defaults to 0.
3370	* @param cb How much to check. Defaults to everything.
3371	*/
3372	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCPTR GCPtr, RTGCPTR cb)
3373	{
3374	#if PGM_SHW_TYPE == PGM_TYPE_NESTED \|\| PGM_SHW_TYPE == PGM_TYPE_EPT
3375	return 0;
3376	#else
3377	unsigned cErrors = 0;
3378
3379	#if PGM_GST_TYPE == PGM_TYPE_PAE
3380	/** @todo currently broken; crashes below somewhere */
3381	AssertFailed();
3382	#endif
3383
3384	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
3385	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
3386	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3387
3388	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3389	bool fBigPagesSupported = true;
3390	# else
3391	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
3392	# endif
3393	PPGM pPGM = &pVM->pgm.s;
3394	RTGCPHYS GCPhysGst; /* page address derived from the guest page tables. */
3395	RTHCPHYS HCPhysShw; /* page address derived from the shadow page tables. */
3396	# ifndef IN_RING0
3397	RTHCPHYS HCPhys; /* general usage. */
3398	# endif
3399	int rc;
3400
3401	/*
3402	* Check that the Guest CR3 and all its mappings are correct.
3403	*/
3404	AssertMsgReturn(pPGM->GCPhysCR3 == (cr3 & GST_CR3_PAGE_MASK),
3405	("Invalid GCPhysCR3=%RGp cr3=%RGp\n", pPGM->GCPhysCR3, (RTGCPHYS)cr3),
3406	false);
3407	# if !defined(IN_RING0) && PGM_GST_TYPE != PGM_TYPE_AMD64
3408	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3409	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGst32BitPdRC, NULL, &HCPhysShw);
3410	# else
3411	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGstPaePdptRC, NULL, &HCPhysShw);
3412	# endif
3413	AssertRCReturn(rc, 1);
3414	HCPhys = NIL_RTHCPHYS;
3415	rc = pgmRamGCPhys2HCPhys(pPGM, cr3 & GST_CR3_PAGE_MASK, &HCPhys);
3416	AssertMsgReturn(HCPhys == HCPhysShw, ("HCPhys=%RHp HCPhyswShw=%RHp (cr3)\n", HCPhys, HCPhysShw), false);
3417	# if PGM_GST_TYPE == PGM_TYPE_32BIT && defined(IN_RING3)
3418	pgmGstGet32bitPDPtr(pPGM);
3419	RTGCPHYS GCPhys;
3420	rc = PGMR3DbgR3Ptr2GCPhys(pVM, pPGM->pGst32BitPdR3, &GCPhys);
3421	AssertRCReturn(rc, 1);
3422	AssertMsgReturn((cr3 & GST_CR3_PAGE_MASK) == GCPhys, ("GCPhys=%RGp cr3=%RGp\n", GCPhys, (RTGCPHYS)cr3), false);
3423	# endif
3424	# endif /* !IN_RING0 */
3425
3426	/*
3427	* Get and check the Shadow CR3.
3428	*/
3429	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3430	unsigned cPDEs = X86_PG_ENTRIES;
3431	unsigned cIncrement = X86_PG_ENTRIES * PAGE_SIZE;
3432	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
3433	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3434	unsigned cPDEs = X86_PG_PAE_ENTRIES * 4; /* treat it as a 2048 entry table. */
3435	# else
3436	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3437	# endif
3438	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3439	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
3440	unsigned cPDEs = X86_PG_PAE_ENTRIES;
3441	unsigned cIncrement = X86_PG_PAE_ENTRIES * PAGE_SIZE;
3442	# endif
3443	if (cb != ~(RTGCPTR)0)
3444	cPDEs = RT_MIN(cb >> SHW_PD_SHIFT, 1);
3445
3446	/** @todo call the other two PGMAssert() functions. /
3447
3448	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3449	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3450	# endif
3451
3452	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3453	unsigned iPml4 = (GCPtr >> X86_PML4_SHIFT) & X86_PML4_MASK;
3454
3455	for (; iPml4 < X86_PG_PAE_ENTRIES; iPml4++)
3456	{
3457	PPGMPOOLPAGE pShwPdpt = NULL;
3458	PX86PML4E pPml4eSrc;
3459	PX86PML4E pPml4eDst;
3460	RTGCPHYS GCPhysPdptSrc;
3461
3462	pPml4eSrc = pgmGstGetLongModePML4EPtr(&pVM->pgm.s, iPml4);
3463	pPml4eDst = pgmShwGetLongModePML4EPtr(&pVM->pgm.s, iPml4);
3464
3465	/* Fetch the pgm pool shadow descriptor if the shadow pml4e is present. */
3466	if (!pPml4eDst->n.u1Present)
3467	{
3468	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3469	continue;
3470	}
3471
3472	pShwPdpt = pgmPoolGetPage(pPool, pPml4eDst->u & X86_PML4E_PG_MASK);
3473	GCPhysPdptSrc = pPml4eSrc->u & X86_PML4E_PG_MASK_FULL;
3474
3475	if (pPml4eSrc->n.u1Present != pPml4eDst->n.u1Present)
3476	{
3477	AssertMsgFailed(("Present bit doesn't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3478	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3479	cErrors++;
3480	continue;
3481	}
3482
3483	if (GCPhysPdptSrc != pShwPdpt->GCPhys)
3484	{
3485	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));
3486	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3487	cErrors++;
3488	continue;
3489	}
3490
3491	if ( pPml4eDst->n.u1User != pPml4eSrc->n.u1User
3492	\|\| pPml4eDst->n.u1Write != pPml4eSrc->n.u1Write
3493	\|\| pPml4eDst->n.u1NoExecute != pPml4eSrc->n.u1NoExecute)
3494	{
3495	AssertMsgFailed(("User/Write/NoExec bits don't match! pPml4eDst.u=%#RX64 pPml4eSrc.u=%RX64\n", pPml4eDst->u, pPml4eSrc->u));
3496	GCPtr += _2M * UINT64_C(512) * UINT64_C(512);
3497	cErrors++;
3498	continue;
3499	}
3500	# else /* PGM_GST_TYPE != PGM_TYPE_AMD64 */
3501	{
3502	# endif /* PGM_GST_TYPE != PGM_TYPE_AMD64 */
3503
3504	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
3505	/*
3506	* Check the PDPTEs too.
3507	*/
3508	unsigned iPdpt = (GCPtr >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
3509
3510	for (;iPdpt <= SHW_PDPT_MASK; iPdpt++)
3511	{
3512	unsigned iPDSrc;
3513	PPGMPOOLPAGE pShwPde = NULL;
3514	PX86PDPE pPdpeDst;
3515	RTGCPHYS GCPhysPdeSrc;
3516	# if PGM_GST_TYPE == PGM_TYPE_PAE
3517	X86PDPE PdpeSrc;
3518	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtr, &iPDSrc, &PdpeSrc);
3519	PX86PDPT pPdptDst = pgmShwGetPaePDPTPtr(&pVM->pgm.s);
3520	# else
3521	PX86PML4E pPml4eSrc;
3522	X86PDPE PdpeSrc;
3523	PX86PDPT pPdptDst;
3524	PX86PDPAE pPDDst;
3525	PGSTPD pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtr, &pPml4eSrc, &PdpeSrc, &iPDSrc);
3526
3527	rc = pgmShwGetLongModePDPtr(pVM, GCPtr, NULL, &pPdptDst, &pPDDst);
3528	if (rc != VINF_SUCCESS)
3529	{
3530	AssertMsg(rc == VERR_PAGE_DIRECTORY_PTR_NOT_PRESENT, ("Unexpected rc=%Rrc\n", rc));
3531	GCPtr += 512 * _2M;
3532	continue; /* next PDPTE */
3533	}
3534	Assert(pPDDst);
3535	# endif
3536	Assert(iPDSrc == 0);
3537
3538	pPdpeDst = &pPdptDst->a[iPdpt];
3539
3540	if (!pPdpeDst->n.u1Present)
3541	{
3542	GCPtr += 512 * _2M;
3543	continue; /* next PDPTE */
3544	}
3545
3546	pShwPde = pgmPoolGetPage(pPool, pPdpeDst->u & X86_PDPE_PG_MASK);
3547	GCPhysPdeSrc = PdpeSrc.u & X86_PDPE_PG_MASK;
3548
3549	if (pPdpeDst->n.u1Present != PdpeSrc.n.u1Present)
3550	{
3551	AssertMsgFailed(("Present bit doesn't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3552	GCPtr += 512 * _2M;
3553	cErrors++;
3554	continue;
3555	}
3556
3557	if (GCPhysPdeSrc != pShwPde->GCPhys)
3558	{
3559	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3560	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));
3561	# else
3562	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));
3563	# endif
3564	GCPtr += 512 * _2M;
3565	cErrors++;
3566	continue;
3567	}
3568
3569	# if PGM_GST_TYPE == PGM_TYPE_AMD64
3570	if ( pPdpeDst->lm.u1User != PdpeSrc.lm.u1User
3571	\|\| pPdpeDst->lm.u1Write != PdpeSrc.lm.u1Write
3572	\|\| pPdpeDst->lm.u1NoExecute != PdpeSrc.lm.u1NoExecute)
3573	{
3574	AssertMsgFailed(("User/Write/NoExec bits don't match! pPdpeDst.u=%#RX64 pPdpeSrc.u=%RX64\n", pPdpeDst->u, PdpeSrc.u));
3575	GCPtr += 512 * _2M;
3576	cErrors++;
3577	continue;
3578	}
3579	# endif
3580
3581	# else /* PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PAE */
3582	{
3583	# endif /* PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PAE */
3584	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3585	GSTPD const *pPDSrc = pgmGstGet32bitPDPtr(&pVM->pgm.s);
3586	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3587	PCX86PD pPDDst = pgmShwGet32BitPDPtr(&pVM->pgm.s);
3588	# endif
3589	# endif /* PGM_GST_TYPE == PGM_TYPE_32BIT */
3590	/*
3591	* Iterate the shadow page directory.
3592	*/
3593	GCPtr = (GCPtr >> SHW_PD_SHIFT) << SHW_PD_SHIFT;
3594	unsigned iPDDst = (GCPtr >> SHW_PD_SHIFT) & SHW_PD_MASK;
3595
3596	for (;
3597	iPDDst < cPDEs;
3598	iPDDst++, GCPtr += cIncrement)
3599	{
3600	# if PGM_SHW_TYPE == PGM_TYPE_PAE
3601	const SHWPDE PdeDst = *pgmShwGetPaePDEPtr(pPGM, GCPtr);
3602	# else
3603	const SHWPDE PdeDst = pPDDst->a[iPDDst];
3604	# endif
3605	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
3606	{
3607	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3608	if ((PdeDst.u & X86_PDE_AVL_MASK) != PGM_PDFLAGS_MAPPING)
3609	{
3610	AssertMsgFailed(("Mapping shall only have PGM_PDFLAGS_MAPPING set! PdeDst.u=%#RX64\n", (uint64_t)PdeDst.u));
3611	cErrors++;
3612	continue;
3613	}
3614	}
3615	else if ( (PdeDst.u & X86_PDE_P)
3616	\|\| ((PdeDst.u & (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY)) == (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY))
3617	)
3618	{
3619	HCPhysShw = PdeDst.u & SHW_PDE_PG_MASK;
3620	PPGMPOOLPAGE pPoolPage = pgmPoolGetPageByHCPhys(pVM, HCPhysShw);
3621	if (!pPoolPage)
3622	{
3623	AssertMsgFailed(("Invalid page table address %RHp at %RGv! PdeDst=%#RX64\n",
3624	HCPhysShw, GCPtr, (uint64_t)PdeDst.u));
3625	cErrors++;
3626	continue;
3627	}
3628	const SHWPT pPTDst = (const SHWPT )PGMPOOL_PAGE_2_PTR(pVM, pPoolPage);
3629
3630	if (PdeDst.u & (X86_PDE4M_PWT \| X86_PDE4M_PCD))
3631	{
3632	AssertMsgFailed(("PDE flags PWT and/or PCD is set at %RGv! These flags are not virtualized! PdeDst=%#RX64\n",
3633	GCPtr, (uint64_t)PdeDst.u));
3634	cErrors++;
3635	}
3636
3637	if (PdeDst.u & (X86_PDE4M_G \| X86_PDE4M_D))
3638	{
3639	AssertMsgFailed(("4K PDE reserved flags at %RGv! PdeDst=%#RX64\n",
3640	GCPtr, (uint64_t)PdeDst.u));
3641	cErrors++;
3642	}
3643
3644	const GSTPDE PdeSrc = pPDSrc->a[(iPDDst >> (GST_PD_SHIFT - SHW_PD_SHIFT)) & GST_PD_MASK];
3645	if (!PdeSrc.n.u1Present)
3646	{
3647	AssertMsgFailed(("Guest PDE at %RGv is not present! PdeDst=%#RX64 PdeSrc=%#RX64\n",
3648	GCPtr, (uint64_t)PdeDst.u, (uint64_t)PdeSrc.u));
3649	cErrors++;
3650	continue;
3651	}
3652
3653	if ( !PdeSrc.b.u1Size
3654	\|\| !fBigPagesSupported)
3655	{
3656	GCPhysGst = PdeSrc.u & GST_PDE_PG_MASK;
3657	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3658	GCPhysGst \|= (iPDDst & 1) * (PAGE_SIZE / 2);
3659	# endif
3660	}
3661	else
3662	{
3663	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3664	if (PdeSrc.u & X86_PDE4M_PG_HIGH_MASK)
3665	{
3666	AssertMsgFailed(("Guest PDE at %RGv is using PSE36 or similar! PdeSrc=%#RX64\n",
3667	GCPtr, (uint64_t)PdeSrc.u));
3668	cErrors++;
3669	continue;
3670	}
3671	# endif
3672	GCPhysGst = GST_GET_PDE_BIG_PG_GCPHYS(PdeSrc);
3673	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3674	GCPhysGst \|= GCPtr & RT_BIT(X86_PAGE_2M_SHIFT);
3675	# endif
3676	}
3677
3678	if ( pPoolPage->enmKind
3679	!= (!PdeSrc.b.u1Size \|\| !fBigPagesSupported ? BTH_PGMPOOLKIND_PT_FOR_PT : BTH_PGMPOOLKIND_PT_FOR_BIG))
3680	{
3681	AssertMsgFailed(("Invalid shadow page table kind %d at %RGv! PdeSrc=%#RX64\n",
3682	pPoolPage->enmKind, GCPtr, (uint64_t)PdeSrc.u));
3683	cErrors++;
3684	}
3685
3686	PPGMPAGE pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3687	if (!pPhysPage)
3688	{
3689	AssertMsgFailed(("Cannot find guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
3690	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3691	cErrors++;
3692	continue;
3693	}
3694
3695	if (GCPhysGst != pPoolPage->GCPhys)
3696	{
3697	AssertMsgFailed(("GCPhysGst=%RGp != pPage->GCPhys=%RGp at %RGv\n",
3698	GCPhysGst, pPoolPage->GCPhys, GCPtr));
3699	cErrors++;
3700	continue;
3701	}
3702
3703	if ( !PdeSrc.b.u1Size
3704	\|\| !fBigPagesSupported)
3705	{
3706	/*
3707	* Page Table.
3708	*/
3709	const GSTPT *pPTSrc;
3710	rc = PGM_GCPHYS_2_PTR(pVM, GCPhysGst & ~(RTGCPHYS)(PAGE_SIZE - 1), &pPTSrc);
3711	if (RT_FAILURE(rc))
3712	{
3713	AssertMsgFailed(("Cannot map/convert guest physical address %RGp in the PDE at %RGv! PdeSrc=%#RX64\n",
3714	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3715	cErrors++;
3716	continue;
3717	}
3718	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/))
3719	!= (PdeDst.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/)))
3720	{
3721	/// @todo We get here a lot on out-of-sync CR3 entries. The access handler should zap them to avoid false alarms here!
3722	// (This problem will go away when/if we shadow multiple CR3s.)
3723	AssertMsgFailed(("4K PDE flags mismatch at %RGv! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3724	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3725	cErrors++;
3726	continue;
3727	}
3728	if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
3729	{
3730	AssertMsgFailed(("4K PDEs cannot have PGM_PDFLAGS_TRACK_DIRTY set! GCPtr=%RGv PdeDst=%#RX64\n",
3731	GCPtr, (uint64_t)PdeDst.u));
3732	cErrors++;
3733	continue;
3734	}
3735
3736	/* iterate the page table. */
3737	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3738	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
3739	const unsigned offPTSrc = ((GCPtr >> SHW_PD_SHIFT) & 1) * 512;
3740	# else
3741	const unsigned offPTSrc = 0;
3742	# endif
3743	for (unsigned iPT = 0, off = 0;
3744	iPT < RT_ELEMENTS(pPTDst->a);
3745	iPT++, off += PAGE_SIZE)
3746	{
3747	const SHWPTE PteDst = pPTDst->a[iPT];
3748
3749	/* skip not-present entries. */
3750	if (!(PteDst.u & (X86_PTE_P \| PGM_PTFLAGS_TRACK_DIRTY))) /** @todo deal with ALL handlers and CSAM !P pages! */
3751	continue;
3752	Assert(PteDst.n.u1Present);
3753
3754	const GSTPTE PteSrc = pPTSrc->a[iPT + offPTSrc];
3755	if (!PteSrc.n.u1Present)
3756	{
3757	# ifdef IN_RING3
3758	PGMAssertHandlerAndFlagsInSync(pVM);
3759	PGMR3DumpHierarchyGC(pVM, cr3, cr4, (PdeSrc.u & GST_PDE_PG_MASK));
3760	# endif
3761	AssertMsgFailed(("Out of sync (!P) PTE at %RGv! PteSrc=%#RX64 PteDst=%#RX64 pPTSrc=%RGv iPTSrc=%x PdeSrc=%x physpte=%RGp\n",
3762	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u, pPTSrc, iPT + offPTSrc, PdeSrc.au32[0],
3763	(PdeSrc.u & GST_PDE_PG_MASK) + (iPT + offPTSrc)*sizeof(PteSrc)));
3764	cErrors++;
3765	continue;
3766	}
3767
3768	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;
3769	# if 1 /** @todo sync accessed bit properly... */
3770	fIgnoreFlags \|= X86_PTE_A;
3771	# endif
3772
3773	/* match the physical addresses */
3774	HCPhysShw = PteDst.u & SHW_PTE_PG_MASK;
3775	GCPhysGst = PteSrc.u & GST_PTE_PG_MASK;
3776
3777	# ifdef IN_RING3
3778	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
3779	if (RT_FAILURE(rc))
3780	{
3781	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3782	{
3783	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
3784	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3785	cErrors++;
3786	continue;
3787	}
3788	}
3789	else if (HCPhysShw != (HCPhys & SHW_PTE_PG_MASK))
3790	{
3791	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3792	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3793	cErrors++;
3794	continue;
3795	}
3796	# endif
3797
3798	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3799	if (!pPhysPage)
3800	{
3801	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
3802	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3803	{
3804	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PteSrc=%#RX64 PteDst=%#RX64\n",
3805	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3806	cErrors++;
3807	continue;
3808	}
3809	# endif
3810	if (PteDst.n.u1Write)
3811	{
3812	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3813	GCPtr + off, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3814	cErrors++;
3815	}
3816	fIgnoreFlags \|= X86_PTE_RW;
3817	}
3818	else if (HCPhysShw != PGM_PAGE_GET_HCPHYS(pPhysPage))
3819	{
3820	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp pPhysPage:%R[pgmpage] GCPhysGst=%RGp PteSrc=%#RX64 PteDst=%#RX64\n",
3821	GCPtr + off, HCPhysShw, pPhysPage, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3822	cErrors++;
3823	continue;
3824	}
3825
3826	/* flags */
3827	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
3828	{
3829	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
3830	{
3831	if (PteDst.n.u1Write)
3832	{
3833	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! pPhysPage=%R[pgmpage] PteSrc=%#RX64 PteDst=%#RX64\n",
3834	GCPtr + off, pPhysPage, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3835	cErrors++;
3836	continue;
3837	}
3838	fIgnoreFlags \|= X86_PTE_RW;
3839	}
3840	else
3841	{
3842	if (PteDst.n.u1Present)
3843	{
3844	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! pPhysPage=%R[pgmpage] PteSrc=%#RX64 PteDst=%#RX64\n",
3845	GCPtr + off, pPhysPage, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3846	cErrors++;
3847	continue;
3848	}
3849	fIgnoreFlags \|= X86_PTE_P;
3850	}
3851	}
3852	else
3853	{
3854	if (!PteSrc.n.u1Dirty && PteSrc.n.u1Write)
3855	{
3856	if (PteDst.n.u1Write)
3857	{
3858	AssertMsgFailed(("!DIRTY page at %RGv is writable! PteSrc=%#RX64 PteDst=%#RX64\n",
3859	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3860	cErrors++;
3861	continue;
3862	}
3863	if (!(PteDst.u & PGM_PTFLAGS_TRACK_DIRTY))
3864	{
3865	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3866	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3867	cErrors++;
3868	continue;
3869	}
3870	if (PteDst.n.u1Dirty)
3871	{
3872	AssertMsgFailed(("!DIRTY page at %RGv is marked DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3873	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3874	cErrors++;
3875	}
3876	# if 0 /** @todo sync access bit properly... */
3877	if (PteDst.n.u1Accessed != PteSrc.n.u1Accessed)
3878	{
3879	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
3880	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3881	cErrors++;
3882	}
3883	fIgnoreFlags \|= X86_PTE_RW;
3884	# else
3885	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
3886	# endif
3887	}
3888	else if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
3889	{
3890	/* access bit emulation (not implemented). */
3891	if (PteSrc.n.u1Accessed \|\| PteDst.n.u1Present)
3892	{
3893	AssertMsgFailed(("PGM_PTFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PteSrc=%#RX64 PteDst=%#RX64\n",
3894	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3895	cErrors++;
3896	continue;
3897	}
3898	if (!PteDst.n.u1Accessed)
3899	{
3900	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PteSrc=%#RX64 PteDst=%#RX64\n",
3901	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3902	cErrors++;
3903	}
3904	fIgnoreFlags \|= X86_PTE_P;
3905	}
3906	# ifdef DEBUG_sandervl
3907	fIgnoreFlags \|= X86_PTE_D \| X86_PTE_A;
3908	# endif
3909	}
3910
3911	if ( (PteSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
3912	&& (PteSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags)
3913	)
3914	{
3915	AssertMsgFailed(("Flags mismatch at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PteSrc=%#RX64 PteDst=%#RX64\n",
3916	GCPtr + off, (uint64_t)PteSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
3917	fIgnoreFlags, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3918	cErrors++;
3919	continue;
3920	}
3921	} /* foreach PTE */
3922	}
3923	else
3924	{
3925	/*
3926	* Big Page.
3927	*/
3928	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;
3929	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
3930	{
3931	if (PdeDst.n.u1Write)
3932	{
3933	AssertMsgFailed(("!DIRTY page at %RGv is writable! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3934	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3935	cErrors++;
3936	continue;
3937	}
3938	if (!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY))
3939	{
3940	AssertMsgFailed(("!DIRTY page at %RGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3941	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3942	cErrors++;
3943	continue;
3944	}
3945	# if 0 /** @todo sync access bit properly... */
3946	if (PdeDst.n.u1Accessed != PdeSrc.b.u1Accessed)
3947	{
3948	AssertMsgFailed(("!DIRTY page at %RGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
3949	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3950	cErrors++;
3951	}
3952	fIgnoreFlags \|= X86_PTE_RW;
3953	# else
3954	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
3955	# endif
3956	}
3957	else if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
3958	{
3959	/* access bit emulation (not implemented). */
3960	if (PdeSrc.b.u1Accessed \|\| PdeDst.n.u1Present)
3961	{
3962	AssertMsgFailed(("PGM_PDFLAGS_TRACK_DIRTY set at %RGv but no accessed bit emulation! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3963	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3964	cErrors++;
3965	continue;
3966	}
3967	if (!PdeDst.n.u1Accessed)
3968	{
3969	AssertMsgFailed(("!ACCESSED page at %RGv is has the accessed bit set! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3970	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3971	cErrors++;
3972	}
3973	fIgnoreFlags \|= X86_PTE_P;
3974	}
3975
3976	if ((PdeSrc.u & ~fIgnoreFlags) != (PdeDst.u & ~fIgnoreFlags))
3977	{
3978	AssertMsgFailed(("Flags mismatch (B) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PdeDst=%#RX64\n",
3979	GCPtr, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PdeDst.u & ~fIgnoreFlags,
3980	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3981	cErrors++;
3982	}
3983
3984	/* iterate the page table. */
3985	for (unsigned iPT = 0, off = 0;
3986	iPT < RT_ELEMENTS(pPTDst->a);
3987	iPT++, off += PAGE_SIZE, GCPhysGst += PAGE_SIZE)
3988	{
3989	const SHWPTE PteDst = pPTDst->a[iPT];
3990
3991	if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
3992	{
3993	AssertMsgFailed(("The PTE at %RGv emulating a 2/4M page is marked TRACK_DIRTY! PdeSrc=%#RX64 PteDst=%#RX64\n",
3994	GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3995	cErrors++;
3996	}
3997
3998	/* skip not-present entries. */
3999	if (!PteDst.n.u1Present) /** @todo deal with ALL handlers and CSAM !P pages! */
4000	continue;
4001
4002	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;
4003
4004	/* match the physical addresses */
4005	HCPhysShw = PteDst.u & X86_PTE_PAE_PG_MASK;
4006
4007	# ifdef IN_RING3
4008	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
4009	if (RT_FAILURE(rc))
4010	{
4011	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4012	{
4013	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
4014	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4015	cErrors++;
4016	}
4017	}
4018	else if (HCPhysShw != (HCPhys & X86_PTE_PAE_PG_MASK))
4019	{
4020	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp HCPhys=%RHp GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4021	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4022	cErrors++;
4023	continue;
4024	}
4025	# endif
4026	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
4027	if (!pPhysPage)
4028	{
4029	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
4030	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
4031	{
4032	AssertMsgFailed(("Cannot find guest physical address %RGp at %RGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
4033	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4034	cErrors++;
4035	continue;
4036	}
4037	# endif
4038	if (PteDst.n.u1Write)
4039	{
4040	AssertMsgFailed(("Invalid guest page at %RGv is writable! GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4041	GCPtr + off, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4042	cErrors++;
4043	}
4044	fIgnoreFlags \|= X86_PTE_RW;
4045	}
4046	else if (HCPhysShw != PGM_PAGE_GET_HCPHYS(pPhysPage))
4047	{
4048	AssertMsgFailed(("Out of sync (phys) at %RGv! HCPhysShw=%RHp pPhysPage=%R[pgmpage] GCPhysGst=%RGp PdeSrc=%#RX64 PteDst=%#RX64\n",
4049	GCPtr + off, HCPhysShw, pPhysPage, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4050	cErrors++;
4051	continue;
4052	}
4053
4054	/* flags */
4055	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
4056	{
4057	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
4058	{
4059	if (PGM_PAGE_GET_HNDL_PHYS_STATE(pPhysPage) != PGM_PAGE_HNDL_PHYS_STATE_DISABLED)
4060	{
4061	if (PteDst.n.u1Write)
4062	{
4063	AssertMsgFailed(("WRITE access flagged at %RGv but the page is writable! pPhysPage=%R[pgmpage] PdeSrc=%#RX64 PteDst=%#RX64\n",
4064	GCPtr + off, pPhysPage, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4065	cErrors++;
4066	continue;
4067	}
4068	fIgnoreFlags \|= X86_PTE_RW;
4069	}
4070	}
4071	else
4072	{
4073	if (PteDst.n.u1Present)
4074	{
4075	AssertMsgFailed(("ALL access flagged at %RGv but the page is present! pPhysPage=%R[pgmpage] PdeSrc=%#RX64 PteDst=%#RX64\n",
4076	GCPtr + off, pPhysPage, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4077	cErrors++;
4078	continue;
4079	}
4080	fIgnoreFlags \|= X86_PTE_P;
4081	}
4082	}
4083
4084	if ( (PdeSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
4085	&& (PdeSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags) /* lazy phys handler dereg. */
4086	)
4087	{
4088	AssertMsgFailed(("Flags mismatch (BT) at %RGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PteDst=%#RX64\n",
4089	GCPtr + off, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
4090	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
4091	cErrors++;
4092	continue;
4093	}
4094	} /* for each PTE */
4095	}
4096	}
4097	/* not present */
4098
4099	} /* for each PDE */
4100
4101	} /* for each PDPTE */
4102
4103	} /* for each PML4E */
4104
4105	# ifdef DEBUG
4106	if (cErrors)
4107	LogFlow(("AssertCR3: cErrors=%d\n", cErrors));
4108	# endif
4109
4110	#endif /* GST == 32BIT, PAE or AMD64 */
4111	return cErrors;
4112
4113	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED && PGM_SHW_TYPE != PGM_TYPE_EPT */
4114	}
4115	#endif /* VBOX_STRICT */
4116
4117
4118	/**
4119	* Sets up the CR3 for shadow paging
4120	*
4121	* @returns Strict VBox status code.
4122	* @retval VINF_SUCCESS.
4123	*
4124	* @param pVM VM handle.
4125	* @param GCPhysCR3 The physical address in the CR3 register.
4126	*/
4127	PGM_BTH_DECL(int, MapCR3)(PVM pVM, RTGCPHYS GCPhysCR3)
4128	{
4129	/* Update guest paging info. */
4130	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
4131	\|\| PGM_GST_TYPE == PGM_TYPE_PAE \
4132	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
4133
4134	LogFlow(("MapCR3: %RGp\n", GCPhysCR3));
4135
4136	/*
4137	* Map the page CR3 points at.
4138	*/
4139	RTHCPTR HCPtrGuestCR3;
4140	RTHCPHYS HCPhysGuestCR3;
4141	# ifdef VBOX_WITH_NEW_PHYS_CODE
4142	pgmLock(pVM);
4143	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, GCPhysCR3);
4144	AssertReturn(pPage, VERR_INTERNAL_ERROR);
4145	HCPhysGuestCR3 = PGM_PAGE_GET_HCPHYS(pPage);
4146	/** @todo this needs some reworking wrt. locking. */
4147	# if defined(IN_RC) \|\| defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0)
4148	HCPtrGuestCR3 = NIL_RTHCPTR;
4149	int rc = VINF_SUCCESS;
4150	# else
4151	int rc = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, GCPhysCR3 & GST_CR3_PAGE_MASK, (void **)&HCPtrGuestCR3);
4152	# endif
4153	pgmUnlock(pVM);
4154	# else /* !VBOX_WITH_NEW_PHYS_CODE */
4155	int rc = pgmRamGCPhys2HCPtrAndHCPhys(&pVM->pgm.s, GCPhysCR3 & GST_CR3_PAGE_MASK, &HCPtrGuestCR3, &HCPhysGuestCR3);
4156	# endif /* !VBOX_WITH_NEW_PHYS_CODE */
4157	if (RT_SUCCESS(rc))
4158	{
4159	rc = PGMMap(pVM, (RTGCPTR)pVM->pgm.s.GCPtrCR3Mapping, HCPhysGuestCR3, PAGE_SIZE, 0);
4160	if (RT_SUCCESS(rc))
4161	{
4162	# ifdef IN_RC
4163	PGM_INVL_PG(pVM->pgm.s.GCPtrCR3Mapping);
4164	# endif
4165	# if PGM_GST_TYPE == PGM_TYPE_32BIT
4166	pVM->pgm.s.pGst32BitPdR3 = (R3PTRTYPE(PX86PD))HCPtrGuestCR3;
4167	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4168	pVM->pgm.s.pGst32BitPdR0 = (R0PTRTYPE(PX86PD))HCPtrGuestCR3;
4169	# endif
4170	pVM->pgm.s.pGst32BitPdRC = (RCPTRTYPE(PX86PD))pVM->pgm.s.GCPtrCR3Mapping;
4171
4172	# elif PGM_GST_TYPE == PGM_TYPE_PAE
4173	unsigned off = GCPhysCR3 & GST_CR3_PAGE_MASK & PAGE_OFFSET_MASK;
4174	pVM->pgm.s.pGstPaePdptR3 = (R3PTRTYPE(PX86PDPT))HCPtrGuestCR3;
4175	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4176	pVM->pgm.s.pGstPaePdptR0 = (R0PTRTYPE(PX86PDPT))HCPtrGuestCR3;
4177	# endif
4178	pVM->pgm.s.pGstPaePdptRC = (RCPTRTYPE(PX86PDPT))((RCPTRTYPE(uint8_t *))pVM->pgm.s.GCPtrCR3Mapping + off);
4179	Log(("Cached mapping %RRv\n", pVM->pgm.s.pGstPaePdptRC));
4180
4181	/*
4182	* Map the 4 PDs too.
4183	*/
4184	PX86PDPT pGuestPDPT = pgmGstGetPaePDPTPtr(&pVM->pgm.s);
4185	RTGCPTR GCPtr = pVM->pgm.s.GCPtrCR3Mapping + PAGE_SIZE;
4186	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++, GCPtr += PAGE_SIZE)
4187	{
4188	if (pGuestPDPT->a[i].n.u1Present)
4189	{
4190	RTHCPTR HCPtr;
4191	RTHCPHYS HCPhys;
4192	RTGCPHYS GCPhys = pGuestPDPT->a[i].u & X86_PDPE_PG_MASK;
4193	# ifdef VBOX_WITH_NEW_PHYS_CODE
4194	pgmLock(pVM);
4195	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, GCPhys);
4196	AssertReturn(pPage, VERR_INTERNAL_ERROR);
4197	HCPhys = PGM_PAGE_GET_HCPHYS(pPage);
4198	# if defined(IN_RC) \|\| defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0)
4199	HCPtr = NIL_RTHCPTR;
4200	int rc2 = VINF_SUCCESS;
4201	# else
4202	int rc2 = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, GCPhys, (void **)&HCPtr);
4203	# endif
4204	pgmUnlock(pVM);
4205	# else /* !VBOX_WITH_NEW_PHYS_CODE */
4206	int rc2 = pgmRamGCPhys2HCPtrAndHCPhys(&pVM->pgm.s, GCPhys, &HCPtr, &HCPhys);
4207	# endif /* !VBOX_WITH_NEW_PHYS_CODE */
4208	if (RT_SUCCESS(rc2))
4209	{
4210	rc = PGMMap(pVM, GCPtr, HCPhys, PAGE_SIZE, 0);
4211	AssertRCReturn(rc, rc);
4212
4213	pVM->pgm.s.apGstPaePDsR3[i] = (R3PTRTYPE(PX86PDPAE))HCPtr;
4214	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4215	pVM->pgm.s.apGstPaePDsR0[i] = (R0PTRTYPE(PX86PDPAE))HCPtr;
4216	# endif
4217	pVM->pgm.s.apGstPaePDsRC[i] = (RCPTRTYPE(PX86PDPAE))GCPtr;
4218	pVM->pgm.s.aGCPhysGstPaePDs[i] = GCPhys;
4219	PGM_INVL_PG(GCPtr); /** @todo This ends up calling HWACCMInvalidatePage, is that correct? */
4220	continue;
4221	}
4222	AssertMsgFailed(("pgmR3Gst32BitMapCR3: rc2=%d GCPhys=%RGp i=%d\n", rc2, GCPhys, i));
4223	}
4224
4225	pVM->pgm.s.apGstPaePDsR3[i] = 0;
4226	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4227	pVM->pgm.s.apGstPaePDsR0[i] = 0;
4228	# endif
4229	pVM->pgm.s.apGstPaePDsRC[i] = 0;
4230	pVM->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
4231	PGM_INVL_PG(GCPtr); /** @todo this shouldn't be necessary? */
4232	}
4233
4234	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
4235	pVM->pgm.s.pGstAmd64Pml4R3 = (R3PTRTYPE(PX86PML4))HCPtrGuestCR3;
4236	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4237	pVM->pgm.s.pGstAmd64Pml4R0 = (R0PTRTYPE(PX86PML4))HCPtrGuestCR3;
4238	# endif
4239	# endif
4240	}
4241	else
4242	AssertMsgFailed(("rc=%Rrc GCPhysGuestPD=%RGp\n", rc, GCPhysCR3));
4243	}
4244	else
4245	AssertMsgFailed(("rc=%Rrc GCPhysGuestPD=%RGp\n", rc, GCPhysCR3));
4246
4247	#else /* prot/real stub */
4248	int rc = VINF_SUCCESS;
4249	#endif
4250
4251	/* Update shadow paging info for guest modes with paging (32, pae, 64). */
4252	# if ( ( PGM_SHW_TYPE == PGM_TYPE_32BIT \
4253	\|\| PGM_SHW_TYPE == PGM_TYPE_PAE \
4254	\|\| PGM_SHW_TYPE == PGM_TYPE_AMD64) \
4255	&& ( PGM_GST_TYPE != PGM_TYPE_REAL \
4256	&& PGM_GST_TYPE != PGM_TYPE_PROT))
4257
4258	Assert(!HWACCMIsNestedPagingActive(pVM));
4259
4260	/*
4261	* Update the shadow root page as well since that's not fixed.
4262	*/
4263	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4264	PPGMPOOLPAGE pOldShwPageCR3 = pVM->pgm.s.CTX_SUFF(pShwPageCR3);
4265	uint32_t iOldShwUserTable = pVM->pgm.s.iShwUserTable;
4266	uint32_t iOldShwUser = pVM->pgm.s.iShwUser;
4267	PPGMPOOLPAGE pNewShwPageCR3;
4268
4269	Assert(!(GCPhysCR3 >> (PAGE_SHIFT + 32)));
4270	rc = pgmPoolAlloc(pVM, GCPhysCR3 & GST_CR3_PAGE_MASK, BTH_PGMPOOLKIND_ROOT, SHW_POOL_ROOT_IDX, GCPhysCR3 >> PAGE_SHIFT, &pNewShwPageCR3);
4271	AssertFatalRC(rc);
4272	rc = VINF_SUCCESS;
4273
4274	/* Mark the page as locked; disallow flushing. */
4275	pgmPoolLockPage(pPool, pNewShwPageCR3);
4276
4277	# ifdef IN_RC
4278	/* NOTE: We can't deal with jumps to ring 3 here as we're now in an inconsistent state! */
4279	bool fLog = VMMGCLogDisable(pVM);
4280	# endif
4281
4282	pVM->pgm.s.iShwUser = SHW_POOL_ROOT_IDX;
4283	pVM->pgm.s.iShwUserTable = GCPhysCR3 >> PAGE_SHIFT;
4284	pVM->pgm.s.CTX_SUFF(pShwPageCR3) = pNewShwPageCR3;
4285	# ifdef IN_RING0
4286	pVM->pgm.s.pShwPageCR3R3 = MMHyperCCToR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4287	pVM->pgm.s.pShwPageCR3RC = MMHyperCCToRC(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4288	# elif defined(IN_RC)
4289	pVM->pgm.s.pShwPageCR3R3 = MMHyperCCToR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4290	pVM->pgm.s.pShwPageCR3R0 = MMHyperCCToR0(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4291	# else
4292	pVM->pgm.s.pShwPageCR3R0 = MMHyperCCToR0(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4293	pVM->pgm.s.pShwPageCR3RC = MMHyperCCToRC(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4294	# endif
4295
4296	# ifndef PGM_WITHOUT_MAPPINGS
4297	/*
4298	* Apply all hypervisor mappings to the new CR3.
4299	* Note that SyncCR3 will be executed in case CR3 is changed in a guest paging mode; this will
4300	* make sure we check for conflicts in the new CR3 root.
4301	*/
4302	# if PGM_WITH_PAGING(PGM_GST_TYPE, PGM_SHW_TYPE)
4303	Assert(VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3_NON_GLOBAL) \|\| VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3));
4304	# endif
4305	rc = pgmMapActivateCR3(pVM, pNewShwPageCR3);
4306	AssertRCReturn(rc, rc);
4307	# endif
4308
4309	/* Set the current hypervisor CR3. */
4310	CPUMSetHyperCR3(pVM, PGMGetHyperCR3(pVM));
4311	SELMShadowCR3Changed(pVM);
4312
4313	# ifdef IN_RC
4314	VMMGCLogRestore(pVM, fLog);
4315	# endif
4316
4317	/* Clean up the old CR3 root. */
4318	if (pOldShwPageCR3)
4319	{
4320	Assert(pOldShwPageCR3->enmKind != PGMPOOLKIND_FREE);
4321	# ifndef PGM_WITHOUT_MAPPINGS
4322	/* Remove the hypervisor mappings from the shadow page table. */
4323	pgmMapDeactivateCR3(pVM, pOldShwPageCR3);
4324	# endif
4325	/* Mark the page as unlocked; allow flushing again. */
4326	pgmPoolUnlockPage(pPool, pOldShwPageCR3);
4327
4328	pgmPoolFreeByPage(pPool, pOldShwPageCR3, iOldShwUser, iOldShwUserTable);
4329	}
4330
4331	# endif
4332
4333	return rc;
4334	}
4335
4336	/**
4337	* Unmaps the shadow CR3.
4338	*
4339	* @returns VBox status, no specials.
4340	* @param pVM VM handle.
4341	*/
4342	PGM_BTH_DECL(int, UnmapCR3)(PVM pVM)
4343	{
4344	LogFlow(("UnmapCR3\n"));
4345
4346	int rc = VINF_SUCCESS;
4347
4348	/* Update guest paging info. */
4349	#if PGM_GST_TYPE == PGM_TYPE_32BIT
4350	pVM->pgm.s.pGst32BitPdR3 = 0;
4351	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4352	pVM->pgm.s.pGst32BitPdR0 = 0;
4353	#endif
4354	pVM->pgm.s.pGst32BitPdRC = 0;
4355
4356	#elif PGM_GST_TYPE == PGM_TYPE_PAE
4357	pVM->pgm.s.pGstPaePdptR3 = 0;
4358	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4359	pVM->pgm.s.pGstPaePdptR0 = 0;
4360	# endif
4361	pVM->pgm.s.pGstPaePdptRC = 0;
4362	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++)
4363	{
4364	pVM->pgm.s.apGstPaePDsR3[i] = 0;
4365	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4366	pVM->pgm.s.apGstPaePDsR0[i] = 0;
4367	# endif
4368	pVM->pgm.s.apGstPaePDsRC[i] = 0;
4369	pVM->pgm.s.aGCPhysGstPaePDs[i] = NIL_RTGCPHYS;
4370	}
4371
4372	#elif PGM_GST_TYPE == PGM_TYPE_AMD64
4373	pVM->pgm.s.pGstAmd64Pml4R3 = 0;
4374	# ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4375	pVM->pgm.s.pGstAmd64Pml4R0 = 0;
4376	# endif
4377
4378	#else /* prot/real mode stub */
4379	/* nothing to do */
4380	#endif
4381
4382	#if !defined(IN_RC) /* In RC we rely on MapCR3 to do the shadow part for us at a safe time */
4383	/* Update shadow paging info. */
4384	# if ( ( PGM_SHW_TYPE == PGM_TYPE_32BIT \
4385	\|\| PGM_SHW_TYPE == PGM_TYPE_PAE \
4386	\|\| PGM_SHW_TYPE == PGM_TYPE_AMD64))
4387
4388	# if PGM_GST_TYPE != PGM_TYPE_REAL
4389	Assert(!HWACCMIsNestedPagingActive(pVM));
4390	# endif
4391
4392	# ifndef PGM_WITHOUT_MAPPINGS
4393	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3))
4394	/* Remove the hypervisor mappings from the shadow page table. */
4395	pgmMapDeactivateCR3(pVM, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4396	# endif
4397
4398	if (pVM->pgm.s.CTX_SUFF(pShwPageCR3))
4399	{
4400	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4401
4402	Assert(pVM->pgm.s.iShwUser != PGMPOOL_IDX_NESTED_ROOT);
4403
4404	/* Mark the page as unlocked; allow flushing again. */
4405	pgmPoolUnlockPage(pPool, pVM->pgm.s.CTX_SUFF(pShwPageCR3));
4406
4407	pgmPoolFreeByPage(pPool, pVM->pgm.s.CTX_SUFF(pShwPageCR3), pVM->pgm.s.iShwUser, pVM->pgm.s.iShwUserTable);
4408	pVM->pgm.s.pShwPageCR3R3 = 0;
4409	pVM->pgm.s.pShwPageCR3R0 = 0;
4410	pVM->pgm.s.pShwPageCR3RC = 0;
4411	pVM->pgm.s.iShwUser = 0;
4412	pVM->pgm.s.iShwUserTable = 0;
4413	}
4414	# endif
4415	#endif /* !IN_RC*/
4416
4417	return rc;
4418	}
4419

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

source: vbox/trunk/src/VBox/VMM/VMMAll/PGMAllBth.h@ 18291

Download in other formats: