PGMAllBth.h@ 9344

Last change on this file since 9344 was 9344, checked in by vboxsync, 17 years ago
Types and string format specifiers.
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 163.8 KB

Line
1	/* $Id: PGMAllBth.h 9344 2008-06-03 09:49:14Z vboxsync $ */
2	/** @file
3	* VBox - Page Manager, Shadow+Guest Paging Template - All context code.
4	*
5	* This file is a big challenge!
6	*/
7
8	/*
9	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
10	*
11	* This file is part of VirtualBox Open Source Edition (OSE), as
12	* available from http://www.virtualbox.org. This file is free software;
13	* you can redistribute it and/or modify it under the terms of the GNU
14	* General Public License (GPL) as published by the Free Software
15	* Foundation, in version 2 as it comes in the "COPYING" file of the
16	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
17	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
18	*
19	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
20	* Clara, CA 95054 USA or visit http://www.sun.com if you need
21	* additional information or have any questions.
22	*/
23
24	/*******************************************************************************
25	* Internal Functions *
26	*******************************************************************************/
27	__BEGIN_DECLS
28	PGM_BTH_DECL(int, Trap0eHandler)(PVM pVM, RTGCUINT uErr, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault);
29	PGM_BTH_DECL(int, InvalidatePage)(PVM pVM, RTGCUINTPTR GCPtrPage);
30	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCUINTPTR GCPtrPage, unsigned cPages, unsigned uErr);
31	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCUINTPTR GCPtrPage);
32	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPD, PGSTPD pPDSrc, RTGCUINTPTR GCPtrPage);
33	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCUINTPTR Addr, unsigned fPage, unsigned uErr);
34	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCUINTPTR GCPtrPage);
35	PGM_BTH_DECL(int, SyncCR3)(PVM pVM, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal);
36	#ifdef VBOX_STRICT
37	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCUINTPTR GCPtr = 0, RTGCUINTPTR cb = ~(RTGCUINTPTR)0);
38	#endif
39	#ifdef PGMPOOL_WITH_USER_TRACKING
40	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys);
41	#endif
42	__END_DECLS
43
44
45	/* Filter out some illegal combinations of guest and shadow paging, so we can remove redundant checks inside functions. */
46	#if PGM_GST_TYPE == PGM_TYPE_PAE && PGM_SHW_TYPE != PGM_TYPE_PAE && PGM_SHW_TYPE != PGM_TYPE_NESTED
47	# error "Invalid combination; PAE guest implies PAE shadow"
48	#endif
49
50	#if (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
51	&& !(PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_AMD64 \|\| PGM_SHW_TYPE == PGM_TYPE_NESTED)
52	# error "Invalid combination; real or protected mode without paging implies 32 bits or PAE shadow paging."
53	#endif
54
55	#if (PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
56	&& !(PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE \|\| PGM_SHW_TYPE == PGM_TYPE_NESTED)
57	# error "Invalid combination; 32 bits guest paging or PAE implies 32 bits or PAE shadow paging."
58	#endif
59
60	#if (PGM_GST_TYPE == PGM_TYPE_AMD64 && PGM_SHW_TYPE != PGM_TYPE_AMD64 && PGM_SHW_TYPE != PGM_TYPE_NESTED) \
61	\|\| (PGM_SHW_TYPE == PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_AMD64 && PGM_GST_TYPE != PGM_TYPE_PROT)
62	# error "Invalid combination; AMD64 guest implies AMD64 shadow and vice versa"
63	#endif
64
65	#ifdef IN_RING0 /* no mappings in VT-x and AMD-V mode */
66	# define PGM_WITHOUT_MAPPINGS
67	#endif
68
69	/**
70	* #PF Handler for raw-mode guest execution.
71	*
72	* @returns VBox status code (appropriate for trap handling and GC return).
73	* @param pVM VM Handle.
74	* @param uErr The trap error code.
75	* @param pRegFrame Trap register frame.
76	* @param pvFault The fault address.
77	*/
78	PGM_BTH_DECL(int, Trap0eHandler)(PVM pVM, RTGCUINT uErr, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault)
79	{
80	#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) \
81	&& PGM_SHW_TYPE != PGM_TYPE_NESTED
82
83	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE != PGM_TYPE_PAE
84	/*
85	* Hide the instruction fetch trap indicator for now.
86	*/
87	/** @todo NXE will change this and we must fix NXE in the switcher too! */
88	if (uErr & X86_TRAP_PF_ID)
89	{
90	uErr &= ~X86_TRAP_PF_ID;
91	TRPMSetErrorCode(pVM, uErr);
92	}
93	# endif
94
95	/*
96	* Get PDs.
97	*/
98	int rc;
99	# if PGM_WITH_PAGING(PGM_GST_TYPE)
100	# if PGM_GST_TYPE == PGM_TYPE_32BIT
101	const unsigned iPDSrc = (RTGCUINTPTR)pvFault >> GST_PD_SHIFT;
102	PGSTPD pPDSrc = CTXSUFF(pVM->pgm.s.pGuestPD);
103
104	# elif PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
105
106	# if PGM_GST_TYPE == PGM_TYPE_PAE
107	unsigned iPDSrc;
108	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, (RTGCUINTPTR)pvFault, &iPDSrc);
109
110	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
111	unsigned iPDSrc;
112	PX86PML4E pPml4e;
113	X86PDPE Pdpe;
114	PGSTPD pPDSrc;
115
116	pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, pvFault, &pPml4e, &Pdpe, &iPDSrc);
117	Assert(pPml4e);
118	# endif
119	/* Quick check for a valid guest trap. */
120	if (!pPDSrc)
121	{
122	LogFlow(("Trap0eHandler: guest PDPTR not present CR3=%VGp\n", (CPUMGetGuestCR3(pVM) & X86_CR3_PAGE_MASK)));
123	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eGuestTrap; });
124	TRPMSetErrorCode(pVM, uErr);
125	return VINF_EM_RAW_GUEST_TRAP;
126	}
127	# endif
128	# else
129	PGSTPD pPDSrc = NULL;
130	const unsigned iPDSrc = 0;
131	# endif
132
133	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
134	const unsigned iPDDst = (RTGCUINTPTR)pvFault >> SHW_PD_SHIFT;
135	PX86PD pPDDst = pVM->pgm.s.CTXMID(p,32BitPD);
136	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
137	const unsigned iPDDst = (RTGCUINTPTR)pvFault >> SHW_PD_SHIFT;
138	PX86PDPAE pPDDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]; /* We treat this as a PD with 2048 entries, so no need to and with SHW_PD_MASK to get iPDDst */
139
140	# if PGM_GST_TYPE == PGM_TYPE_PAE
141	/* Did we mark the PDPT as not present in SyncCR3? */
142	unsigned iPDPTE = ((RTGCUINTPTR)pvFault >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
143	if (!pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPDPTE].n.u1Present)
144	pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPDPTE].n.u1Present = 1;
145
146	# endif
147
148	# elif PGM_SHW_TYPE == PGM_TYPE_AMD64
149	const unsigned iPDDst = (((RTGCUINTPTR)pvFault >> SHW_PD_SHIFT) & SHW_PD_MASK);
150	PX86PDPAE pPDDst;
151
152	rc = PGMShwGetLongModePDPtr(pVM, (RTGCUINTPTR)pvFault, &pPDDst);
153	if (rc != VINF_SUCCESS)
154	{
155	AssertMsg(rc == VINF_PGM_SYNC_CR3, ("Unexpected rc=%Vrc\n", rc));
156	return rc;
157	}
158	Assert(pPDDst);
159	# endif
160
161	# if PGM_WITH_PAGING(PGM_GST_TYPE)
162	/*
163	* If we successfully correct the write protection fault due to dirty bit
164	* tracking, or this page fault is a genuine one, then return immediately.
165	*/
166	STAM_PROFILE_START(&pVM->pgm.s.StatCheckPageFault, e);
167	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, &pPDDst->a[iPDDst], &pPDSrc->a[iPDSrc], (RTGCUINTPTR)pvFault);
168	STAM_PROFILE_STOP(&pVM->pgm.s.StatCheckPageFault, e);
169	if ( rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT
170	\|\| rc == VINF_EM_RAW_GUEST_TRAP)
171	{
172	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution)
173	= rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? &pVM->pgm.s.StatTrap0eDirtyAndAccessedBits : &pVM->pgm.s.StatTrap0eGuestTrap; });
174	LogBird(("Trap0eHandler: returns %s\n", rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? "VINF_SUCCESS" : "VINF_EM_RAW_GUEST_TRAP"));
175	return rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT ? VINF_SUCCESS : rc;
176	}
177
178	STAM_COUNTER_INC(&pVM->pgm.s.StatGCTrap0ePD[iPDSrc]);
179	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
180
181	/*
182	* A common case is the not-present error caused by lazy page table syncing.
183	*
184	* It is IMPORTANT that we weed out any access to non-present shadow PDEs here
185	* so we can safely assume that the shadow PT is present when calling SyncPage later.
186	*
187	* On failure, we ASSUME that SyncPT is out of memory or detected some kind
188	* of mapping conflict and defer to SyncCR3 in R3.
189	* (Again, we do NOT support access handlers for non-present guest pages.)
190	*
191	*/
192	# if PGM_WITH_PAGING(PGM_GST_TYPE)
193	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
194	# else
195	GSTPDE PdeSrc;
196	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
197	PdeSrc.n.u1Present = 1;
198	PdeSrc.n.u1Write = 1;
199	PdeSrc.n.u1Accessed = 1;
200	PdeSrc.n.u1User = 1;
201	# endif
202	if ( !(uErr & X86_TRAP_PF_P) /* not set means page not present instead of page protection violation */
203	&& !pPDDst->a[iPDDst].n.u1Present
204	&& PdeSrc.n.u1Present
205	)
206
207	{
208	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eSyncPT; });
209	STAM_PROFILE_START(&pVM->pgm.s.StatLazySyncPT, f);
210	LogFlow(("=>SyncPT %04x = %08x\n", iPDSrc, PdeSrc.au32[0]));
211	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, (RTGCUINTPTR)pvFault);
212	if (VBOX_SUCCESS(rc))
213	{
214	STAM_PROFILE_STOP(&pVM->pgm.s.StatLazySyncPT, f);
215	return rc;
216	}
217	Log(("SyncPT: %d failed!! rc=%d\n", iPDSrc, rc));
218	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync, right? */
219	STAM_PROFILE_STOP(&pVM->pgm.s.StatLazySyncPT, f);
220	return VINF_PGM_SYNC_CR3;
221	}
222
223	# if PGM_WITH_PAGING(PGM_GST_TYPE)
224	/*
225	* Check if this address is within any of our mappings.
226	*
227	* This is very fast and it's gonna save us a bit of effort below and prevent
228	* us from screwing ourself with MMIO2 pages which have a GC Mapping (VRam).
229	* (BTW, it's impossible to have physical access handlers in a mapping.)
230	*/
231	if (pgmMapAreMappingsEnabled(&pVM->pgm.s))
232	{
233	STAM_PROFILE_START(&pVM->pgm.s.StatMapping, a);
234	PPGMMAPPING pMapping = CTXALLSUFF(pVM->pgm.s.pMappings);
235	for ( ; pMapping; pMapping = CTXALLSUFF(pMapping->pNext))
236	{
237	if ((RTGCUINTPTR)pvFault < (RTGCUINTPTR)pMapping->GCPtr)
238	break;
239	if ((RTGCUINTPTR)pvFault - (RTGCUINTPTR)pMapping->GCPtr < pMapping->cb)
240	{
241	/*
242	* The first thing we check is if we've got an undetected conflict.
243	*/
244	if (!pVM->pgm.s.fMappingsFixed)
245	{
246	unsigned iPT = pMapping->cb >> GST_PD_SHIFT;
247	while (iPT-- > 0)
248	if (pPDSrc->a[iPDSrc + iPT].n.u1Present)
249	{
250	STAM_COUNTER_INC(&pVM->pgm.s.StatGCTrap0eConflicts);
251	Log(("Trap0e: Detected Conflict %VGv-%VGv\n", pMapping->GCPtr, pMapping->GCPtrLast));
252	VM_FF_SET(pVM, VM_FF_PGM_SYNC_CR3); /** @todo no need to do global sync,right? */
253	STAM_PROFILE_STOP(&pVM->pgm.s.StatMapping, a);
254	return VINF_PGM_SYNC_CR3;
255	}
256	}
257
258	/*
259	* Check if the fault address is in a virtual page access handler range.
260	*/
261	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&CTXSUFF(pVM->pgm.s.pTrees)->HyperVirtHandlers, pvFault);
262	if ( pCur
263	&& (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->GCPtr < pCur->cb
264	&& uErr & X86_TRAP_PF_RW)
265	{
266	# ifdef IN_GC
267	STAM_PROFILE_START(&pCur->Stat, h);
268	rc = CTXSUFF(pCur->pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->GCPtr, (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->GCPtr);
269	STAM_PROFILE_STOP(&pCur->Stat, h);
270	# else
271	AssertFailed();
272	rc = VINF_EM_RAW_EMULATE_INSTR; /* can't happen with VMX */
273	# endif
274	STAM_COUNTER_INC(&pVM->pgm.s.StatTrap0eMapHandler);
275	STAM_PROFILE_STOP(&pVM->pgm.s.StatMapping, a);
276	return rc;
277	}
278
279	/*
280	* Pretend we're not here and let the guest handle the trap.
281	*/
282	TRPMSetErrorCode(pVM, uErr & ~X86_TRAP_PF_P);
283	STAM_COUNTER_INC(&pVM->pgm.s.StatGCTrap0eMap);
284	LogFlow(("PGM: Mapping access -> route trap to recompiler!\n"));
285	STAM_PROFILE_STOP(&pVM->pgm.s.StatMapping, a);
286	return VINF_EM_RAW_GUEST_TRAP;
287	}
288	}
289	STAM_PROFILE_STOP(&pVM->pgm.s.StatMapping, a);
290	} /* pgmAreMappingsEnabled(&pVM->pgm.s) */
291	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
292
293	/*
294	* Check if this fault address is flagged for special treatment,
295	* which means we'll have to figure out the physical address and
296	* check flags associated with it.
297	*
298	* ASSUME that we can limit any special access handling to pages
299	* in page tables which the guest believes to be present.
300	*/
301	if (PdeSrc.n.u1Present)
302	{
303	RTGCPHYS GCPhys = NIL_RTGCPHYS;
304
305	# if PGM_WITH_PAGING(PGM_GST_TYPE)
306	uint32_t cr4 = CPUMGetGuestCR4(pVM);
307	if ( PdeSrc.b.u1Size
308	&& (cr4 & X86_CR4_PSE))
309	GCPhys = (PdeSrc.u & GST_PDE_BIG_PG_MASK)
310	\| ((RTGCPHYS)pvFault & (GST_BIG_PAGE_OFFSET_MASK ^ PAGE_OFFSET_MASK));
311	else
312	{
313	PGSTPT pPTSrc;
314	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
315	if (VBOX_SUCCESS(rc))
316	{
317	unsigned iPTESrc = ((RTGCUINTPTR)pvFault >> GST_PT_SHIFT) & GST_PT_MASK;
318	if (pPTSrc->a[iPTESrc].n.u1Present)
319	GCPhys = pPTSrc->a[iPTESrc].u & GST_PTE_PG_MASK;
320	}
321	}
322	# else
323	/* No paging so the fault address is the physical address */
324	GCPhys = (RTGCPHYS)((RTGCUINTPTR)pvFault & ~PAGE_OFFSET_MASK);
325	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
326
327	/*
328	* If we have a GC address we'll check if it has any flags set.
329	*/
330	if (GCPhys != NIL_RTGCPHYS)
331	{
332	STAM_PROFILE_START(&pVM->pgm.s.StatHandlers, b);
333
334	PPGMPAGE pPage;
335	rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
336	if (VBOX_SUCCESS(rc))
337	{
338	if (PGM_PAGE_HAS_ANY_HANDLERS(pPage))
339	{
340	if (PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage))
341	{
342	/*
343	* Physical page access handler.
344	*/
345	const RTGCPHYS GCPhysFault = GCPhys \| ((RTGCUINTPTR)pvFault & PAGE_OFFSET_MASK);
346	PPGMPHYSHANDLER pCur = (PPGMPHYSHANDLER)RTAvlroGCPhysRangeGet(&CTXSUFF(pVM->pgm.s.pTrees)->PhysHandlers, GCPhysFault);
347	if (pCur)
348	{
349	# ifdef PGM_SYNC_N_PAGES
350	/*
351	* If the region is write protected and we got a page not present fault, then sync
352	* the pages. If the fault was caused by a read, then restart the instruction.
353	* In case of write access continue to the GC write handler.
354	*
355	* ASSUMES that there is only one handler per page or that they have similar write properties.
356	*/
357	if ( pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
358	&& !(uErr & X86_TRAP_PF_P))
359	{
360	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, PGM_SYNC_NR_PAGES, uErr);
361	if ( VBOX_FAILURE(rc)
362	\|\| !(uErr & X86_TRAP_PF_RW)
363	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE)
364	{
365	AssertRC(rc);
366	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersOutOfSync);
367	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
368	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eOutOfSyncHndPhys; });
369	return rc;
370	}
371	}
372	# endif
373
374	AssertMsg( pCur->enmType != PGMPHYSHANDLERTYPE_PHYSICAL_WRITE
375	\|\| (pCur->enmType == PGMPHYSHANDLERTYPE_PHYSICAL_WRITE && (uErr & X86_TRAP_PF_RW)),
376	("Unexpected trap for physical handler: %08X (phys=%08x) HCPhys=%X uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
377
378	#if defined(IN_GC) \|\| defined(IN_RING0)
379	if (CTXALLSUFF(pCur->pfnHandler))
380	{
381	STAM_PROFILE_START(&pCur->Stat, h);
382	rc = pCur->CTXALLSUFF(pfnHandler)(pVM, uErr, pRegFrame, pvFault, GCPhysFault, CTXALLSUFF(pCur->pvUser));
383	STAM_PROFILE_STOP(&pCur->Stat, h);
384	}
385	else
386	#endif
387	rc = VINF_EM_RAW_EMULATE_INSTR;
388	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersPhysical);
389	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
390	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eHndPhys; });
391	return rc;
392	}
393	}
394	# if PGM_WITH_PAGING(PGM_GST_TYPE)
395	else
396	{
397	# ifdef PGM_SYNC_N_PAGES
398	/*
399	* If the region is write protected and we got a page not present fault, then sync
400	* the pages. If the fault was caused by a read, then restart the instruction.
401	* In case of write access continue to the GC write handler.
402	*/
403	if ( PGM_PAGE_GET_HNDL_VIRT_STATE(pPage) < PGM_PAGE_HNDL_PHYS_STATE_ALL
404	&& !(uErr & X86_TRAP_PF_P))
405	{
406	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, PGM_SYNC_NR_PAGES, uErr);
407	if ( VBOX_FAILURE(rc)
408	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
409	\|\| !(uErr & X86_TRAP_PF_RW))
410	{
411	AssertRC(rc);
412	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersOutOfSync);
413	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
414	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eOutOfSyncHndVirt; });
415	return rc;
416	}
417	}
418	# endif
419	/*
420	* Ok, it's an virtual page access handler.
421	*
422	* Since it's faster to search by address, we'll do that first
423	* and then retry by GCPhys if that fails.
424	*/
425	/** @todo r=bird: perhaps we should consider looking up by physical address directly now? */
426	/** @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
427	* page was changed without us noticing it (not-present -> present without invlpg or mov cr3, xxx)
428	*/
429	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&CTXSUFF(pVM->pgm.s.pTrees)->VirtHandlers, pvFault);
430	if (pCur)
431	{
432	AssertMsg(!((RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->GCPtr < pCur->cb)
433	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
434	\|\| !(uErr & X86_TRAP_PF_P)
435	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
436	("Unexpected trap for virtual handler: %VGv (phys=%VGp) HCPhys=%HGp uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
437
438	if ( (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->GCPtr < pCur->cb
439	&& ( uErr & X86_TRAP_PF_RW
440	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
441	{
442	# ifdef IN_GC
443	STAM_PROFILE_START(&pCur->Stat, h);
444	rc = CTXSUFF(pCur->pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->GCPtr, (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->GCPtr);
445	STAM_PROFILE_STOP(&pCur->Stat, h);
446	# else
447	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
448	# endif
449	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersVirtual);
450	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
451	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eHndVirt; });
452	return rc;
453	}
454	/* Unhandled part of a monitored page */
455	}
456	else
457	{
458	/* Check by physical address. */
459	PPGMVIRTHANDLER pCur;
460	unsigned iPage;
461	rc = pgmHandlerVirtualFindByPhysAddr(pVM, GCPhys + ((RTGCUINTPTR)pvFault & PAGE_OFFSET_MASK),
462	&pCur, &iPage);
463	Assert(VBOX_SUCCESS(rc) \|\| !pCur);
464	if ( pCur
465	&& ( uErr & X86_TRAP_PF_RW
466	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
467	{
468	Assert((pCur->aPhysToVirt[iPage].Core.Key & X86_PTE_PAE_PG_MASK) == GCPhys);
469	# ifdef IN_GC
470	RTGCUINTPTR off = (iPage << PAGE_SHIFT) + ((RTGCUINTPTR)pvFault & PAGE_OFFSET_MASK) - ((RTGCUINTPTR)pCur->GCPtr & PAGE_OFFSET_MASK);
471	Assert(off < pCur->cb);
472	STAM_PROFILE_START(&pCur->Stat, h);
473	rc = CTXSUFF(pCur->pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->GCPtr, off);
474	STAM_PROFILE_STOP(&pCur->Stat, h);
475	# else
476	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
477	# endif
478	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersVirtualByPhys);
479	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
480	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eHndVirt; });
481	return rc;
482	}
483	}
484	}
485	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
486
487	/*
488	* There is a handled area of the page, but this fault doesn't belong to it.
489	* We must emulate the instruction.
490	*
491	* To avoid crashing (non-fatal) in the interpreter and go back to the recompiler
492	* we first check if this was a page-not-present fault for a page with only
493	* write access handlers. Restart the instruction if it wasn't a write access.
494	*/
495	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersUnhandled);
496
497	if ( !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
498	&& !(uErr & X86_TRAP_PF_P))
499	{
500	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, PGM_SYNC_NR_PAGES, uErr);
501	if ( VBOX_FAILURE(rc)
502	\|\| rc == VINF_PGM_SYNCPAGE_MODIFIED_PDE
503	\|\| !(uErr & X86_TRAP_PF_RW))
504	{
505	AssertRC(rc);
506	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersOutOfSync);
507	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
508	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eOutOfSyncHndPhys; });
509	return rc;
510	}
511	}
512
513	/** @todo This particular case can cause quite a lot of overhead. E.g. early stage of kernel booting in Ubuntu 6.06
514	* It's writing to an unhandled part of the LDT page several million times.
515	*/
516	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
517	LogFlow(("PGM: PGMInterpretInstruction -> rc=%d HCPhys=%RHp%s%s\n",
518	rc, pPage->HCPhys,
519	PGM_PAGE_HAS_ANY_PHYSICAL_HANDLERS(pPage) ? " phys" : "",
520	PGM_PAGE_HAS_ANY_VIRTUAL_HANDLERS(pPage) ? " virt" : ""));
521	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
522	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eHndUnhandled; });
523	return rc;
524	} /* if any kind of handler */
525
526	# if PGM_WITH_PAGING(PGM_GST_TYPE)
527	if (uErr & X86_TRAP_PF_P)
528	{
529	/*
530	* The page isn't marked, but it might still be monitored by a virtual page access handler.
531	* (ASSUMES no temporary disabling of virtual handlers.)
532	*/
533	/** @todo r=bird: Since the purpose is to catch out of sync pages with virtual handler(s) here,
534	* we should correct both the shadow page table and physical memory flags, and not only check for
535	* accesses within the handler region but for access to pages with virtual handlers. */
536	PPGMVIRTHANDLER pCur = (PPGMVIRTHANDLER)RTAvlroGCPtrRangeGet(&CTXSUFF(pVM->pgm.s.pTrees)->VirtHandlers, pvFault);
537	if (pCur)
538	{
539	AssertMsg( !((RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->GCPtr < pCur->cb)
540	\|\| ( pCur->enmType != PGMVIRTHANDLERTYPE_WRITE
541	\|\| !(uErr & X86_TRAP_PF_P)
542	\|\| (pCur->enmType == PGMVIRTHANDLERTYPE_WRITE && (uErr & X86_TRAP_PF_RW))),
543	("Unexpected trap for virtual handler: %08X (phys=%08x) HCPhys=%X uErr=%X, enum=%d\n", pvFault, GCPhys, pPage->HCPhys, uErr, pCur->enmType));
544
545	if ( (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->GCPtr < pCur->cb
546	&& ( uErr & X86_TRAP_PF_RW
547	\|\| pCur->enmType != PGMVIRTHANDLERTYPE_WRITE ) )
548	{
549	# ifdef IN_GC
550	STAM_PROFILE_START(&pCur->Stat, h);
551	rc = CTXSUFF(pCur->pfnHandler)(pVM, uErr, pRegFrame, pvFault, pCur->GCPtr, (RTGCUINTPTR)pvFault - (RTGCUINTPTR)pCur->GCPtr);
552	STAM_PROFILE_STOP(&pCur->Stat, h);
553	# else
554	rc = VINF_EM_RAW_EMULATE_INSTR; /** @todo for VMX */
555	# endif
556	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersVirtualUnmarked);
557	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
558	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eHndVirt; });
559	return rc;
560	}
561	}
562	}
563	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
564	}
565	else
566	{
567	/* When the guest accesses invalid physical memory (e.g. probing of RAM or accessing a remapped MMIO range), then we'll fall
568	* back to the recompiler to emulate the instruction.
569	*/
570	LogFlow(("pgmPhysGetPageEx %VGp failed with %Vrc\n", GCPhys, rc));
571	STAM_COUNTER_INC(&pVM->pgm.s.StatHandlersInvalid);
572	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
573	return VINF_EM_RAW_EMULATE_INSTR;
574	}
575
576	STAM_PROFILE_STOP(&pVM->pgm.s.StatHandlers, b);
577
578	# ifdef PGM_OUT_OF_SYNC_IN_GC
579	/*
580	* We are here only if page is present in Guest page tables and trap is not handled
581	* by our handlers.
582	* Check it for page out-of-sync situation.
583	*/
584	STAM_PROFILE_START(&pVM->pgm.s.StatOutOfSync, c);
585
586	if (!(uErr & X86_TRAP_PF_P))
587	{
588	/*
589	* Page is not present in our page tables.
590	* Try to sync it!
591	* BTW, fPageShw is invalid in this branch!
592	*/
593	if (uErr & X86_TRAP_PF_US)
594	STAM_COUNTER_INC(&pVM->pgm.s.StatGCPageOutOfSyncUser);
595	else /* supervisor */
596	STAM_COUNTER_INC(&pVM->pgm.s.StatGCPageOutOfSyncSupervisor);
597
598	# if defined(LOG_ENABLED) && !defined(IN_RING0)
599	RTGCPHYS GCPhys;
600	uint64_t fPageGst;
601	PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
602	Log(("Page out of sync: %VGv eip=%08x PdeSrc.n.u1User=%d fPageGst=%08llx GCPhys=%VGp scan=%d\n",
603	pvFault, pRegFrame->eip, PdeSrc.n.u1User, fPageGst, GCPhys, CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)));
604	# endif /* LOG_ENABLED */
605
606	# if PGM_WITH_PAGING(PGM_GST_TYPE) && !defined(IN_RING0)
607	if (CPUMGetGuestCPL(pVM, pRegFrame) == 0)
608	{
609	uint64_t fPageGst;
610	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
611	if ( VBOX_SUCCESS(rc)
612	&& !(fPageGst & X86_PTE_US))
613	{
614	/* Note: can't check for X86_TRAP_ID bit, because that requires execute disable support on the CPU */
615	if ( pvFault == (RTGCPTR)pRegFrame->eip
616	\|\| (RTGCUINTPTR)pvFault - pRegFrame->eip < 8 /* instruction crossing a page boundary */
617	# ifdef CSAM_DETECT_NEW_CODE_PAGES
618	\|\| ( !PATMIsPatchGCAddr(pVM, (RTGCPTR)pRegFrame->eip)
619	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)pRegFrame->eip)) /* any new code we encounter here */
620	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
621	)
622	{
623	LogFlow(("CSAMExecFault %VGv\n", pRegFrame->eip));
624	rc = CSAMExecFault(pVM, (RTRCPTR)pRegFrame->eip);
625	if (rc != VINF_SUCCESS)
626	{
627	/*
628	* CSAM needs to perform a job in ring 3.
629	*
630	* Sync the page before going to the host context; otherwise we'll end up in a loop if
631	* CSAM fails (e.g. instruction crosses a page boundary and the next page is not present)
632	*/
633	LogFlow(("CSAM ring 3 job\n"));
634	int rc2 = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, 1, uErr);
635	AssertRC(rc2);
636
637	STAM_PROFILE_STOP(&pVM->pgm.s.StatOutOfSync, c);
638	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eCSAM; });
639	return rc;
640	}
641	}
642	# ifdef CSAM_DETECT_NEW_CODE_PAGES
643	else
644	if ( uErr == X86_TRAP_PF_RW
645	&& pRegFrame->ecx >= 0x100 /* early check for movswd count */
646	&& pRegFrame->ecx < 0x10000
647	)
648	{
649	/* In case of a write to a non-present supervisor shadow page, we'll take special precautions
650	* to detect loading of new code pages.
651	*/
652
653	/*
654	* Decode the instruction.
655	*/
656	RTGCPTR PC;
657	rc = SELMValidateAndConvertCSAddr(pVM, pRegFrame->eflags, pRegFrame->ss, pRegFrame->cs, &pRegFrame->csHid, (RTGCPTR)pRegFrame->eip, &PC);
658	if (rc == VINF_SUCCESS)
659	{
660	DISCPUSTATE Cpu;
661	uint32_t cbOp;
662	rc = EMInterpretDisasOneEx(pVM, (RTGCUINTPTR)PC, pRegFrame, &Cpu, &cbOp);
663
664	/* For now we'll restrict this to rep movsw/d instructions */
665	if ( rc == VINF_SUCCESS
666	&& Cpu.pCurInstr->opcode == OP_MOVSWD
667	&& (Cpu.prefix & PREFIX_REP))
668	{
669	CSAMMarkPossibleCodePage(pVM, pvFault);
670	}
671	}
672	}
673	# endif /* CSAM_DETECT_NEW_CODE_PAGES */
674
675	/*
676	* Mark this page as safe.
677	*/
678	/** @todo not correct for pages that contain both code and data!! */
679	Log2(("CSAMMarkPage %VGv; scanned=%d\n", pvFault, true));
680	CSAMMarkPage(pVM, (RTRCPTR)pvFault, true);
681	}
682	}
683	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE) && !defined(IN_RING0) */
684	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, PGM_SYNC_NR_PAGES, uErr);
685	if (VBOX_SUCCESS(rc))
686	{
687	/* The page was successfully synced, return to the guest. */
688	STAM_PROFILE_STOP(&pVM->pgm.s.StatOutOfSync, c);
689	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eOutOfSync; });
690	return VINF_SUCCESS;
691	}
692	}
693	else
694	{
695	/*
696	* A side effect of not flushing global PDEs are out of sync pages due
697	* to physical monitored regions, that are no longer valid.
698	* Assume for now it only applies to the read/write flag
699	*/
700	if (VBOX_SUCCESS(rc) && (uErr & X86_TRAP_PF_RW))
701	{
702	if (uErr & X86_TRAP_PF_US)
703	STAM_COUNTER_INC(&pVM->pgm.s.StatGCPageOutOfSyncUser);
704	else /* supervisor */
705	STAM_COUNTER_INC(&pVM->pgm.s.StatGCPageOutOfSyncSupervisor);
706
707
708	/*
709	* Note: Do NOT use PGM_SYNC_NR_PAGES here. That only works if the page is not present, which is not true in this case.
710	*/
711	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)pvFault, 1, uErr);
712	if (VBOX_SUCCESS(rc))
713	{
714	/*
715	* Page was successfully synced, return to guest.
716	*/
717	# ifdef VBOX_STRICT
718	RTGCPHYS GCPhys;
719	uint64_t fPageGst;
720	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, &GCPhys);
721	Assert(VBOX_SUCCESS(rc) && fPageGst & X86_PTE_RW);
722	LogFlow(("Obsolete physical monitor page out of sync %VGv - phys %VGp flags=%08llx\n", pvFault, GCPhys, (uint64_t)fPageGst));
723
724	uint64_t fPageShw;
725	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
726	AssertMsg(VBOX_SUCCESS(rc) && fPageShw & X86_PTE_RW, ("rc=%Vrc fPageShw=%VX64\n", rc, fPageShw));
727	# endif /* VBOX_STRICT */
728	STAM_PROFILE_STOP(&pVM->pgm.s.StatOutOfSync, c);
729	STAM_STATS({ pVM->pgm.s.CTXSUFF(pStatTrap0eAttribution) = &pVM->pgm.s.StatTrap0eOutOfSyncObsHnd; });
730	return VINF_SUCCESS;
731	}
732
733	/* Check to see if we need to emulate the instruction as X86_CR0_WP has been cleared. */
734	if ( CPUMGetGuestCPL(pVM, pRegFrame) == 0
735	&& ((CPUMGetGuestCR0(pVM) & (X86_CR0_WP\|X86_CR0_PG)) == X86_CR0_PG)
736	&& (uErr & (X86_TRAP_PF_RW \| X86_TRAP_PF_P)) == (X86_TRAP_PF_RW \| X86_TRAP_PF_P))
737	{
738	uint64_t fPageGst;
739	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
740	if ( VBOX_SUCCESS(rc)
741	&& !(fPageGst & X86_PTE_RW))
742	{
743	rc = PGMInterpretInstruction(pVM, pRegFrame, pvFault);
744	if (VBOX_SUCCESS(rc))
745	STAM_COUNTER_INC(&pVM->pgm.s.StatTrap0eWPEmulGC);
746	else
747	STAM_COUNTER_INC(&pVM->pgm.s.StatTrap0eWPEmulR3);
748	return rc;
749	}
750	else
751	AssertMsgFailed(("Unexpected r/w page %x flag=%x\n", pvFault, (uint32_t)fPageGst));
752	}
753
754	}
755
756	# if PGM_WITH_PAGING(PGM_GST_TYPE)
757	# ifdef VBOX_STRICT
758	/*
759	* Check for VMM page flags vs. Guest page flags consistency.
760	* Currently only for debug purposes.
761	*/
762	if (VBOX_SUCCESS(rc))
763	{
764	/* Get guest page flags. */
765	uint64_t fPageGst;
766	rc = PGMGstGetPage(pVM, pvFault, &fPageGst, NULL);
767	if (VBOX_SUCCESS(rc))
768	{
769	uint64_t fPageShw;
770	rc = PGMShwGetPage(pVM, pvFault, &fPageShw, NULL);
771
772	/*
773	* Compare page flags.
774	* Note: we have AVL, A, D bits desynched.
775	*/
776	AssertMsg((fPageShw & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)) == (fPageGst & ~(X86_PTE_A \| X86_PTE_D \| X86_PTE_AVL_MASK)),
777	("Page flags mismatch! pvFault=%VGv GCPhys=%VGp fPageShw=%08llx fPageGst=%08llx\n", pvFault, GCPhys, fPageShw, fPageGst));
778	}
779	else
780	AssertMsgFailed(("PGMGstGetPage rc=%Vrc\n", rc));
781	}
782	else
783	AssertMsgFailed(("PGMGCGetPage rc=%Vrc\n", rc));
784	# endif /* VBOX_STRICT */
785	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
786	}
787	STAM_PROFILE_STOP(&pVM->pgm.s.StatOutOfSync, c);
788	# endif /* PGM_OUT_OF_SYNC_IN_GC */
789	}
790	else
791	{
792	/*
793	* Page not present in Guest OS or invalid page table address.
794	* This is potential virtual page access handler food.
795	*
796	* For the present we'll say that our access handlers don't
797	* work for this case - we've already discarded the page table
798	* not present case which is identical to this.
799	*
800	* When we perchance find we need this, we will probably have AVL
801	* trees (offset based) to operate on and we can measure their speed
802	* agains mapping a page table and probably rearrange this handling
803	* a bit. (Like, searching virtual ranges before checking the
804	* physical address.)
805	*/
806	}
807	}
808
809
810	# if PGM_WITH_PAGING(PGM_GST_TYPE)
811	/*
812	* Conclusion, this is a guest trap.
813	*/
814	LogFlow(("PGM: Unhandled #PF -> route trap to recompiler!\n"));
815	STAM_COUNTER_INC(&pVM->pgm.s.StatGCTrap0eUnhandled);
816	return VINF_EM_RAW_GUEST_TRAP;
817	# else
818	/* present, but not a monitored page; perhaps the guest is probing physical memory */
819	return VINF_EM_RAW_EMULATE_INSTR;
820	# endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
821
822
823	#else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
824
825	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
826	return VERR_INTERNAL_ERROR;
827	#endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
828	}
829
830
831	/**
832	* Emulation of the invlpg instruction.
833	*
834	*
835	* @returns VBox status code.
836	*
837	* @param pVM VM handle.
838	* @param GCPtrPage Page to invalidate.
839	*
840	* @remark ASSUMES that the guest is updating before invalidating. This order
841	* isn't required by the CPU, so this is speculative and could cause
842	* trouble.
843	*
844	* @todo Flush page or page directory only if necessary!
845	* @todo Add a #define for simply invalidating the page.
846	*/
847	PGM_BTH_DECL(int, InvalidatePage)(PVM pVM, RTGCUINTPTR GCPtrPage)
848	{
849	#if PGM_WITH_PAGING(PGM_GST_TYPE) \
850	&& PGM_SHW_TYPE != PGM_TYPE_NESTED
851	int rc;
852
853	LogFlow(("InvalidatePage %x\n", GCPtrPage));
854	/*
855	* Get the shadow PD entry and skip out if this PD isn't present.
856	* (Guessing that it is frequent for a shadow PDE to not be present, do this first.)
857	*/
858	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
859	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
860	PX86PDE pPdeDst = &pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst];
861	# elif PGM_SHW_TYPE == PGM_TYPE_PAE
862	PX86PDEPAE pPdeDst = &pVM->pgm.s.CTXMID(ap,PaePDs[0])->a[iPDDst];
863	# else /* AMD64 */
864	/* PML4 */
865	const unsigned iPml4 = ((RTGCUINTPTR64)GCPtrPage >> X86_PML4_SHIFT) & X86_PML4_MASK;
866	PX86PML4E pPml4eDst = &CTXMID(pVM->pgm.s.p,PaePML4)->a[iPml4];
867	if (!pPml4eDst->n.u1Present)
868	{
869	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePageSkipped));
870	return VINF_SUCCESS;
871	}
872
873	/* PDPT */
874	PX86PDPT pPDPT;
875	rc = PGM_HCPHYS_2_PTR(pVM, pPml4eDst->u & X86_PML4E_PG_MASK, &pPDPT);
876	if (VBOX_FAILURE(rc))
877	return rc;
878	const unsigned iPDPT = ((RTGCUINTPTR64)GCPtrPage >> SHW_PDPT_SHIFT) & SHW_PDPT_MASK;
879	PX86PDPE pPdpeDst = &pPDPT->a[iPDPT];
880	if (!pPdpeDst->n.u1Present)
881	{
882	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePageSkipped));
883	return VINF_SUCCESS;
884	}
885
886	/* PD */
887	PX86PDPAE pPd;
888	rc = PGM_HCPHYS_2_PTR(pVM, pPdpeDst->u & X86_PDPE_PG_MASK, &pPd);
889	if (VBOX_FAILURE(rc))
890	return rc;
891	const unsigned iPd = (GCPtrPage >> SHW_PD_SHIFT) & SHW_PD_MASK;
892	PX86PDEPAE pPdeDst = &pPd->a[iPd];
893	# endif
894
895	const SHWPDE PdeDst = *pPdeDst;
896	if (!PdeDst.n.u1Present)
897	{
898	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePageSkipped));
899	return VINF_SUCCESS;
900	}
901
902	/*
903	* Get the guest PD entry and calc big page.
904	*/
905	# if PGM_GST_TYPE == PGM_TYPE_32BIT
906	PX86PD pPDSrc = CTXSUFF(pVM->pgm.s.pGuestPD);
907	const unsigned iPDSrc = GCPtrPage >> GST_PD_SHIFT;
908	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
909	# else
910	unsigned iPDSrc;
911	# if PGM_GST_TYPE == PGM_TYPE_PAE
912	PX86PDPAE pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc);
913	# else /* AMD64 */
914	PX86PML4E pPml4eSrc;
915	X86PDPE PdpeSrc;
916	PX86PDPAE pPDSrc = pgmGstGetLongModePDPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc, &PdpeSrc, &iPDSrc);
917	# endif
918	GSTPDE PdeSrc;
919
920	if (pPDSrc)
921	PdeSrc = pPDSrc->a[iPDSrc];
922	else
923	PdeSrc.u = 0;
924	# endif
925
926	const uint32_t cr4 = CPUMGetGuestCR4(pVM);
927	const bool fIsBigPage = PdeSrc.b.u1Size && (cr4 & X86_CR4_PSE);
928
929	# ifdef IN_RING3
930	/*
931	* If a CR3 Sync is pending we may ignore the invalidate page operation
932	* depending on the kind of sync and if it's a global page or not.
933	* This doesn't make sense in GC/R0 so we'll skip it entirely there.
934	*/
935	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
936	if ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3)
937	\|\| ( VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3_NON_GLOBAL)
938	&& fIsBigPage
939	&& PdeSrc.b.u1Global
940	&& (cr4 & X86_CR4_PGE)
941	)
942	)
943	# else
944	if (VM_FF_ISPENDING(pVM, VM_FF_PGM_SYNC_CR3 \| VM_FF_PGM_SYNC_CR3_NON_GLOBAL) )
945	# endif
946	{
947	STAM_COUNTER_INC(&pVM->pgm.s.StatHCInvalidatePageSkipped);
948	return VINF_SUCCESS;
949	}
950	# endif /* IN_RING3 */
951
952
953	# if PGM_GST_TYPE == PGM_TYPE_AMD64
954	Assert(pPml4eDst->n.u1Present && pPml4eDst->u & SHW_PDPT_MASK);
955	if (pPml4eSrc->n.u1Present)
956	{
957	if ( pPml4eSrc->n.u1User != pPml4eDst->n.u1User
958	\|\| (!pPml4eSrc->n.u1Write && pPml4eDst->n.u1Write))
959	{
960	/*
961	* Mark not present so we can resync the PML4E when it's used.
962	*/
963	LogFlow(("InvalidatePage: Out-of-sync PML4E at %VGp Pml4eSrc=%RX64 Pml4eDst=%RX64\n",
964	GCPtrPage, (uint64_t)pPml4eSrc->u, (uint64_t)pPml4eDst->u));
965	pgmPoolFree(pVM, pPml4eDst->u & X86_PML4E_PG_MASK, PGMPOOL_IDX_PML4, iPml4);
966	pPml4eDst->u = 0;
967	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDOutOfSync));
968	PGM_INVL_GUEST_TLBS();
969	}
970	else if (!pPml4eSrc->n.u1Accessed)
971	{
972	/*
973	* Mark not present so we can set the accessed bit.
974	*/
975	pgmPoolFree(pVM, pPml4eDst->u & X86_PML4E_PG_MASK, PGMPOOL_IDX_PML4, iPml4);
976	pPml4eDst->u = 0;
977	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDNAs));
978	PGM_INVL_GUEST_TLBS();
979	}
980	}
981	else
982	{
983	pgmPoolFree(pVM, pPml4eDst->u & X86_PML4E_PG_MASK, PGMPOOL_IDX_PML4, iPml4);
984	pPml4eDst->u = 0;
985	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDNPs));
986	PGM_INVL_PG(GCPtrPage);
987	return VINF_SUCCESS;
988	}
989
990	Assert(pPdpeDst->n.u1Present && pPdpeDst->u & SHW_PDPT_MASK);
991	if (PdpeSrc.n.u1Present)
992	{
993	if ( PdpeSrc.lm.u1User != pPdpeDst->lm.u1User
994	\|\| (!PdpeSrc.lm.u1Write && pPdpeDst->lm.u1Write))
995	{
996	/*
997	* Mark not present so we can resync the PML4E when it's used.
998	*/
999	LogFlow(("InvalidatePage: Out-of-sync PDPE at %VGp PdpeSrc=%RX64 PdpeDst=%RX64\n",
1000	GCPtrPage, (uint64_t)PdpeSrc.u, (uint64_t)pPdpeDst->u));
1001	pgmPoolFree(pVM, pPdpeDst->u & SHW_PDPT_MASK, PGMPOOL_IDX_PML4, iPml4);
1002	pPdpeDst->u = 0;
1003	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDOutOfSync));
1004	PGM_INVL_GUEST_TLBS();
1005	}
1006	else if (!PdpeSrc.lm.u1Accessed)
1007	{
1008	/*
1009	* Mark not present so we can set the accessed bit.
1010	*/
1011	pgmPoolFree(pVM, pPdpeDst->u & SHW_PDPT_MASK, PGMPOOL_IDX_PML4, iPml4);
1012	pPdpeDst->u = 0;
1013	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDNAs));
1014	PGM_INVL_GUEST_TLBS();
1015	}
1016	}
1017	else
1018	{
1019	pgmPoolFree(pVM, pPdpeDst->u & SHW_PDPT_MASK, PGMPOOL_IDX_PDPT, iPDDst);
1020	pPdpeDst->u = 0;
1021	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDNPs));
1022	PGM_INVL_PG(GCPtrPage);
1023	return VINF_SUCCESS;
1024	}
1025	# endif
1026
1027	/*
1028	* Deal with the Guest PDE.
1029	*/
1030	rc = VINF_SUCCESS;
1031	if (PdeSrc.n.u1Present)
1032	{
1033	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
1034	{
1035	/*
1036	* Conflict - Let SyncPT deal with it to avoid duplicate code.
1037	*/
1038	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1039	Assert(PGMGetGuestMode(pVM) <= PGMMODE_32_BIT);
1040	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
1041	}
1042	else if ( PdeSrc.n.u1User != PdeDst.n.u1User
1043	\|\| (!PdeSrc.n.u1Write && PdeDst.n.u1Write))
1044	{
1045	/*
1046	* Mark not present so we can resync the PDE when it's used.
1047	*/
1048	LogFlow(("InvalidatePage: Out-of-sync at %VGp PdeSrc=%RX64 PdeDst=%RX64\n",
1049	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1050	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1051	pPdeDst->u = 0;
1052	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDOutOfSync));
1053	PGM_INVL_GUEST_TLBS();
1054	}
1055	else if (!PdeSrc.n.u1Accessed)
1056	{
1057	/*
1058	* Mark not present so we can set the accessed bit.
1059	*/
1060	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1061	pPdeDst->u = 0;
1062	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDNAs));
1063	PGM_INVL_GUEST_TLBS();
1064	}
1065	else if (!fIsBigPage)
1066	{
1067	/*
1068	* 4KB - page.
1069	*/
1070	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1071	RTGCPHYS GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1072	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1073	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1074	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1075	# endif
1076	if (pShwPage->GCPhys == GCPhys)
1077	{
1078	# if 0 /* likely cause of a major performance regression; must be SyncPageWorkerTrackDeref then */
1079	const unsigned iPTEDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1080	PSHWPT pPT = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1081	if (pPT->a[iPTEDst].n.u1Present)
1082	{
1083	# ifdef PGMPOOL_WITH_USER_TRACKING
1084	/* This is very unlikely with caching/monitoring enabled. */
1085	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPT->a[iPTEDst].u & SHW_PTE_PG_MASK);
1086	# endif
1087	pPT->a[iPTEDst].u = 0;
1088	}
1089	# else /* Syncing it here isn't 100% safe and it's probably not worth spending time syncing it. */
1090	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
1091	if (VBOX_SUCCESS(rc))
1092	rc = VINF_SUCCESS;
1093	# endif
1094	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePage4KBPages));
1095	PGM_INVL_PG(GCPtrPage);
1096	}
1097	else
1098	{
1099	/*
1100	* The page table address changed.
1101	*/
1102	LogFlow(("InvalidatePage: Out-of-sync at %VGp PdeSrc=%RX64 PdeDst=%RX64 ShwGCPhys=%VGp iPDDst=%#x\n",
1103	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u, pShwPage->GCPhys, iPDDst));
1104	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1105	pPdeDst->u = 0;
1106	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDOutOfSync));
1107	PGM_INVL_GUEST_TLBS();
1108	}
1109	}
1110	else
1111	{
1112	/*
1113	* 2/4MB - page.
1114	*/
1115	/* Before freeing the page, check if anything really changed. */
1116	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1117	RTGCPHYS GCPhys = PdeSrc.u & GST_PDE_BIG_PG_MASK;
1118	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1119	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1120	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1121	# endif
1122	if ( pShwPage->GCPhys == GCPhys
1123	&& pShwPage->enmKind == BTH_PGMPOOLKIND_PT_FOR_BIG)
1124	{
1125	/* ASSUMES a the given bits are identical for 4M and normal PDEs */
1126	/** @todo PAT */
1127	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1128	== (PdeDst.u & (X86_PDE_P \| X86_PDE_RW \| X86_PDE_US \| X86_PDE_PWT \| X86_PDE_PCD))
1129	&& ( PdeSrc.b.u1Dirty /** @todo rainy day: What about read-only 4M pages? not very common, but still... */
1130	\|\| (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)))
1131	{
1132	LogFlow(("Skipping flush for big page containing %VGv (PD=%X .u=%VX64)-> nothing has changed!\n", GCPtrPage, iPDSrc, PdeSrc.u));
1133	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePage4MBPagesSkip));
1134	return VINF_SUCCESS;
1135	}
1136	}
1137
1138	/*
1139	* Ok, the page table is present and it's been changed in the guest.
1140	* If we're in host context, we'll just mark it as not present taking the lazy approach.
1141	* We could do this for some flushes in GC too, but we need an algorithm for
1142	* deciding which 4MB pages containing code likely to be executed very soon.
1143	*/
1144	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1145	pPdeDst->u = 0;
1146	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePage4MBPages));
1147	PGM_INVL_BIG_PG(GCPtrPage);
1148	}
1149	}
1150	else
1151	{
1152	/*
1153	* Page directory is not present, mark shadow PDE not present.
1154	*/
1155	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
1156	{
1157	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1158	pPdeDst->u = 0;
1159	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDNPs));
1160	PGM_INVL_PG(GCPtrPage);
1161	}
1162	else
1163	{
1164	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
1165	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,InvalidatePagePDMappings));
1166	}
1167	}
1168
1169	return rc;
1170
1171	#else /* guest real and protected mode */
1172	/* There's no such thing as InvalidatePage when paging is disabled, so just ignore. */
1173	return VINF_SUCCESS;
1174	#endif
1175	}
1176
1177
1178	#ifdef PGMPOOL_WITH_USER_TRACKING
1179	/**
1180	* Update the tracking of shadowed pages.
1181	*
1182	* @param pVM The VM handle.
1183	* @param pShwPage The shadow page.
1184	* @param HCPhys The physical page we is being dereferenced.
1185	*/
1186	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackDeref)(PVM pVM, PPGMPOOLPAGE pShwPage, RTHCPHYS HCPhys)
1187	{
1188	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1189	STAM_PROFILE_START(&pVM->pgm.s.StatTrackDeref, a);
1190	LogFlow(("SyncPageWorkerTrackDeref: Damn HCPhys=%VHp pShwPage->idx=%#x!!!\n", HCPhys, pShwPage->idx));
1191
1192	/** @todo If this turns out to be a bottle neck (very likely) two things can be done:
1193	* 1. have a medium sized HCPhys -> GCPhys TLB (hash?)
1194	* 2. write protect all shadowed pages. I.e. implement caching.
1195	*/
1196	/*
1197	* Find the guest address.
1198	*/
1199	for (PPGMRAMRANGE pRam = CTXALLSUFF(pVM->pgm.s.pRamRanges);
1200	pRam;
1201	pRam = CTXALLSUFF(pRam->pNext))
1202	{
1203	unsigned iPage = pRam->cb >> PAGE_SHIFT;
1204	while (iPage-- > 0)
1205	{
1206	if (PGM_PAGE_GET_HCPHYS(&pRam->aPages[iPage]) == HCPhys)
1207	{
1208	PPGMPOOL pPool = pVM->pgm.s.CTXSUFF(pPool);
1209	pgmTrackDerefGCPhys(pPool, pShwPage, &pRam->aPages[iPage]);
1210	pShwPage->cPresent--;
1211	pPool->cPresent--;
1212	STAM_PROFILE_STOP(&pVM->pgm.s.StatTrackDeref, a);
1213	return;
1214	}
1215	}
1216	}
1217
1218	for (;;)
1219	AssertReleaseMsgFailed(("HCPhys=%VHp wasn't found!\n", HCPhys));
1220	# else /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1221	pShwPage->cPresent--;
1222	pVM->pgm.s.CTXSUFF(pPool)->cPresent--;
1223	# endif /* !PGMPOOL_WITH_GCPHYS_TRACKING */
1224	}
1225
1226
1227	/**
1228	* Update the tracking of shadowed pages.
1229	*
1230	* @param pVM The VM handle.
1231	* @param pShwPage The shadow page.
1232	* @param u16 The top 16-bit of the pPage->HCPhys.
1233	* @param pPage Pointer to the guest page. this will be modified.
1234	* @param iPTDst The index into the shadow table.
1235	*/
1236	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorkerTrackAddref)(PVM pVM, PPGMPOOLPAGE pShwPage, uint16_t u16, PPGMPAGE pPage, const unsigned iPTDst)
1237	{
1238	# ifdef PGMPOOL_WITH_GCPHYS_TRACKING
1239	/*
1240	* We're making certain assumptions about the placement of cRef and idx.
1241	*/
1242	Assert(MM_RAM_FLAGS_IDX_SHIFT == 48);
1243	Assert(MM_RAM_FLAGS_CREFS_SHIFT > MM_RAM_FLAGS_IDX_SHIFT);
1244
1245	/*
1246	* Just deal with the simple first time here.
1247	*/
1248	if (!u16)
1249	{
1250	STAM_COUNTER_INC(&pVM->pgm.s.StatTrackVirgin);
1251	u16 = (1 << (MM_RAM_FLAGS_CREFS_SHIFT - MM_RAM_FLAGS_IDX_SHIFT)) \| pShwPage->idx;
1252	}
1253	else
1254	u16 = pgmPoolTrackPhysExtAddref(pVM, u16, pShwPage->idx);
1255
1256	/* write back, trying to be clever... */
1257	Log2(("SyncPageWorkerTrackAddRef: u16=%#x pPage->HCPhys=%VHp->%VHp iPTDst=%#x\n",
1258	u16, pPage->HCPhys, (pPage->HCPhys & MM_RAM_FLAGS_NO_REFS_MASK) \| ((uint64_t)u16 << MM_RAM_FLAGS_CREFS_SHIFT), iPTDst));
1259	((uint16_t )&pPage->HCPhys + 3) = u16; /** @todo PAGE FLAGS */
1260	# endif /* PGMPOOL_WITH_GCPHYS_TRACKING */
1261
1262	/* update statistics. */
1263	pVM->pgm.s.CTXSUFF(pPool)->cPresent++;
1264	pShwPage->cPresent++;
1265	if (pShwPage->iFirstPresent > iPTDst)
1266	pShwPage->iFirstPresent = iPTDst;
1267	}
1268	#endif /* PGMPOOL_WITH_USER_TRACKING */
1269
1270
1271	/**
1272	* Creates a 4K shadow page for a guest page.
1273	*
1274	* For 4M pages the caller must convert the PDE4M to a PTE, this includes adjusting the
1275	* physical address. The PdeSrc argument only the flags are used. No page structured
1276	* will be mapped in this function.
1277	*
1278	* @param pVM VM handle.
1279	* @param pPteDst Destination page table entry.
1280	* @param PdeSrc Source page directory entry (i.e. Guest OS page directory entry).
1281	* Can safely assume that only the flags are being used.
1282	* @param PteSrc Source page table entry (i.e. Guest OS page table entry).
1283	* @param pShwPage Pointer to the shadow page.
1284	* @param iPTDst The index into the shadow table.
1285	*
1286	* @remark Not used for 2/4MB pages!
1287	*/
1288	DECLINLINE(void) PGM_BTH_NAME(SyncPageWorker)(PVM pVM, PSHWPTE pPteDst, GSTPDE PdeSrc, GSTPTE PteSrc, PPGMPOOLPAGE pShwPage, unsigned iPTDst)
1289	{
1290	if (PteSrc.n.u1Present)
1291	{
1292	/*
1293	* Find the ram range.
1294	*/
1295	PPGMPAGE pPage;
1296	int rc = pgmPhysGetPageEx(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK, &pPage);
1297	if (VBOX_SUCCESS(rc))
1298	{
1299	/** @todo investiage PWT, PCD and PAT. */
1300	/*
1301	* Make page table entry.
1302	*/
1303	const RTHCPHYS HCPhys = pPage->HCPhys; /** @todo FLAGS */
1304	SHWPTE PteDst;
1305	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1306	{
1307	/** @todo r=bird: Are we actually handling dirty and access bits for pages with access handlers correctly? No. */
1308	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1309	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))
1310	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1311	else
1312	{
1313	LogFlow(("SyncPageWorker: monitored page (%VGp) -> mark not present\n", HCPhys));
1314	PteDst.u = 0;
1315	}
1316	/** @todo count these two kinds. */
1317	}
1318	else
1319	{
1320	/*
1321	* If the page or page directory entry is not marked accessed,
1322	* we mark the page not present.
1323	*/
1324	if (!PteSrc.n.u1Accessed \|\| !PdeSrc.n.u1Accessed)
1325	{
1326	LogFlow(("SyncPageWorker: page and or page directory not accessed -> mark not present\n"));
1327	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,AccessedPage));
1328	PteDst.u = 0;
1329	}
1330	else
1331	/*
1332	* If the page is not flagged as dirty and is writable, then make it read-only, so we can set the dirty bit
1333	* when the page is modified.
1334	*/
1335	if (!PteSrc.n.u1Dirty && (PdeSrc.n.u1Write & PteSrc.n.u1Write))
1336	{
1337	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,DirtyPage));
1338	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))
1339	\| (HCPhys & X86_PTE_PAE_PG_MASK)
1340	\| PGM_PTFLAGS_TRACK_DIRTY;
1341	}
1342	else
1343	{
1344	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,DirtyPageSkipped));
1345	PteDst.u = (PteSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1346	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1347	}
1348	}
1349
1350	#ifdef PGMPOOL_WITH_USER_TRACKING
1351	/*
1352	* Keep user track up to date.
1353	*/
1354	if (PteDst.n.u1Present)
1355	{
1356	if (!pPteDst->n.u1Present)
1357	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1358	else if ((pPteDst->u & SHW_PTE_PG_MASK) != (PteDst.u & SHW_PTE_PG_MASK))
1359	{
1360	Log2(("SyncPageWorker: deref! *pPteDst=%RX64 PteDst=%RX64\n", (uint64_t)pPteDst->u, (uint64_t)PteDst.u));
1361	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1362	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1363	}
1364	}
1365	else if (pPteDst->n.u1Present)
1366	{
1367	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1368	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1369	}
1370	#endif /* PGMPOOL_WITH_USER_TRACKING */
1371
1372	/*
1373	* Update statistics and commit the entry.
1374	*/
1375	if (!PteSrc.n.u1Global)
1376	pShwPage->fSeenNonGlobal = true;
1377	*pPteDst = PteDst;
1378	}
1379	/* else MMIO or invalid page, we must handle them manually in the #PF handler. */
1380	/** @todo count these. */
1381	}
1382	else
1383	{
1384	/*
1385	* Page not-present.
1386	*/
1387	LogFlow(("SyncPageWorker: page not present in Pte\n"));
1388	#ifdef PGMPOOL_WITH_USER_TRACKING
1389	/* Keep user track up to date. */
1390	if (pPteDst->n.u1Present)
1391	{
1392	Log2(("SyncPageWorker: deref! *pPteDst=%RX64\n", (uint64_t)pPteDst->u));
1393	PGM_BTH_NAME(SyncPageWorkerTrackDeref)(pVM, pShwPage, pPteDst->u & SHW_PTE_PG_MASK);
1394	}
1395	#endif /* PGMPOOL_WITH_USER_TRACKING */
1396	pPteDst->u = 0;
1397	/** @todo count these. */
1398	}
1399	}
1400
1401
1402	/**
1403	* Syncs a guest OS page.
1404	*
1405	* There are no conflicts at this point, neither is there any need for
1406	* page table allocations.
1407	*
1408	* @returns VBox status code.
1409	* @returns VINF_PGM_SYNCPAGE_MODIFIED_PDE if it modifies the PDE in any way.
1410	* @param pVM VM handle.
1411	* @param PdeSrc Page directory entry of the guest.
1412	* @param GCPtrPage Guest context page address.
1413	* @param cPages Number of pages to sync (PGM_SYNC_N_PAGES) (default=1).
1414	* @param uErr Fault error (X86_TRAP_PF_*).
1415	*/
1416	PGM_BTH_DECL(int, SyncPage)(PVM pVM, GSTPDE PdeSrc, RTGCUINTPTR GCPtrPage, unsigned cPages, unsigned uErr)
1417	{
1418	LogFlow(("SyncPage: GCPtrPage=%VGv cPages=%d uErr=%#x\n", GCPtrPage, cPages, uErr));
1419
1420	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
1421	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
1422	&& PGM_SHW_TYPE != PGM_TYPE_NESTED
1423
1424	# if PGM_WITH_NX(PGM_GST_TYPE)
1425	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1426	# endif
1427
1428	/*
1429	* Assert preconditions.
1430	*/
1431	STAM_COUNTER_INC(&pVM->pgm.s.StatGCSyncPagePD[(GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK]);
1432	Assert(PdeSrc.n.u1Present);
1433	Assert(cPages);
1434
1435	/*
1436	* Get the shadow PDE, find the shadow page table in the pool.
1437	*/
1438	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
1439	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1440	X86PDE PdeDst = pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst];
1441	# else /* PAE */
1442	X86PDEPAE PdeDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[iPDDst];
1443	# endif
1444	Assert(PdeDst.n.u1Present);
1445	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1446
1447	/*
1448	* Check that the page is present and that the shadow PDE isn't out of sync.
1449	*/
1450	const bool fBigPage = PdeSrc.b.u1Size && (CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1451	RTGCPHYS GCPhys;
1452	if (!fBigPage)
1453	{
1454	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
1455	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1456	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1457	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE/2);
1458	# endif
1459	}
1460	else
1461	{
1462	GCPhys = PdeSrc.u & GST_PDE_BIG_PG_MASK;
1463	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1464	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
1465	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
1466	# endif
1467	}
1468	if ( pShwPage->GCPhys == GCPhys
1469	&& PdeSrc.n.u1Present
1470	&& (PdeSrc.n.u1User == PdeDst.n.u1User)
1471	&& (PdeSrc.n.u1Write == PdeDst.n.u1Write \|\| !PdeDst.n.u1Write)
1472	# if PGM_WITH_NX(PGM_GST_TYPE)
1473	&& (!fNoExecuteBitValid \|\| PdeSrc.n.u1NoExecute == PdeDst.n.u1NoExecute)
1474	# endif
1475	)
1476	{
1477	/*
1478	* Check that the PDE is marked accessed already.
1479	* Since we set the accessed bit before getting here on a #PF, this
1480	* check is only meant for dealing with non-#PF'ing paths.
1481	*/
1482	if (PdeSrc.n.u1Accessed)
1483	{
1484	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1485	if (!fBigPage)
1486	{
1487	/*
1488	* 4KB Page - Map the guest page table.
1489	*/
1490	PGSTPT pPTSrc;
1491	int rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
1492	if (VBOX_SUCCESS(rc))
1493	{
1494	# ifdef PGM_SYNC_N_PAGES
1495	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1496	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1497	{
1498	/*
1499	* This code path is currently only taken when the caller is PGMTrap0eHandler
1500	* for non-present pages!
1501	*
1502	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1503	* deal with locality.
1504	*/
1505	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1506	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
1507	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1508	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
1509	# else
1510	const unsigned offPTSrc = 0;
1511	# endif
1512	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, ELEMENTS(pPTDst->a));
1513	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1514	iPTDst = 0;
1515	else
1516	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1517	for (; iPTDst < iPTDstEnd; iPTDst++)
1518	{
1519	if (!pPTDst->a[iPTDst].n.u1Present)
1520	{
1521	GSTPTE PteSrc = pPTSrc->a[offPTSrc + iPTDst];
1522	RTGCUINTPTR GCPtrCurPage = ((RTGCUINTPTR)GCPtrPage & ~(RTGCUINTPTR)(GST_PT_MASK << GST_PT_SHIFT)) \| ((offPTSrc + iPTDst) << PAGE_SHIFT);
1523	NOREF(GCPtrCurPage);
1524	#ifndef IN_RING0
1525	/*
1526	* Assuming kernel code will be marked as supervisor - and not as user level
1527	* and executed using a conforming code selector - And marked as readonly.
1528	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
1529	*/
1530	PPGMPAGE pPage;
1531	if ( ((PdeSrc.u & PteSrc.u) & (X86_PTE_RW \| X86_PTE_US))
1532	\|\| iPTDst == ((GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK) /* always sync GCPtrPage */
1533	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)GCPtrCurPage)
1534	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
1535	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1536	)
1537	#endif /* else: CSAM not active */
1538	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1539	Log2(("SyncPage: 4K+ %VGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1540	GCPtrCurPage, PteSrc.n.u1Present,
1541	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1542	PteSrc.n.u1User & PdeSrc.n.u1User,
1543	(uint64_t)PteSrc.u,
1544	(uint64_t)pPTDst->a[iPTDst].u,
1545	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1546	}
1547	}
1548	}
1549	else
1550	# endif /* PGM_SYNC_N_PAGES */
1551	{
1552	const unsigned iPTSrc = (GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK;
1553	GSTPTE PteSrc = pPTSrc->a[iPTSrc];
1554	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1555	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1556	Log2(("SyncPage: 4K %VGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}%s\n",
1557	GCPtrPage, PteSrc.n.u1Present,
1558	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1559	PteSrc.n.u1User & PdeSrc.n.u1User,
1560	(uint64_t)PteSrc.u,
1561	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1562	}
1563	}
1564	else /* MMIO or invalid page: emulated in #PF handler. */
1565	{
1566	LogFlow(("PGM_GCPHYS_2_PTR %VGp failed with %Vrc\n", GCPhys, rc));
1567	Assert(!pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK].n.u1Present);
1568	}
1569	}
1570	else
1571	{
1572	/*
1573	* 4/2MB page - lazy syncing shadow 4K pages.
1574	* (There are many causes of getting here, it's no longer only CSAM.)
1575	*/
1576	/* Calculate the GC physical address of this 4KB shadow page. */
1577	RTGCPHYS GCPhys = (PdeSrc.u & GST_PDE_BIG_PG_MASK) \| ((RTGCUINTPTR)GCPtrPage & GST_BIG_PAGE_OFFSET_MASK);
1578	/* Find ram range. */
1579	PPGMPAGE pPage;
1580	int rc = pgmPhysGetPageEx(&pVM->pgm.s, GCPhys, &pPage);
1581	if (VBOX_SUCCESS(rc))
1582	{
1583	/*
1584	* Make shadow PTE entry.
1585	*/
1586	const RTHCPHYS HCPhys = pPage->HCPhys; /** @todo PAGE FLAGS */
1587	SHWPTE PteDst;
1588	PteDst.u = (PdeSrc.u & ~(X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT))
1589	\| (HCPhys & X86_PTE_PAE_PG_MASK);
1590	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
1591	{
1592	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
1593	PteDst.n.u1Write = 0;
1594	else
1595	PteDst.u = 0;
1596	}
1597	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1598	# ifdef PGMPOOL_WITH_USER_TRACKING
1599	if (PteDst.n.u1Present && !pPTDst->a[iPTDst].n.u1Present)
1600	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst);
1601	# endif
1602	pPTDst->a[iPTDst] = PteDst;
1603
1604
1605	/*
1606	* If the page is not flagged as dirty and is writable, then make it read-only
1607	* at PD level, so we can set the dirty bit when the page is modified.
1608	*
1609	* ASSUMES that page access handlers are implemented on page table entry level.
1610	* Thus we will first catch the dirty access and set PDE.D and restart. If
1611	* there is an access handler, we'll trap again and let it work on the problem.
1612	*/
1613	/** @todo r=bird: figure out why we need this here, SyncPT should've taken care of this already.
1614	* As for invlpg, it simply frees the whole shadow PT.
1615	* ...It's possibly because the guest clears it and the guest doesn't really tell us... */
1616	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
1617	{
1618	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,DirtyPageBig));
1619	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
1620	PdeDst.n.u1Write = 0;
1621	}
1622	else
1623	{
1624	PdeDst.au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
1625	PdeDst.n.u1Write = PdeSrc.n.u1Write;
1626	}
1627	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1628	pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst] = PdeDst;
1629	# else /* PAE */
1630	pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[iPDDst] = PdeDst;
1631	# endif
1632	Log2(("SyncPage: BIG %VGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} GCPhys=%VGp%s\n",
1633	GCPtrPage, PdeSrc.n.u1Present, PdeSrc.n.u1Write, PdeSrc.n.u1User, (uint64_t)PdeSrc.u, GCPhys,
1634	PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1635	}
1636	else
1637	LogFlow(("PGM_GCPHYS_2_PTR %VGp (big) failed with %Vrc\n", GCPhys, rc));
1638	}
1639	return VINF_SUCCESS;
1640	}
1641	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncPagePDNAs));
1642	}
1643	else
1644	{
1645	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncPagePDOutOfSync));
1646	Log2(("SyncPage: Out-Of-Sync PDE at %VGp PdeSrc=%RX64 PdeDst=%RX64\n",
1647	GCPtrPage, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
1648	}
1649
1650	/*
1651	* Mark the PDE not present. Restart the instruction and let #PF call SyncPT.
1652	* Yea, I'm lazy.
1653	*/
1654	pgmPoolFree(pVM, PdeDst.u & SHW_PDE_PG_MASK, SHW_POOL_ROOT_IDX, iPDDst);
1655	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1656	pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst].u = 0;
1657	# else /* PAE */
1658	pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[iPDDst].u = 0;
1659	# endif
1660	PGM_INVL_GUEST_TLBS();
1661	return VINF_PGM_SYNCPAGE_MODIFIED_PDE;
1662
1663	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
1664	&& PGM_SHW_TYPE != PGM_TYPE_NESTED
1665
1666	# ifdef PGM_SYNC_N_PAGES
1667	/*
1668	* Get the shadow PDE, find the shadow page table in the pool.
1669	*/
1670	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
1671	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
1672	X86PDE PdeDst = pVM->pgm.s.CTXMID(p,32BitPD)->a[iPDDst];
1673	# else /* PAE */
1674	X86PDEPAE PdeDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[iPDDst];
1675	# endif
1676	Assert(PdeDst.n.u1Present);
1677	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, PdeDst.u & SHW_PDE_PG_MASK);
1678	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1679
1680	# if PGM_SHW_TYPE == PGM_TYPE_PAE
1681	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
1682	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
1683	# else
1684	const unsigned offPTSrc = 0;
1685	# endif
1686
1687	Assert(cPages == 1 \|\| !(uErr & X86_TRAP_PF_P));
1688	if (cPages > 1 && !(uErr & X86_TRAP_PF_P))
1689	{
1690	/*
1691	* This code path is currently only taken when the caller is PGMTrap0eHandler
1692	* for non-present pages!
1693	*
1694	* We're setting PGM_SYNC_NR_PAGES pages around the faulting page to sync it and
1695	* deal with locality.
1696	*/
1697	unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1698	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, ELEMENTS(pPTDst->a));
1699	if (iPTDst < PGM_SYNC_NR_PAGES / 2)
1700	iPTDst = 0;
1701	else
1702	iPTDst -= PGM_SYNC_NR_PAGES / 2;
1703	for (; iPTDst < iPTDstEnd; iPTDst++)
1704	{
1705	if (!pPTDst->a[iPTDst].n.u1Present)
1706	{
1707	GSTPTE PteSrc;
1708
1709	RTGCUINTPTR GCPtrCurPage = ((RTGCUINTPTR)GCPtrPage & ~(RTGCUINTPTR)(GST_PT_MASK << GST_PT_SHIFT)) \| ((offPTSrc + iPTDst) << PAGE_SHIFT);
1710
1711	/* Fake the page table entry */
1712	PteSrc.u = GCPtrCurPage;
1713	PteSrc.n.u1Present = 1;
1714	PteSrc.n.u1Dirty = 1;
1715	PteSrc.n.u1Accessed = 1;
1716	PteSrc.n.u1Write = 1;
1717	PteSrc.n.u1User = 1;
1718
1719	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1720
1721	Log2(("SyncPage: 4K+ %VGv PteSrc:{P=%d RW=%d U=%d raw=%08llx} PteDst=%08llx%s\n",
1722	GCPtrCurPage, PteSrc.n.u1Present,
1723	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1724	PteSrc.n.u1User & PdeSrc.n.u1User,
1725	(uint64_t)PteSrc.u,
1726	(uint64_t)pPTDst->a[iPTDst].u,
1727	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1728	}
1729	}
1730	}
1731	else
1732	# endif /* PGM_SYNC_N_PAGES */
1733	{
1734	GSTPTE PteSrc;
1735	const unsigned iPTDst = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
1736	RTGCUINTPTR GCPtrCurPage = ((RTGCUINTPTR)GCPtrPage & ~(RTGCUINTPTR)(GST_PT_MASK << GST_PT_SHIFT)) \| ((offPTSrc + iPTDst) << PAGE_SHIFT);
1737
1738	/* Fake the page table entry */
1739	PteSrc.u = GCPtrCurPage;
1740	PteSrc.n.u1Present = 1;
1741	PteSrc.n.u1Dirty = 1;
1742	PteSrc.n.u1Accessed = 1;
1743	PteSrc.n.u1Write = 1;
1744	PteSrc.n.u1User = 1;
1745	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
1746
1747	Log2(("SyncPage: 4K %VGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}PteDst=%08llx%s\n",
1748	GCPtrPage, PteSrc.n.u1Present,
1749	PteSrc.n.u1Write & PdeSrc.n.u1Write,
1750	PteSrc.n.u1User & PdeSrc.n.u1User,
1751	(uint64_t)PteSrc.u,
1752	(uint64_t)pPTDst->a[iPTDst].u,
1753	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
1754	}
1755	return VINF_SUCCESS;
1756
1757	#else /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
1758	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
1759	return VERR_INTERNAL_ERROR;
1760	#endif /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
1761	}
1762
1763
1764
1765	#if PGM_WITH_PAGING(PGM_GST_TYPE)
1766
1767	/**
1768	* Investigate page fault and handle write protection page faults caused by
1769	* dirty bit tracking.
1770	*
1771	* @returns VBox status code.
1772	* @param pVM VM handle.
1773	* @param uErr Page fault error code.
1774	* @param pPdeDst Shadow page directory entry.
1775	* @param pPdeSrc Guest page directory entry.
1776	* @param GCPtrPage Guest context page address.
1777	*/
1778	PGM_BTH_DECL(int, CheckPageFault)(PVM pVM, uint32_t uErr, PSHWPDE pPdeDst, PGSTPDE pPdeSrc, RTGCUINTPTR GCPtrPage)
1779	{
1780	bool fWriteProtect = !!(CPUMGetGuestCR0(pVM) & X86_CR0_WP);
1781	bool fUserLevelFault = !!(uErr & X86_TRAP_PF_US);
1782	bool fWriteFault = !!(uErr & X86_TRAP_PF_RW);
1783	bool fBigPagesSupported = !!(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
1784	# if PGM_WITH_NX(PGM_GST_TYPE)
1785	bool fNoExecuteBitValid = !!(CPUMGetGuestEFER(pVM) & MSR_K6_EFER_NXE);
1786	# endif
1787	unsigned uPageFaultLevel;
1788	int rc;
1789
1790	STAM_PROFILE_START(&pVM->pgm.s.CTXMID(Stat, DirtyBitTracking), a);
1791	LogFlow(("CheckPageFault: GCPtrPage=%VGv uErr=%#x PdeSrc=%08x\n", GCPtrPage, uErr, pPdeSrc->u));
1792
1793	# if PGM_GST_TYPE == PGM_TYPE_PAE \
1794	\|\| PGM_GST_TYPE == PGM_TYPE_AMD64
1795
1796	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1797	PX86PML4E pPml4eSrc;
1798	PX86PDPE pPdpeSrc;
1799
1800	pPdpeSrc = pgmGstGetLongModePDPTPtr(&pVM->pgm.s, GCPtrPage, &pPml4eSrc);
1801	Assert(pPml4eSrc);
1802
1803	/*
1804	* Real page fault? (PML4E level)
1805	*/
1806	if ( (uErr & X86_TRAP_PF_RSVD)
1807	\|\| !pPml4eSrc->n.u1Present
1808	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPml4eSrc->n.u1NoExecute)
1809	\|\| (fWriteFault && !pPml4eSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
1810	\|\| (fUserLevelFault && !pPml4eSrc->n.u1User)
1811	)
1812	{
1813	uPageFaultLevel = 0;
1814	goto UpperLevelPageFault;
1815	}
1816	Assert(pPdpeSrc);
1817
1818	# else /* PAE */
1819	PX86PDPE pPdpeSrc = &pVM->pgm.s.CTXSUFF(pGstPaePDPT)->a[(GCPtrPage >> GST_PDPT_SHIFT) & GST_PDPT_MASK];
1820	# endif
1821
1822	/*
1823	* Real page fault? (PDPE level)
1824	*/
1825	if ( (uErr & X86_TRAP_PF_RSVD)
1826	\|\| !pPdpeSrc->n.u1Present
1827	# if PGM_GST_TYPE == PGM_TYPE_AMD64 /* NX, r/w, u/s bits in the PDPE are long mode only */
1828	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdpeSrc->lm.u1NoExecute)
1829	\|\| (fWriteFault && !pPdpeSrc->lm.u1Write && (fUserLevelFault \|\| fWriteProtect))
1830	\|\| (fUserLevelFault && !pPdpeSrc->lm.u1User)
1831	# endif
1832	)
1833	{
1834	uPageFaultLevel = 1;
1835	goto UpperLevelPageFault;
1836	}
1837	# endif
1838
1839	/*
1840	* Real page fault? (PDE level)
1841	*/
1842	if ( (uErr & X86_TRAP_PF_RSVD)
1843	\|\| !pPdeSrc->n.u1Present
1844	# if PGM_WITH_NX(PGM_GST_TYPE)
1845	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && pPdeSrc->n.u1NoExecute)
1846	# endif
1847	\|\| (fWriteFault && !pPdeSrc->n.u1Write && (fUserLevelFault \|\| fWriteProtect))
1848	\|\| (fUserLevelFault && !pPdeSrc->n.u1User) )
1849	{
1850	uPageFaultLevel = 2;
1851	goto UpperLevelPageFault;
1852	}
1853
1854	/*
1855	* First check the easy case where the page directory has been marked read-only to track
1856	* the dirty bit of an emulated BIG page
1857	*/
1858	if (pPdeSrc->b.u1Size && fBigPagesSupported)
1859	{
1860	/* Mark guest page directory as accessed */
1861	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1862	pPml4eSrc->n.u1Accessed = 1;
1863	pPdpeSrc->lm.u1Accessed = 1;
1864	# endif
1865	pPdeSrc->b.u1Accessed = 1;
1866
1867	/*
1868	* Only write protection page faults are relevant here.
1869	*/
1870	if (fWriteFault)
1871	{
1872	/* Mark guest page directory as dirty (BIG page only). */
1873	pPdeSrc->b.u1Dirty = 1;
1874
1875	if (pPdeDst->n.u1Present && (pPdeDst->u & PGM_PDFLAGS_TRACK_DIRTY))
1876	{
1877	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,DirtyPageTrap));
1878
1879	Assert(pPdeSrc->b.u1Write);
1880
1881	pPdeDst->n.u1Write = 1;
1882	pPdeDst->n.u1Accessed = 1;
1883	pPdeDst->au32[0] &= ~PGM_PDFLAGS_TRACK_DIRTY;
1884	PGM_INVL_BIG_PG(GCPtrPage);
1885	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,DirtyBitTracking), a);
1886	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
1887	}
1888	}
1889	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,DirtyBitTracking), a);
1890	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
1891	}
1892	/* else: 4KB page table */
1893
1894	/*
1895	* Map the guest page table.
1896	*/
1897	PGSTPT pPTSrc;
1898	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
1899	if (VBOX_SUCCESS(rc))
1900	{
1901	/*
1902	* Real page fault?
1903	*/
1904	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
1905	const GSTPTE PteSrc = *pPteSrc;
1906	if ( !PteSrc.n.u1Present
1907	# if PGM_WITH_NX(PGM_GST_TYPE)
1908	\|\| (fNoExecuteBitValid && (uErr & X86_TRAP_PF_ID) && PteSrc.n.u1NoExecute)
1909	# endif
1910	\|\| (fWriteFault && !PteSrc.n.u1Write && (fUserLevelFault \|\| fWriteProtect))
1911	\|\| (fUserLevelFault && !PteSrc.n.u1User)
1912	)
1913	{
1914	# ifdef IN_GC
1915	STAM_COUNTER_INC(&pVM->pgm.s.StatGCDirtyTrackRealPF);
1916	# endif
1917	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,DirtyBitTracking), a);
1918	LogFlow(("CheckPageFault: real page fault at %VGv PteSrc.u=%08x (2)\n", GCPtrPage, PteSrc.u));
1919
1920	/* Check the present bit as the shadow tables can cause different error codes by being out of sync.
1921	* See the 2nd case above as well.
1922	*/
1923	if (pPdeSrc->n.u1Present && pPteSrc->n.u1Present)
1924	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
1925
1926	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,DirtyBitTracking), a);
1927	return VINF_EM_RAW_GUEST_TRAP;
1928	}
1929	LogFlow(("CheckPageFault: page fault at %VGv PteSrc.u=%08x\n", GCPtrPage, PteSrc.u));
1930
1931	/*
1932	* Set the accessed bits in the page directory and the page table.
1933	*/
1934	# if PGM_GST_TYPE == PGM_TYPE_AMD64
1935	pPml4eSrc->n.u1Accessed = 1;
1936	pPdpeSrc->lm.u1Accessed = 1;
1937	# endif
1938	pPdeSrc->n.u1Accessed = 1;
1939	pPteSrc->n.u1Accessed = 1;
1940
1941	/*
1942	* Only write protection page faults are relevant here.
1943	*/
1944	if (fWriteFault)
1945	{
1946	/* Write access, so mark guest entry as dirty. */
1947	# if defined(IN_GC) && defined(VBOX_WITH_STATISTICS)
1948	if (!pPteSrc->n.u1Dirty)
1949	STAM_COUNTER_INC(&pVM->pgm.s.StatGCDirtiedPage);
1950	else
1951	STAM_COUNTER_INC(&pVM->pgm.s.StatGCPageAlreadyDirty);
1952	# endif
1953
1954	pPteSrc->n.u1Dirty = 1;
1955
1956	if (pPdeDst->n.u1Present)
1957	{
1958	/* Bail out here as pgmPoolGetPageByHCPhys will return NULL and we'll crash below.
1959	* Our individual shadow handlers will provide more information and force a fatal exit.
1960	*/
1961	if (MMHyperIsInsideArea(pVM, (RTGCPTR)GCPtrPage))
1962	{
1963	LogRel(("CheckPageFault: write to hypervisor region %VGv\n", GCPtrPage));
1964	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,DirtyBitTracking), a);
1965	return VINF_SUCCESS;
1966	}
1967
1968	/*
1969	* Map shadow page table.
1970	*/
1971	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
1972	if (pShwPage)
1973	{
1974	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
1975	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
1976	if ( pPteDst->n.u1Present /** @todo Optimize accessed bit emulation? */
1977	&& (pPteDst->u & PGM_PTFLAGS_TRACK_DIRTY))
1978	{
1979	LogFlow(("DIRTY page trap addr=%VGv\n", GCPtrPage));
1980	# ifdef VBOX_STRICT
1981	PPGMPAGE pPage = pgmPhysGetPage(&pVM->pgm.s, pPteSrc->u & GST_PTE_PG_MASK);
1982	if (pPage)
1983	AssertMsg(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage),
1984	("Unexpected dirty bit tracking on monitored page %VGv (phys %VGp)!!!!!!\n", GCPtrPage, pPteSrc->u & X86_PTE_PAE_PG_MASK));
1985	# endif
1986	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,DirtyPageTrap));
1987
1988	Assert(pPteSrc->n.u1Write);
1989
1990	pPteDst->n.u1Write = 1;
1991	pPteDst->n.u1Dirty = 1;
1992	pPteDst->n.u1Accessed = 1;
1993	pPteDst->au32[0] &= ~PGM_PTFLAGS_TRACK_DIRTY;
1994	PGM_INVL_PG(GCPtrPage);
1995
1996	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,DirtyBitTracking), a);
1997	return VINF_PGM_HANDLED_DIRTY_BIT_FAULT;
1998	}
1999	}
2000	else
2001	AssertMsgFailed(("pgmPoolGetPageByHCPhys %VGp failed!\n", pPdeDst->u & SHW_PDE_PG_MASK));
2002	}
2003	}
2004	/** @todo Optimize accessed bit emulation? */
2005	# ifdef VBOX_STRICT
2006	/*
2007	* Sanity check.
2008	*/
2009	else if ( !pPteSrc->n.u1Dirty
2010	&& (pPdeSrc->n.u1Write & pPteSrc->n.u1Write)
2011	&& pPdeDst->n.u1Present)
2012	{
2013	PPGMPOOLPAGE pShwPage = pgmPoolGetPageByHCPhys(pVM, pPdeDst->u & SHW_PDE_PG_MASK);
2014	PSHWPT pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2015	PSHWPTE pPteDst = &pPTDst->a[(GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK];
2016	if ( pPteDst->n.u1Present
2017	&& pPteDst->n.u1Write)
2018	LogFlow(("Writable present page %VGv not marked for dirty bit tracking!!!\n", GCPtrPage));
2019	}
2020	# endif /* VBOX_STRICT */
2021	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,DirtyBitTracking), a);
2022	return VINF_PGM_NO_DIRTY_BIT_TRACKING;
2023	}
2024	AssertRC(rc);
2025	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,DirtyBitTracking), a);
2026	return rc;
2027
2028
2029	UpperLevelPageFault:
2030	/* Pagefault detected while checking the PML4E, PDPE or PDE.
2031	* Single exit handler to get rid of duplicate code paths.
2032	*/
2033	# ifdef IN_GC
2034	STAM_COUNTER_INC(&pVM->pgm.s.StatGCDirtyTrackRealPF);
2035	# endif
2036	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat, DirtyBitTracking), a);
2037	LogFlow(("CheckPageFault: real page fault at %VGv (%d)\n", GCPtrPage, uPageFaultLevel));
2038
2039	if (
2040	# if PGM_GST_TYPE == PGM_TYPE_AMD64
2041	pPml4eSrc->n.u1Present &&
2042	# endif
2043	# if PGM_GST_TYPE == PGM_TYPE_AMD64 \|\| PGM_GST_TYPE == PGM_TYPE_PAE
2044	pPdpeSrc->n.u1Present &&
2045	# endif
2046	pPdeSrc->n.u1Present)
2047	{
2048	/* Check the present bit as the shadow tables can cause different error codes by being out of sync. */
2049	if (pPdeSrc->b.u1Size && fBigPagesSupported)
2050	{
2051	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2052	}
2053	else
2054	{
2055	/*
2056	* Map the guest page table.
2057	*/
2058	PGSTPT pPTSrc;
2059	rc = PGM_GCPHYS_2_PTR(pVM, pPdeSrc->u & GST_PDE_PG_MASK, &pPTSrc);
2060	if (VBOX_SUCCESS(rc))
2061	{
2062	PGSTPTE pPteSrc = &pPTSrc->a[(GCPtrPage >> GST_PT_SHIFT) & GST_PT_MASK];
2063	const GSTPTE PteSrc = *pPteSrc;
2064	if (pPteSrc->n.u1Present)
2065	TRPMSetErrorCode(pVM, uErr \| X86_TRAP_PF_P); /* page-level protection violation */
2066	}
2067	AssertRC(rc);
2068	}
2069	}
2070	return VINF_EM_RAW_GUEST_TRAP;
2071	}
2072
2073	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
2074
2075
2076	/**
2077	* Sync a shadow page table.
2078	*
2079	* The shadow page table is not present. This includes the case where
2080	* there is a conflict with a mapping.
2081	*
2082	* @returns VBox status code.
2083	* @param pVM VM handle.
2084	* @param iPD Page directory index.
2085	* @param pPDSrc Source page directory (i.e. Guest OS page directory).
2086	* Assume this is a temporary mapping.
2087	* @param GCPtrPage GC Pointer of the page that caused the fault
2088	*/
2089	PGM_BTH_DECL(int, SyncPT)(PVM pVM, unsigned iPDSrc, PGSTPD pPDSrc, RTGCUINTPTR GCPtrPage)
2090	{
2091	STAM_PROFILE_START(&pVM->pgm.s.CTXMID(Stat,SyncPT), a);
2092	STAM_COUNTER_INC(&pVM->pgm.s.StatGCSyncPtPD[iPDSrc]);
2093	LogFlow(("SyncPT: GCPtrPage=%VGv\n", GCPtrPage));
2094
2095	#if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
2096	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
2097	&& PGM_SHW_TYPE != PGM_TYPE_NESTED
2098
2099	/*
2100	* Validate input a little bit.
2101	*/
2102	AssertMsg(iPDSrc == ((GCPtrPage >> GST_PD_SHIFT) & GST_PD_MASK), ("iPDSrc=%x GCPtrPage=%VGv\n", iPDSrc, GCPtrPage));
2103	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2104	PX86PD pPDDst = pVM->pgm.s.CTXMID(p,32BitPD);
2105	# else
2106	PX86PDPAE pPDDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0];
2107	# endif
2108	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
2109	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2110	SHWPDE PdeDst = *pPdeDst;
2111
2112	# ifndef PGM_WITHOUT_MAPPINGS
2113	/*
2114	* Check for conflicts.
2115	* GC: In case of a conflict we'll go to Ring-3 and do a full SyncCR3.
2116	* HC: Simply resolve the conflict.
2117	*/
2118	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
2119	{
2120	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
2121	# ifndef IN_RING3
2122	Log(("SyncPT: Conflict at %VGv\n", GCPtrPage));
2123	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,SyncPT), a);
2124	return VERR_ADDRESS_CONFLICT;
2125	# else
2126	PPGMMAPPING pMapping = pgmGetMapping(pVM, (RTGCPTR)GCPtrPage);
2127	Assert(pMapping);
2128	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2129	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2130	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2131	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, GCPtrPage & (GST_PD_MASK << GST_PD_SHIFT));
2132	# endif
2133	if (VBOX_FAILURE(rc))
2134	{
2135	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,SyncPT), a);
2136	return rc;
2137	}
2138	PdeDst = *pPdeDst;
2139	# endif
2140	}
2141	# else /* PGM_WITHOUT_MAPPINGS */
2142	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
2143	# endif /* PGM_WITHOUT_MAPPINGS */
2144	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2145
2146	/*
2147	* Sync page directory entry.
2148	*/
2149	int rc = VINF_SUCCESS;
2150	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2151	if (PdeSrc.n.u1Present)
2152	{
2153	/*
2154	* Allocate & map the page table.
2155	*/
2156	PSHWPT pPTDst;
2157	const bool fPageTable = !PdeSrc.b.u1Size \|\| !(CPUMGetGuestCR4(pVM) & X86_CR4_PSE);
2158	PPGMPOOLPAGE pShwPage;
2159	RTGCPHYS GCPhys;
2160	if (fPageTable)
2161	{
2162	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
2163	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2164	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2165	GCPhys \|= (iPDDst & 1) * (PAGE_SIZE / 2);
2166	# endif
2167	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2168	}
2169	else
2170	{
2171	GCPhys = PdeSrc.u & GST_PDE_BIG_PG_MASK;
2172	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2173	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
2174	GCPhys \|= GCPtrPage & (1 << X86_PD_PAE_SHIFT);
2175	# endif
2176	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_BIG, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2177	}
2178	if (rc == VINF_SUCCESS)
2179	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2180	else if (rc == VINF_PGM_CACHED_PAGE)
2181	{
2182	/*
2183	* The PT was cached, just hook it up.
2184	*/
2185	if (fPageTable)
2186	PdeDst.u = pShwPage->Core.Key
2187	\| (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));
2188	else
2189	{
2190	PdeDst.u = pShwPage->Core.Key
2191	\| (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));
2192	/* (see explanation and assumptions further down.) */
2193	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2194	{
2195	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,DirtyPageBig));
2196	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2197	PdeDst.b.u1Write = 0;
2198	}
2199	}
2200	*pPdeDst = PdeDst;
2201	return VINF_SUCCESS;
2202	}
2203	else if (rc == VERR_PGM_POOL_FLUSHED)
2204	return VINF_PGM_SYNC_CR3;
2205	else
2206	AssertMsgFailedReturn(("rc=%Vrc\n", rc), VERR_INTERNAL_ERROR);
2207	PdeDst.u &= X86_PDE_AVL_MASK;
2208	PdeDst.u \|= pShwPage->Core.Key;
2209
2210	/*
2211	* Page directory has been accessed (this is a fault situation, remember).
2212	*/
2213	pPDSrc->a[iPDSrc].n.u1Accessed = 1;
2214	if (fPageTable)
2215	{
2216	/*
2217	* Page table - 4KB.
2218	*
2219	* Sync all or just a few entries depending on PGM_SYNC_N_PAGES.
2220	*/
2221	Log2(("SyncPT: 4K %VGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx}\n",
2222	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u));
2223	PGSTPT pPTSrc;
2224	rc = PGM_GCPHYS_2_PTR(pVM, PdeSrc.u & GST_PDE_PG_MASK, &pPTSrc);
2225	if (VBOX_SUCCESS(rc))
2226	{
2227	/*
2228	* Start by syncing the page directory entry so CSAM's TLB trick works.
2229	*/
2230	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| X86_PDE_AVL_MASK))
2231	\| (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));
2232	*pPdeDst = PdeDst;
2233
2234	/*
2235	* Directory/page user or supervisor privilege: (same goes for read/write)
2236	*
2237	* Directory Page Combined
2238	* U/S U/S U/S
2239	* 0 0 0
2240	* 0 1 0
2241	* 1 0 0
2242	* 1 1 1
2243	*
2244	* Simple AND operation. Table listed for completeness.
2245	*
2246	*/
2247	STAM_COUNTER_INC(CTXSUFF(&pVM->pgm.s.StatSynPT4k));
2248	# ifdef PGM_SYNC_N_PAGES
2249	unsigned iPTBase = (GCPtrPage >> SHW_PT_SHIFT) & SHW_PT_MASK;
2250	unsigned iPTDst = iPTBase;
2251	const unsigned iPTDstEnd = RT_MIN(iPTDst + PGM_SYNC_NR_PAGES / 2, ELEMENTS(pPTDst->a));
2252	if (iPTDst <= PGM_SYNC_NR_PAGES / 2)
2253	iPTDst = 0;
2254	else
2255	iPTDst -= PGM_SYNC_NR_PAGES / 2;
2256	# else /* !PGM_SYNC_N_PAGES */
2257	unsigned iPTDst = 0;
2258	const unsigned iPTDstEnd = ELEMENTS(pPTDst->a);
2259	# endif /* !PGM_SYNC_N_PAGES */
2260	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2261	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2262	const unsigned offPTSrc = ((GCPtrPage >> SHW_PD_SHIFT) & 1) * 512;
2263	# else
2264	const unsigned offPTSrc = 0;
2265	# endif
2266	for (; iPTDst < iPTDstEnd; iPTDst++)
2267	{
2268	const unsigned iPTSrc = iPTDst + offPTSrc;
2269	const GSTPTE PteSrc = pPTSrc->a[iPTSrc];
2270
2271	if (PteSrc.n.u1Present) /* we've already cleared it above */
2272	{
2273	# ifndef IN_RING0
2274	/*
2275	* Assuming kernel code will be marked as supervisor - and not as user level
2276	* and executed using a conforming code selector - And marked as readonly.
2277	* Also assume that if we're monitoring a page, it's of no interest to CSAM.
2278	*/
2279	PPGMPAGE pPage;
2280	if ( ((PdeSrc.u & pPTSrc->a[iPTSrc].u) & (X86_PTE_RW \| X86_PTE_US))
2281	\|\| !CSAMDoesPageNeedScanning(pVM, (RTRCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)))
2282	\|\| ( (pPage = pgmPhysGetPage(&pVM->pgm.s, PteSrc.u & GST_PTE_PG_MASK))
2283	&& PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2284	)
2285	# endif
2286	PGM_BTH_NAME(SyncPageWorker)(pVM, &pPTDst->a[iPTDst], PdeSrc, PteSrc, pShwPage, iPTDst);
2287	Log2(("SyncPT: 4K+ %VGv PteSrc:{P=%d RW=%d U=%d raw=%08llx}%s dst.raw=%08llx iPTSrc=%x PdeSrc.u=%x physpte=%VGp\n",
2288	(RTGCPTR)((iPDSrc << GST_PD_SHIFT) \| (iPTSrc << PAGE_SHIFT)),
2289	PteSrc.n.u1Present,
2290	PteSrc.n.u1Write & PdeSrc.n.u1Write,
2291	PteSrc.n.u1User & PdeSrc.n.u1User,
2292	(uint64_t)PteSrc.u,
2293	pPTDst->a[iPTDst].u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : "", pPTDst->a[iPTDst].u, iPTSrc, PdeSrc.au32[0],
2294	(PdeSrc.u & GST_PDE_PG_MASK) + iPTSrc*sizeof(PteSrc)));
2295	}
2296	} /* for PTEs */
2297	}
2298	}
2299	else
2300	{
2301	/*
2302	* Big page - 2/4MB.
2303	*
2304	* We'll walk the ram range list in parallel and optimize lookups.
2305	* We will only sync on shadow page table at a time.
2306	*/
2307	STAM_COUNTER_INC(CTXSUFF(&pVM->pgm.s.StatSynPT4M));
2308
2309	/**
2310	* @todo It might be more efficient to sync only a part of the 4MB page (similar to what we do for 4kb PDs).
2311	*/
2312
2313	/*
2314	* Start by syncing the page directory entry.
2315	*/
2316	PdeDst.u = (PdeDst.u & (SHW_PDE_PG_MASK \| (X86_PDE_AVL_MASK & ~PGM_PDFLAGS_TRACK_DIRTY)))
2317	\| (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));
2318
2319	/*
2320	* If the page is not flagged as dirty and is writable, then make it read-only
2321	* at PD level, so we can set the dirty bit when the page is modified.
2322	*
2323	* ASSUMES that page access handlers are implemented on page table entry level.
2324	* Thus we will first catch the dirty access and set PDE.D and restart. If
2325	* there is an access handler, we'll trap again and let it work on the problem.
2326	*/
2327	/** @todo move the above stuff to a section in the PGM documentation. */
2328	Assert(!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY));
2329	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
2330	{
2331	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,DirtyPageBig));
2332	PdeDst.u \|= PGM_PDFLAGS_TRACK_DIRTY;
2333	PdeDst.b.u1Write = 0;
2334	}
2335	*pPdeDst = PdeDst;
2336
2337	/*
2338	* Fill the shadow page table.
2339	*/
2340	/* Get address and flags from the source PDE. */
2341	SHWPTE PteDstBase;
2342	PteDstBase.u = PdeSrc.u & ~(GST_PDE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PAT \| X86_PTE_PCD \| X86_PTE_PWT);
2343
2344	/* Loop thru the entries in the shadow PT. */
2345	const RTGCUINTPTR GCPtr = (GCPtrPage >> SHW_PD_SHIFT) << SHW_PD_SHIFT; NOREF(GCPtr);
2346	Log2(("SyncPT: BIG %VGv PdeSrc:{P=%d RW=%d U=%d raw=%08llx} Shw=%VGv GCPhys=%VGp %s\n",
2347	GCPtrPage, PdeSrc.b.u1Present, PdeSrc.b.u1Write, PdeSrc.b.u1User, (uint64_t)PdeSrc.u, GCPtr,
2348	GCPhys, PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2349	PPGMRAMRANGE pRam = CTXALLSUFF(pVM->pgm.s.pRamRanges);
2350	unsigned iPTDst = 0;
2351	while (iPTDst < ELEMENTS(pPTDst->a))
2352	{
2353	/* Advance ram range list. */
2354	while (pRam && GCPhys > pRam->GCPhysLast)
2355	pRam = CTXALLSUFF(pRam->pNext);
2356	if (pRam && GCPhys >= pRam->GCPhys)
2357	{
2358	unsigned iHCPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2359	do
2360	{
2361	/* Make shadow PTE. */
2362	PPGMPAGE pPage = &pRam->aPages[iHCPage];
2363	SHWPTE PteDst;
2364
2365	/* Make sure the RAM has already been allocated. */
2366	if (pRam->fFlags & MM_RAM_FLAGS_DYNAMIC_ALLOC) /** @todo PAGE FLAGS */
2367	{
2368	if (RT_UNLIKELY(!PGM_PAGE_GET_HCPHYS(pPage)))
2369	{
2370	# ifdef IN_RING3
2371	int rc = pgmr3PhysGrowRange(pVM, GCPhys);
2372	# else
2373	int rc = CTXALLMID(VMM, CallHost)(pVM, VMMCALLHOST_PGM_RAM_GROW_RANGE, GCPhys);
2374	# endif
2375	if (rc != VINF_SUCCESS)
2376	return rc;
2377	}
2378	}
2379
2380	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
2381	{
2382	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
2383	{
2384	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2385	PteDst.n.u1Write = 0;
2386	}
2387	else
2388	PteDst.u = 0;
2389	}
2390	# ifndef IN_RING0
2391	/*
2392	* Assuming kernel code will be marked as supervisor and not as user level and executed
2393	* using a conforming code selector. Don't check for readonly, as that implies the whole
2394	* 4MB can be code or readonly data. Linux enables write access for its large pages.
2395	*/
2396	else if ( !PdeSrc.n.u1User
2397	&& CSAMDoesPageNeedScanning(pVM, (RTRCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT))))
2398	PteDst.u = 0;
2399	# endif
2400	else
2401	PteDst.u = PGM_PAGE_GET_HCPHYS(pPage) \| PteDstBase.u;
2402	# ifdef PGMPOOL_WITH_USER_TRACKING
2403	if (PteDst.n.u1Present)
2404	PGM_BTH_NAME(SyncPageWorkerTrackAddref)(pVM, pShwPage, pPage->HCPhys >> MM_RAM_FLAGS_IDX_SHIFT, pPage, iPTDst); /** @todo PAGE FLAGS */
2405	# endif
2406	/* commit it */
2407	pPTDst->a[iPTDst] = PteDst;
2408	Log4(("SyncPT: BIG %VGv PteDst:{P=%d RW=%d U=%d raw=%08llx}%s\n",
2409	(RTGCPTR)(GCPtr \| (iPTDst << SHW_PT_SHIFT)), PteDst.n.u1Present, PteDst.n.u1Write, PteDst.n.u1User, (uint64_t)PteDst.u,
2410	PteDst.u & PGM_PTFLAGS_TRACK_DIRTY ? " Track-Dirty" : ""));
2411
2412	/* advance */
2413	GCPhys += PAGE_SIZE;
2414	iHCPage++;
2415	iPTDst++;
2416	} while ( iPTDst < ELEMENTS(pPTDst->a)
2417	&& GCPhys <= pRam->GCPhysLast);
2418	}
2419	else if (pRam)
2420	{
2421	Log(("Invalid pages at %VGp\n", GCPhys));
2422	do
2423	{
2424	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2425	GCPhys += PAGE_SIZE;
2426	iPTDst++;
2427	} while ( iPTDst < ELEMENTS(pPTDst->a)
2428	&& GCPhys < pRam->GCPhys);
2429	}
2430	else
2431	{
2432	Log(("Invalid pages at %VGp (2)\n", GCPhys));
2433	for ( ; iPTDst < ELEMENTS(pPTDst->a); iPTDst++)
2434	pPTDst->a[iPTDst].u = 0; /* MMIO or invalid page, we must handle them manually. */
2435	}
2436	} /* while more PTEs */
2437	} /* 4KB / 4MB */
2438	}
2439	else
2440	AssertRelease(!PdeDst.n.u1Present);
2441
2442	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,SyncPT), a);
2443	# ifdef IN_GC
2444	if (VBOX_FAILURE(rc))
2445	STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncPTFailed));
2446	# endif
2447	return rc;
2448
2449	#elif (PGM_GST_TYPE == PGM_TYPE_REAL \|\| PGM_GST_TYPE == PGM_TYPE_PROT) \
2450	&& PGM_SHW_TYPE != PGM_TYPE_NESTED
2451
2452	int rc = VINF_SUCCESS;
2453
2454	/*
2455	* Validate input a little bit.
2456	*/
2457	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2458	PX86PD pPDDst = pVM->pgm.s.CTXMID(p,32BitPD);
2459	# else
2460	PX86PDPAE pPDDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0];
2461	# endif
2462	const unsigned iPDDst = GCPtrPage >> SHW_PD_SHIFT;
2463	PSHWPDE pPdeDst = &pPDDst->a[iPDDst];
2464	SHWPDE PdeDst = *pPdeDst;
2465
2466	Assert(!(PdeDst.u & PGM_PDFLAGS_MAPPING));
2467	Assert(!PdeDst.n.u1Present); /* We're only supposed to call SyncPT on PDE!P and conflicts.*/
2468
2469	GSTPDE PdeSrc;
2470	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2471	PdeSrc.n.u1Present = 1;
2472	PdeSrc.n.u1Write = 1;
2473	PdeSrc.n.u1Accessed = 1;
2474	PdeSrc.n.u1User = 1;
2475
2476	/*
2477	* Allocate & map the page table.
2478	*/
2479	PSHWPT pPTDst;
2480	PPGMPOOLPAGE pShwPage;
2481	RTGCPHYS GCPhys;
2482
2483	/* Virtual address = physical address */
2484	GCPhys = GCPtrPage & X86_PAGE_4K_BASE_MASK_32;
2485	rc = pgmPoolAlloc(pVM, GCPhys, BTH_PGMPOOLKIND_PT_FOR_PT, SHW_POOL_ROOT_IDX, iPDDst, &pShwPage);
2486
2487	if ( rc == VINF_SUCCESS
2488	\|\| rc == VINF_PGM_CACHED_PAGE)
2489	pPTDst = (PSHWPT)PGMPOOL_PAGE_2_PTR(pVM, pShwPage);
2490	else
2491	AssertMsgFailedReturn(("rc=%Vrc\n", rc), VERR_INTERNAL_ERROR);
2492
2493	PdeDst.u &= X86_PDE_AVL_MASK;
2494	PdeDst.u \|= pShwPage->Core.Key;
2495	PdeDst.n.u1Present = 1;
2496	PdeDst.n.u1Write = 1;
2497	PdeDst.n.u1User = 1;
2498	*pPdeDst = PdeDst;
2499
2500	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, (RTGCUINTPTR)GCPtrPage, PGM_SYNC_NR_PAGES, 0 /* page not present */);
2501	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,SyncPT), a);
2502	return rc;
2503
2504	#else /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
2505	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
2506	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,SyncPT), a);
2507	return VERR_INTERNAL_ERROR;
2508	#endif /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
2509	}
2510
2511
2512
2513	/**
2514	* Prefetch a page/set of pages.
2515	*
2516	* Typically used to sync commonly used pages before entering raw mode
2517	* after a CR3 reload.
2518	*
2519	* @returns VBox status code.
2520	* @param pVM VM handle.
2521	* @param GCPtrPage Page to invalidate.
2522	*/
2523	PGM_BTH_DECL(int, PrefetchPage)(PVM pVM, RTGCUINTPTR GCPtrPage)
2524	{
2525	#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) \
2526	&& PGM_SHW_TYPE != PGM_TYPE_AMD64 && PGM_SHW_TYPE != PGM_TYPE_NESTED
2527	/*
2528	* Check that all Guest levels thru the PDE are present, getting the
2529	* PD and PDE in the processes.
2530	*/
2531	int rc = VINF_SUCCESS;
2532	# if PGM_WITH_PAGING(PGM_GST_TYPE)
2533	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2534	const unsigned iPDSrc = (RTGCUINTPTR)GCPtrPage >> GST_PD_SHIFT;
2535	PGSTPD pPDSrc = CTXSUFF(pVM->pgm.s.pGuestPD);
2536	# else /* PAE */
2537	unsigned iPDSrc;
2538	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc);
2539	if (!pPDSrc)
2540	return VINF_SUCCESS; /* not present */
2541	# endif
2542	const GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2543	# else
2544	PGSTPD pPDSrc = NULL;
2545	const unsigned iPDSrc = 0;
2546	GSTPDE PdeSrc;
2547
2548	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2549	PdeSrc.n.u1Present = 1;
2550	PdeSrc.n.u1Write = 1;
2551	PdeSrc.n.u1Accessed = 1;
2552	PdeSrc.n.u1User = 1;
2553	# endif
2554
2555	if (PdeSrc.n.u1Present && PdeSrc.n.u1Accessed)
2556	{
2557	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2558	const X86PDE PdeDst = pVM->pgm.s.CTXMID(p,32BitPD)->a[GCPtrPage >> SHW_PD_SHIFT];
2559	# else
2560	const X86PDEPAE PdeDst = pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[GCPtrPage >> SHW_PD_SHIFT];
2561	# endif
2562	if (!(PdeDst.u & PGM_PDFLAGS_MAPPING))
2563	{
2564	if (!PdeDst.n.u1Present)
2565	/** r=bird: This guy will set the A bit on the PDE, probably harmless. */
2566	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
2567	else
2568	{
2569	/** @note We used to sync PGM_SYNC_NR_PAGES pages, which triggered assertions in CSAM, because
2570	* R/W attributes of nearby pages were reset. Not sure how that could happen. Anyway, it
2571	* makes no sense to prefetch more than one page.
2572	*/
2573	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
2574	if (VBOX_SUCCESS(rc))
2575	rc = VINF_SUCCESS;
2576	}
2577	}
2578	}
2579	return rc;
2580	#elif PGM_SHW_TYPE == PGM_TYPE_NESTED
2581	return VINF_SUCCESS; /* ignore */
2582	#else /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
2583
2584	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_SHW_TYPE, PGM_GST_TYPE));
2585	return VERR_INTERNAL_ERROR;
2586	#endif /* PGM_GST_TYPE == PGM_TYPE_AMD64 */
2587	}
2588
2589
2590
2591
2592	/**
2593	* Syncs a page during a PGMVerifyAccess() call.
2594	*
2595	* @returns VBox status code (informational included).
2596	* @param GCPtrPage The address of the page to sync.
2597	* @param fPage The effective guest page flags.
2598	* @param uErr The trap error code.
2599	*/
2600	PGM_BTH_DECL(int, VerifyAccessSyncPage)(PVM pVM, RTGCUINTPTR GCPtrPage, unsigned fPage, unsigned uErr)
2601	{
2602	LogFlow(("VerifyAccessSyncPage: GCPtrPage=%VGv fPage=%#x uErr=%#x\n", GCPtrPage, fPage, uErr));
2603
2604	Assert(!HWACCMIsNestedPagingActive(pVM));
2605	#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) \
2606	&& PGM_SHW_TYPE != PGM_TYPE_AMD64 && PGM_SHW_TYPE != PGM_TYPE_NESTED
2607
2608	# ifndef IN_RING0
2609	if (!(fPage & X86_PTE_US))
2610	{
2611	/*
2612	* Mark this page as safe.
2613	*/
2614	/** @todo not correct for pages that contain both code and data!! */
2615	Log(("CSAMMarkPage %VGv; scanned=%d\n", GCPtrPage, true));
2616	CSAMMarkPage(pVM, (RTRCPTR)GCPtrPage, true);
2617	}
2618	# endif
2619	/*
2620	* Get guest PD and index.
2621	*/
2622
2623	# if PGM_WITH_PAGING(PGM_GST_TYPE)
2624	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2625	const unsigned iPDSrc = (RTGCUINTPTR)GCPtrPage >> GST_PD_SHIFT;
2626	PGSTPD pPDSrc = CTXSUFF(pVM->pgm.s.pGuestPD);
2627	# else /* PAE */
2628	unsigned iPDSrc;
2629	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, GCPtrPage, &iPDSrc);
2630
2631	if (pPDSrc)
2632	{
2633	Log(("PGMVerifyAccess: access violation for %VGv due to non-present PDPTR\n", GCPtrPage));
2634	return VINF_EM_RAW_GUEST_TRAP;
2635	}
2636	# endif
2637	# else
2638	PGSTPD pPDSrc = NULL;
2639	const unsigned iPDSrc = 0;
2640	# endif
2641	int rc = VINF_SUCCESS;
2642
2643	/*
2644	* First check if the shadow pd is present.
2645	*/
2646	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2647	PX86PDE pPdeDst = &pVM->pgm.s.CTXMID(p,32BitPD)->a[GCPtrPage >> SHW_PD_SHIFT];
2648	# else
2649	PX86PDEPAE pPdeDst = &pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[GCPtrPage >> SHW_PD_SHIFT];
2650	# endif
2651	if (!pPdeDst->n.u1Present)
2652	{
2653	rc = PGM_BTH_NAME(SyncPT)(pVM, iPDSrc, pPDSrc, GCPtrPage);
2654	AssertRC(rc);
2655	if (rc != VINF_SUCCESS)
2656	return rc;
2657	}
2658
2659	# if PGM_WITH_PAGING(PGM_GST_TYPE)
2660	/* Check for dirty bit fault */
2661	rc = PGM_BTH_NAME(CheckPageFault)(pVM, uErr, pPdeDst, &pPDSrc->a[iPDSrc], GCPtrPage);
2662	if (rc == VINF_PGM_HANDLED_DIRTY_BIT_FAULT)
2663	Log(("PGMVerifyAccess: success (dirty)\n"));
2664	else
2665	{
2666	GSTPDE PdeSrc = pPDSrc->a[iPDSrc];
2667	#else
2668	{
2669	GSTPDE PdeSrc;
2670	PdeSrc.au32[0] = 0; /* faked so we don't have to #ifdef everything */
2671	PdeSrc.n.u1Present = 1;
2672	PdeSrc.n.u1Write = 1;
2673	PdeSrc.n.u1Accessed = 1;
2674	PdeSrc.n.u1User = 1;
2675
2676	#endif /* PGM_WITH_PAGING(PGM_GST_TYPE) */
2677	Assert(rc != VINF_EM_RAW_GUEST_TRAP);
2678	if (uErr & X86_TRAP_PF_US)
2679	STAM_COUNTER_INC(&pVM->pgm.s.StatGCPageOutOfSyncUser);
2680	else /* supervisor */
2681	STAM_COUNTER_INC(&pVM->pgm.s.StatGCPageOutOfSyncSupervisor);
2682
2683	rc = PGM_BTH_NAME(SyncPage)(pVM, PdeSrc, GCPtrPage, 1, 0);
2684	if (VBOX_SUCCESS(rc))
2685	{
2686	/* Page was successfully synced */
2687	Log2(("PGMVerifyAccess: success (sync)\n"));
2688	rc = VINF_SUCCESS;
2689	}
2690	else
2691	{
2692	Log(("PGMVerifyAccess: access violation for %VGv rc=%d\n", GCPtrPage, rc));
2693	return VINF_EM_RAW_GUEST_TRAP;
2694	}
2695	}
2696	return rc;
2697
2698	#else /* PGM_GST_TYPE != PGM_TYPE_32BIT */
2699
2700	AssertReleaseMsgFailed(("Shw=%d Gst=%d is not implemented!\n", PGM_GST_TYPE, PGM_SHW_TYPE));
2701	return VERR_INTERNAL_ERROR;
2702	#endif /* PGM_GST_TYPE != PGM_TYPE_32BIT */
2703	}
2704
2705
2706	#if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE
2707	# if PGM_SHW_TYPE == PGM_TYPE_32BIT \|\| PGM_SHW_TYPE == PGM_TYPE_PAE
2708	/**
2709	* Figures out which kind of shadow page this guest PDE warrants.
2710	*
2711	* @returns Shadow page kind.
2712	* @param pPdeSrc The guest PDE in question.
2713	* @param cr4 The current guest cr4 value.
2714	*/
2715	DECLINLINE(PGMPOOLKIND) PGM_BTH_NAME(CalcPageKind)(const GSTPDE *pPdeSrc, uint32_t cr4)
2716	{
2717	if (!pPdeSrc->n.u1Size \|\| !(cr4 & X86_CR4_PSE))
2718	return BTH_PGMPOOLKIND_PT_FOR_PT;
2719	//switch (pPdeSrc->u & (X86_PDE4M_RW \| X86_PDE4M_US /\| X86_PDE4M_PAE_NX/))
2720	//{
2721	// case 0:
2722	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RO;
2723	// case X86_PDE4M_RW:
2724	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW;
2725	// case X86_PDE4M_US:
2726	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US;
2727	// case X86_PDE4M_RW \| X86_PDE4M_US:
2728	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US;
2729	# if 0
2730	// case X86_PDE4M_PAE_NX:
2731	// return BTH_PGMPOOLKIND_PT_FOR_BIG_NX;
2732	// case X86_PDE4M_RW \| X86_PDE4M_PAE_NX:
2733	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_NX;
2734	// case X86_PDE4M_US \| X86_PDE4M_PAE_NX:
2735	// return BTH_PGMPOOLKIND_PT_FOR_BIG_US_NX;
2736	// case X86_PDE4M_RW \| X86_PDE4M_US \| X86_PDE4M_PAE_NX:
2737	// return BTH_PGMPOOLKIND_PT_FOR_BIG_RW_US_NX;
2738	# endif
2739	return BTH_PGMPOOLKIND_PT_FOR_BIG;
2740	//}
2741	}
2742	# endif
2743	#endif
2744
2745	#undef MY_STAM_COUNTER_INC
2746	#define MY_STAM_COUNTER_INC(a) do { } while (0)
2747
2748
2749	/**
2750	* Syncs the paging hierarchy starting at CR3.
2751	*
2752	* @returns VBox status code, no specials.
2753	* @param pVM The virtual machine.
2754	* @param cr0 Guest context CR0 register
2755	* @param cr3 Guest context CR3 register
2756	* @param cr4 Guest context CR4 register
2757	* @param fGlobal Including global page directories or not
2758	*/
2759	PGM_BTH_DECL(int, SyncCR3)(PVM pVM, uint64_t cr0, uint64_t cr3, uint64_t cr4, bool fGlobal)
2760	{
2761	#if PGM_SHW_TYPE == PGM_TYPE_NESTED
2762	/** @todo check if this is really necessary */
2763	HWACCMFlushTLB(pVM);
2764	return VINF_SUCCESS;
2765
2766	#else /* PGM_SHW_TYPE != PGM_TYPE_NESTED */
2767	if (VM_FF_ISSET(pVM, VM_FF_PGM_SYNC_CR3))
2768	fGlobal = true; /* Change this CR3 reload to be a global one. */
2769
2770	/*
2771	* Update page access handlers.
2772	* The virtual are always flushed, while the physical are only on demand.
2773	* WARNING: We are incorrectly not doing global flushing on Virtual Handler updates. We'll
2774	* have to look into that later because it will have a bad influence on the performance.
2775	* @note SvL: There's no need for that. Just invalidate the virtual range(s).
2776	* bird: Yes, but that won't work for aliases.
2777	*/
2778	/** @todo this MUST go away. See #1557. */
2779	STAM_PROFILE_START(&pVM->pgm.s.CTXMID(Stat,SyncCR3Handlers), h);
2780	PGM_GST_NAME(HandlerVirtualUpdate)(pVM, cr4);
2781	STAM_PROFILE_STOP(&pVM->pgm.s.CTXMID(Stat,SyncCR3Handlers), h);
2782
2783	# ifdef PGMPOOL_WITH_MONITORING
2784	/*
2785	* When monitoring shadowed pages, we reset the modification counters on CR3 sync.
2786	* Occationally we will have to clear all the shadow page tables because we wanted
2787	* to monitor a page which was mapped by too many shadowed page tables. This operation
2788	* sometimes refered to as a 'lightweight flush'.
2789	*/
2790	if (!(pVM->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL))
2791	pgmPoolMonitorModifiedClearAll(pVM);
2792	else
2793	{
2794	# ifdef IN_RING3
2795	pVM->pgm.s.fSyncFlags &= ~PGM_SYNC_CLEAR_PGM_POOL;
2796	pgmPoolClearAll(pVM);
2797	# else
2798	LogFlow(("SyncCR3: PGM_SYNC_CLEAR_PGM_POOL is set -> VINF_PGM_SYNC_CR3\n"));
2799	return VINF_PGM_SYNC_CR3;
2800	# endif
2801	}
2802	# endif
2803
2804	Assert(fGlobal \|\| (cr4 & X86_CR4_PGE));
2805	MY_STAM_COUNTER_INC(fGlobal ? &pVM->pgm.s.CTXMID(Stat,SyncCR3Global) : &pVM->pgm.s.CTXMID(Stat,SyncCR3NotGlobal));
2806
2807	# if PGM_GST_TYPE == PGM_TYPE_32BIT \|\| PGM_GST_TYPE == PGM_TYPE_PAE
2808	/*
2809	* Get page directory addresses.
2810	*/
2811	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2812	PX86PDE pPDEDst = &pVM->pgm.s.CTXMID(p,32BitPD)->a[0];
2813	# else /* PGM_SHW_TYPE == PGM_TYPE_PAE */
2814	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2815	PX86PDEPAE pPDEDst = &pVM->pgm.s.CTXMID(ap,PaePDs)[0]->a[0];
2816	# endif
2817	# endif
2818
2819	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2820	PGSTPD pPDSrc = CTXSUFF(pVM->pgm.s.pGuestPD);
2821	Assert(pPDSrc);
2822	# ifndef IN_GC
2823	Assert(MMPhysGCPhys2HCVirt(pVM, (RTGCPHYS)(cr3 & GST_CR3_PAGE_MASK), sizeof(*pPDSrc)) == pPDSrc);
2824	# endif
2825	# endif
2826
2827	/*
2828	* Iterate the page directory.
2829	*/
2830	PPGMMAPPING pMapping;
2831	unsigned iPdNoMapping;
2832	const bool fRawR0Enabled = EMIsRawRing0Enabled(pVM);
2833	PPGMPOOL pPool = pVM->pgm.s.CTXSUFF(pPool);
2834
2835	/* Only check mappings if they are supposed to be put into the shadow page table. */
2836	if (pgmMapAreMappingsEnabled(&pVM->pgm.s))
2837	{
2838	pMapping = pVM->pgm.s.CTXALLSUFF(pMappings);
2839	iPdNoMapping = (pMapping) ? (pMapping->GCPtr >> GST_PD_SHIFT) : ~0U;
2840	}
2841	else
2842	{
2843	pMapping = 0;
2844	iPdNoMapping = ~0U;
2845	}
2846	# if PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
2847	for (unsigned iPDPTE = 0; iPDPTE < GST_PDPE_ENTRIES; iPDPTE++)
2848	{
2849	unsigned iPDSrc;
2850	# if PGM_SHW_TYPE == PGM_TYPE_PAE
2851	PX86PDPAE pPDPAE = pVM->pgm.s.CTXMID(ap,PaePDs)[0];
2852	# else
2853	AssertFailed(); /* @todo */
2854	PX86PDPE pPDPAE = pVM->pgm.s.CTXMID(ap,PaePDs)[iPDPTE * X86_PG_AMD64_ENTRIES];
2855	# endif
2856	PX86PDEPAE pPDEDst = &pPDPAE->a[iPDPTE * X86_PG_PAE_ENTRIES];
2857	PGSTPD pPDSrc = pgmGstGetPaePDPtr(&pVM->pgm.s, iPDPTE << X86_PDPT_SHIFT, &iPDSrc);
2858
2859	if (pPDSrc == NULL)
2860	{
2861	/* PDPT not present */
2862	if (pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPDPTE].n.u1Present)
2863	{
2864	for (unsigned iPD = 0; iPD < ELEMENTS(pPDSrc->a); iPD++)
2865	{
2866	if ( pPDEDst[iPD].n.u1Present
2867	&& !(pPDEDst[iPD].u & PGM_PDFLAGS_MAPPING))
2868	{
2869	pgmPoolFreeByPage(pPool, pgmPoolGetPage(pPool, pPDEDst[iPD].u & SHW_PDE_PG_MASK), SHW_POOL_ROOT_IDX, iPDPTE * X86_PG_PAE_ENTRIES + iPD);
2870	pPDEDst[iPD].u = 0;
2871	}
2872	}
2873	}
2874	if (!(pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPDPTE].u & PGM_PLXFLAGS_MAPPING))
2875	pVM->pgm.s.CTXMID(p,PaePDPT)->a[iPDPTE].n.u1Present = 0;
2876	continue;
2877	}
2878	# else /* PGM_GST_TYPE != PGM_TYPE_PAE && PGM_GST_TYPE != PGM_TYPE_AMD64 */
2879	{
2880	# endif /* PGM_GST_TYPE != PGM_TYPE_PAE && PGM_GST_TYPE != PGM_TYPE_AMD64 */
2881	for (unsigned iPD = 0; iPD < ELEMENTS(pPDSrc->a); iPD++)
2882	{
2883	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
2884	Assert(&pVM->pgm.s.CTXMID(p,32BitPD)->a[iPD] == pPDEDst);
2885	# elif PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2886	AssertMsg(&pVM->pgm.s.CTXMID(ap,PaePDs)[iPD * 2 / 512]->a[iPD * 2 % 512] == pPDEDst, ("%p vs %p\n", &pVM->pgm.s.CTXMID(ap,PaePDs)[iPD * 2 / 512]->a[iPD * 2 % 512], pPDEDst));
2887	# endif
2888	register GSTPDE PdeSrc = pPDSrc->a[iPD];
2889	if ( PdeSrc.n.u1Present
2890	&& (PdeSrc.n.u1User \|\| fRawR0Enabled))
2891	{
2892	# if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
2893	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
2894	&& !defined(PGM_WITHOUT_MAPPINGS)
2895
2896	/*
2897	* Check for conflicts with GC mappings.
2898	*/
2899	# if PGM_GST_TYPE == PGM_TYPE_PAE
2900	if (iPD + iPDPTE * X86_PG_PAE_ENTRIES == iPdNoMapping)
2901	# else
2902	if (iPD == iPdNoMapping)
2903	# endif
2904	{
2905	if (pVM->pgm.s.fMappingsFixed)
2906	{
2907	/* It's fixed, just skip the mapping. */
2908	const unsigned cPTs = pMapping->cb >> GST_PD_SHIFT;
2909	iPD += cPTs - 1;
2910	pPDEDst += cPTs + (PGM_GST_TYPE != PGM_SHW_TYPE) * cPTs; /* Only applies to the pae shadow and 32 bits guest case */
2911	pMapping = pMapping->CTXALLSUFF(pNext);
2912	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
2913	continue;
2914	}
2915	# ifdef IN_RING3
2916	# if PGM_GST_TYPE == PGM_TYPE_32BIT
2917	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, iPD << GST_PD_SHIFT);
2918	# elif PGM_GST_TYPE == PGM_TYPE_PAE
2919	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, (iPDPTE << GST_PDPT_SHIFT) + (iPD << GST_PD_SHIFT));
2920	# endif
2921	if (VBOX_FAILURE(rc))
2922	return rc;
2923
2924	/*
2925	* Update iPdNoMapping and pMapping.
2926	*/
2927	pMapping = pVM->pgm.s.pMappingsR3;
2928	while (pMapping && pMapping->GCPtr < (iPD << GST_PD_SHIFT))
2929	pMapping = pMapping->pNextR3;
2930	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
2931	# else
2932	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
2933	return VINF_PGM_SYNC_CR3;
2934	# endif
2935	}
2936	# else /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
2937	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
2938	# endif /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
2939	/*
2940	* Sync page directory entry.
2941	*
2942	* The current approach is to allocated the page table but to set
2943	* the entry to not-present and postpone the page table synching till
2944	* it's actually used.
2945	*/
2946	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2947	for (unsigned i = 0, iPdShw = iPD * 2; i < 2; i++, iPdShw++) /* pray that the compiler unrolls this */
2948	# elif PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
2949	const unsigned iPdShw = iPD + iPDPTE * X86_PG_PAE_ENTRIES; NOREF(iPdShw);
2950	# else
2951	const unsigned iPdShw = iPD; NOREF(iPdShw);
2952	# endif
2953	{
2954	SHWPDE PdeDst = *pPDEDst;
2955	if (PdeDst.n.u1Present)
2956	{
2957	PPGMPOOLPAGE pShwPage = pgmPoolGetPage(pPool, PdeDst.u & SHW_PDE_PG_MASK);
2958	RTGCPHYS GCPhys;
2959	if ( !PdeSrc.b.u1Size
2960	\|\| !(cr4 & X86_CR4_PSE))
2961	{
2962	GCPhys = PdeSrc.u & GST_PDE_PG_MASK;
2963	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2964	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
2965	GCPhys \|= i * (PAGE_SIZE / 2);
2966	# endif
2967	}
2968	else
2969	{
2970	GCPhys = PdeSrc.u & GST_PDE_BIG_PG_MASK;
2971	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
2972	/* Select the right PDE as we're emulating a 4MB page directory with two 2 MB shadow PDEs.*/
2973	GCPhys \|= i * X86_PAGE_2M_SIZE;
2974	# endif
2975	}
2976
2977	if ( pShwPage->GCPhys == GCPhys
2978	&& pShwPage->enmKind == PGM_BTH_NAME(CalcPageKind)(&PdeSrc, cr4)
2979	&& ( pShwPage->fCached
2980	\|\| ( !fGlobal
2981	&& ( false
2982	# ifdef PGM_SKIP_GLOBAL_PAGEDIRS_ON_NONGLOBAL_FLUSH
2983	\|\| ( (PdeSrc.u & (X86_PDE4M_PS \| X86_PDE4M_G)) == (X86_PDE4M_PS \| X86_PDE4M_G)
2984	&& (cr4 & (X86_CR4_PGE \| X86_CR4_PSE)) == (X86_CR4_PGE \| X86_CR4_PSE)) /* global 2/4MB page. */
2985	\|\| ( !pShwPage->fSeenNonGlobal
2986	&& (cr4 & X86_CR4_PGE))
2987	# endif
2988	)
2989	)
2990	)
2991	&& ( (PdeSrc.u & (X86_PDE_US \| X86_PDE_RW)) == (PdeDst.u & (X86_PDE_US \| X86_PDE_RW))
2992	\|\| ( (cr4 & X86_CR4_PSE)
2993	&& ((PdeSrc.u & (X86_PDE_US \| X86_PDE4M_PS \| X86_PDE4M_D)) \| PGM_PDFLAGS_TRACK_DIRTY)
2994	== ((PdeDst.u & (X86_PDE_US \| X86_PDE_RW \| PGM_PDFLAGS_TRACK_DIRTY)) \| X86_PDE4M_PS))
2995	)
2996	)
2997	{
2998	# ifdef VBOX_WITH_STATISTICS
2999	if ( !fGlobal
3000	&& (PdeSrc.u & (X86_PDE4M_PS \| X86_PDE4M_G)) == (X86_PDE4M_PS \| X86_PDE4M_G)
3001	&& (cr4 & (X86_CR4_PGE \| X86_CR4_PSE)) == (X86_CR4_PGE \| X86_CR4_PSE))
3002	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncCR3DstSkippedGlobalPD));
3003	else if (!fGlobal && !pShwPage->fSeenNonGlobal && (cr4 & X86_CR4_PGE))
3004	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncCR3DstSkippedGlobalPT));
3005	else
3006	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncCR3DstCacheHit));
3007	# endif /* VBOX_WITH_STATISTICS */
3008	/** @todo a replacement strategy isn't really needed unless we're using a very small pool < 512 pages.
3009	* The whole ageing stuff should be put in yet another set of #ifdefs. For now, let's just skip it. */
3010	//# ifdef PGMPOOL_WITH_CACHE
3011	// pgmPoolCacheUsed(pPool, pShwPage);
3012	//# endif
3013	}
3014	else
3015	{
3016	pgmPoolFreeByPage(pPool, pShwPage, SHW_POOL_ROOT_IDX, iPdShw);
3017	pPDEDst->u = 0;
3018	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncCR3DstFreed));
3019	}
3020	}
3021	else
3022	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncCR3DstNotPresent));
3023	pPDEDst++;
3024	}
3025	}
3026	# if PGM_GST_TYPE == PGM_TYPE_PAE
3027	else if (iPD + iPDPTE * X86_PG_PAE_ENTRIES != iPdNoMapping)
3028	# else
3029	else if (iPD != iPdNoMapping)
3030	# endif
3031	{
3032	/*
3033	* Check if there is any page directory to mark not present here.
3034	*/
3035	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3036	for (unsigned i = 0, iPdShw = iPD * 2; i < 2; i++, iPdShw++) /* pray that the compiler unrolls this */
3037	# elif PGM_GST_TYPE == PGM_TYPE_PAE \|\| PGM_GST_TYPE == PGM_TYPE_AMD64
3038	const unsigned iPdShw = iPD + iPDPTE * X86_PG_PAE_ENTRIES; NOREF(iPdShw);
3039	# else
3040	const unsigned iPdShw = iPD; NOREF(iPdShw);
3041	# endif
3042	{
3043	if (pPDEDst->n.u1Present)
3044	{
3045	pgmPoolFreeByPage(pPool, pgmPoolGetPage(pPool, pPDEDst->u & SHW_PDE_PG_MASK), SHW_POOL_ROOT_IDX, iPdShw);
3046	pPDEDst->u = 0;
3047	MY_STAM_COUNTER_INC(&pVM->pgm.s.CTXMID(Stat,SyncCR3DstFreedSrcNP));
3048	}
3049	pPDEDst++;
3050	}
3051	}
3052	else
3053	{
3054	# if ( PGM_GST_TYPE == PGM_TYPE_32BIT \
3055	\|\| PGM_GST_TYPE == PGM_TYPE_PAE) \
3056	&& !defined(PGM_WITHOUT_MAPPINGS)
3057
3058	const unsigned cPTs = pMapping->cb >> GST_PD_SHIFT;
3059
3060	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3061	if (pVM->pgm.s.fMappingsFixed)
3062	{
3063	/* It's fixed, just skip the mapping. */
3064	pMapping = pMapping->CTXALLSUFF(pNext);
3065	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3066	}
3067	else
3068	{
3069	/*
3070	* Check for conflicts for subsequent pagetables
3071	* and advance to the next mapping.
3072	*/
3073	iPdNoMapping = ~0U;
3074	unsigned iPT = cPTs;
3075	while (iPT-- > 1)
3076	{
3077	if ( pPDSrc->a[iPD + iPT].n.u1Present
3078	&& (pPDSrc->a[iPD + iPT].n.u1User \|\| fRawR0Enabled))
3079	{
3080	# ifdef IN_RING3
3081	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3082	int rc = pgmR3SyncPTResolveConflict(pVM, pMapping, pPDSrc, iPD << GST_PD_SHIFT);
3083	# elif PGM_GST_TYPE == PGM_TYPE_PAE
3084	int rc = pgmR3SyncPTResolveConflictPAE(pVM, pMapping, (iPDPTE << GST_PDPT_SHIFT) + (iPD << GST_PD_SHIFT));
3085	# endif
3086	if (VBOX_FAILURE(rc))
3087	return rc;
3088
3089	/*
3090	* Update iPdNoMapping and pMapping.
3091	*/
3092	pMapping = pVM->pgm.s.CTXALLSUFF(pMappings);
3093	while (pMapping && pMapping->GCPtr < (iPD << GST_PD_SHIFT))
3094	pMapping = pMapping->CTXALLSUFF(pNext);
3095	iPdNoMapping = pMapping ? pMapping->GCPtr >> GST_PD_SHIFT : ~0U;
3096	break;
3097	# else
3098	LogFlow(("SyncCR3: detected conflict -> VINF_PGM_SYNC_CR3\n"));
3099	return VINF_PGM_SYNC_CR3;
3100	# endif
3101	}
3102	}
3103	if (iPdNoMapping == ~0U && pMapping)
3104	{
3105	pMapping = pMapping->CTXALLSUFF(pNext);
3106	if (pMapping)
3107	iPdNoMapping = pMapping->GCPtr >> GST_PD_SHIFT;
3108	}
3109	}
3110
3111	/* advance. */
3112	iPD += cPTs - 1;
3113	pPDEDst += cPTs + (PGM_GST_TYPE != PGM_SHW_TYPE) * cPTs; /* Only applies to the pae shadow and 32 bits guest case */
3114	# if PGM_GST_TYPE != PGM_SHW_TYPE
3115	AssertCompile(PGM_GST_TYPE == PGM_TYPE_32BIT && PGM_SHW_TYPE == PGM_TYPE_PAE);
3116	# endif
3117	# else /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3118	Assert(!pgmMapAreMappingsEnabled(&pVM->pgm.s));
3119	# endif /* (PGM_GST_TYPE != PGM_TYPE_32BIT && PGM_GST_TYPE != PGM_TYPE_PAE) \|\| PGM_WITHOUT_MAPPINGS */
3120	}
3121
3122	} /* for iPD */
3123	} /* for each PDPTE (PAE) */
3124
3125	return VINF_SUCCESS;
3126
3127	# elif PGM_GST_TYPE == PGM_TYPE_AMD64
3128	//# error not implemented
3129	return VERR_INTERNAL_ERROR;
3130	# else /* guest real and protected mode */
3131	return VINF_SUCCESS;
3132	# endif
3133	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED */
3134	}
3135
3136
3137
3138
3139	#ifdef VBOX_STRICT
3140	#ifdef IN_GC
3141	# undef AssertMsgFailed
3142	# define AssertMsgFailed Log
3143	#endif
3144	#ifdef IN_RING3
3145	# include <VBox/dbgf.h>
3146
3147	/**
3148	* Dumps a page table hierarchy use only physical addresses and cr4/lm flags.
3149	*
3150	* @returns VBox status code (VINF_SUCCESS).
3151	* @param pVM The VM handle.
3152	* @param cr3 The root of the hierarchy.
3153	* @param crr The cr4, only PAE and PSE is currently used.
3154	* @param fLongMode Set if long mode, false if not long mode.
3155	* @param cMaxDepth Number of levels to dump.
3156	* @param pHlp Pointer to the output functions.
3157	*/
3158	__BEGIN_DECLS
3159	PGMR3DECL(int) PGMR3DumpHierarchyHC(PVM pVM, uint32_t cr3, uint32_t cr4, bool fLongMode, unsigned cMaxDepth, PCDBGFINFOHLP pHlp);
3160	__END_DECLS
3161
3162	#endif
3163
3164	/**
3165	* Checks that the shadow page table is in sync with the guest one.
3166	*
3167	* @returns The number of errors.
3168	* @param pVM The virtual machine.
3169	* @param cr3 Guest context CR3 register
3170	* @param cr4 Guest context CR4 register
3171	* @param GCPtr Where to start. Defaults to 0.
3172	* @param cb How much to check. Defaults to everything.
3173	*/
3174	PGM_BTH_DECL(unsigned, AssertCR3)(PVM pVM, uint64_t cr3, uint64_t cr4, RTGCUINTPTR GCPtr, RTGCUINTPTR cb)
3175	{
3176	#if PGM_SHW_TYPE == PGM_TYPE_NESTED
3177	return 0;
3178	#else
3179	unsigned cErrors = 0;
3180
3181	#if PGM_GST_TYPE == PGM_TYPE_32BIT \
3182	\|\| PGM_GST_TYPE == PGM_TYPE_PAE
3183
3184	PPGM pPGM = &pVM->pgm.s;
3185	RTGCPHYS GCPhysGst; /* page address derived from the guest page tables. */
3186	RTHCPHYS HCPhysShw; /* page address derived from the shadow page tables. */
3187	# ifndef IN_RING0
3188	RTHCPHYS HCPhys; /* general usage. */
3189	# endif
3190	int rc;
3191
3192	/*
3193	* Check that the Guest CR3 and all its mappings are correct.
3194	*/
3195	AssertMsgReturn(pPGM->GCPhysCR3 == (cr3 & GST_CR3_PAGE_MASK),
3196	("Invalid GCPhysCR3=%VGp cr3=%VGp\n", pPGM->GCPhysCR3, (RTGCPHYS)cr3),
3197	false);
3198	# ifndef IN_RING0
3199	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3200	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGuestPDGC, NULL, &HCPhysShw);
3201	# else
3202	rc = PGMShwGetPage(pVM, (RTGCPTR)pPGM->pGstPaePDPTGC, NULL, &HCPhysShw);
3203	# endif
3204	AssertRCReturn(rc, 1);
3205	HCPhys = NIL_RTHCPHYS;
3206	rc = pgmRamGCPhys2HCPhys(pPGM, cr3 & GST_CR3_PAGE_MASK, &HCPhys);
3207	AssertMsgReturn(HCPhys == HCPhysShw, ("HCPhys=%VHp HCPhyswShw=%VHp (cr3)\n", HCPhys, HCPhysShw), false);
3208	# if PGM_GST_TYPE == PGM_TYPE_32BIT && defined(IN_RING3)
3209	RTGCPHYS GCPhys;
3210	rc = PGMR3DbgHCPtr2GCPhys(pVM, pPGM->pGuestPDHC, &GCPhys);
3211	AssertRCReturn(rc, 1);
3212	AssertMsgReturn((cr3 & GST_CR3_PAGE_MASK) == GCPhys, ("GCPhys=%VGp cr3=%VGp\n", GCPhys, (RTGCPHYS)cr3), false);
3213	# endif
3214	#endif /* !IN_RING0 */
3215
3216	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3217	const GSTPD *pPDSrc = CTXSUFF(pPGM->pGuestPD);
3218	# endif
3219
3220	/*
3221	* Get and check the Shadow CR3.
3222	*/
3223	# if PGM_SHW_TYPE == PGM_TYPE_32BIT
3224	const X86PD *pPDDst = pPGM->CTXMID(p,32BitPD);
3225	unsigned cPDEs = ELEMENTS(pPDDst->a);
3226	# else
3227	const X86PDPAE pPDDst = pPGM->CTXMID(ap,PaePDs[0]); / use it as a 2048 entry PD */
3228	unsigned cPDEs = ELEMENTS(pPDDst->a) * ELEMENTS(pPGM->apHCPaePDs);
3229	# endif
3230	if (cb != ~(RTGCUINTPTR)0)
3231	cPDEs = RT_MIN(cb >> SHW_PD_SHIFT, 1);
3232
3233	/** @todo call the other two PGMAssert() functions. /
3234
3235	# if PGM_GST_TYPE == PGM_TYPE_PAE
3236	/*
3237	* Check the 4 PDPTs too.
3238	*/
3239	for (unsigned i = 0; i < 4; i++)
3240	{
3241	RTHCPTR HCPtr;
3242	RTHCPHYS HCPhys;
3243	RTGCPHYS GCPhys = pVM->pgm.s.CTXSUFF(pGstPaePDPT)->a[i].u & X86_PDPE_PG_MASK;
3244	int rc2 = pgmRamGCPhys2HCPtrAndHCPhysWithFlags(&pVM->pgm.s, GCPhys, &HCPtr, &HCPhys);
3245	if (VBOX_SUCCESS(rc2))
3246	{
3247	AssertMsg( pVM->pgm.s.apGstPaePDsHC[i] == (R3R0PTRTYPE(PX86PDPAE))HCPtr
3248	&& pVM->pgm.s.aGCPhysGstPaePDs[i] == GCPhys,
3249	("idx %d apGstPaePDsHC %VHv vs %VHv aGCPhysGstPaePDs %VGp vs %VGp\n",
3250	i, pVM->pgm.s.apGstPaePDsHC[i], HCPtr, pVM->pgm.s.aGCPhysGstPaePDs[i], GCPhys));
3251	}
3252	}
3253	# endif
3254
3255	/*
3256	* Iterate the shadow page directory.
3257	*/
3258	GCPtr = (GCPtr >> SHW_PD_SHIFT) << SHW_PD_SHIFT;
3259	unsigned iPDDst = GCPtr >> SHW_PD_SHIFT;
3260	cPDEs += iPDDst;
3261	for (;
3262	iPDDst < cPDEs;
3263	iPDDst++, GCPtr += _4G / cPDEs)
3264	{
3265	# if PGM_GST_TYPE == PGM_TYPE_PAE
3266	uint32_t iPDSrc;
3267	PGSTPD pPDSrc = pgmGstGetPaePDPtr(pPGM, (RTGCUINTPTR)GCPtr, &iPDSrc);
3268	if (!pPDSrc)
3269	{
3270	AssertMsg(!pVM->pgm.s.CTXSUFF(pGstPaePDPT)->a[(GCPtr >> GST_PDPT_SHIFT) & GST_PDPT_MASK].n.u1Present, ("Guest PDTPR not present, shadow PDPTR %VX64\n", pVM->pgm.s.CTXSUFF(pGstPaePDPT)->a[(GCPtr >> GST_PDPT_SHIFT) & GST_PDPT_MASK].u));
3271	continue;
3272	}
3273	#endif
3274
3275	const SHWPDE PdeDst = pPDDst->a[iPDDst];
3276	if (PdeDst.u & PGM_PDFLAGS_MAPPING)
3277	{
3278	Assert(pgmMapAreMappingsEnabled(&pVM->pgm.s));
3279	if ((PdeDst.u & X86_PDE_AVL_MASK) != PGM_PDFLAGS_MAPPING)
3280	{
3281	AssertMsgFailed(("Mapping shall only have PGM_PDFLAGS_MAPPING set! PdeDst.u=%#RX64\n", (uint64_t)PdeDst.u));
3282	cErrors++;
3283	continue;
3284	}
3285	}
3286	else if ( (PdeDst.u & X86_PDE_P)
3287	\|\| ((PdeDst.u & (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY)) == (X86_PDE_P \| PGM_PDFLAGS_TRACK_DIRTY))
3288	)
3289	{
3290	HCPhysShw = PdeDst.u & SHW_PDE_PG_MASK;
3291	PPGMPOOLPAGE pPoolPage = pgmPoolGetPageByHCPhys(pVM, HCPhysShw);
3292	if (!pPoolPage)
3293	{
3294	AssertMsgFailed(("Invalid page table address %VGp at %VGv! PdeDst=%#RX64\n",
3295	HCPhysShw, GCPtr, (uint64_t)PdeDst.u));
3296	cErrors++;
3297	continue;
3298	}
3299	const SHWPT pPTDst = (const SHWPT )PGMPOOL_PAGE_2_PTR(pVM, pPoolPage);
3300
3301	if (PdeDst.u & (X86_PDE4M_PWT \| X86_PDE4M_PCD))
3302	{
3303	AssertMsgFailed(("PDE flags PWT and/or PCD is set at %VGv! These flags are not virtualized! PdeDst=%#RX64\n",
3304	GCPtr, (uint64_t)PdeDst.u));
3305	cErrors++;
3306	}
3307
3308	if (PdeDst.u & (X86_PDE4M_G \| X86_PDE4M_D))
3309	{
3310	AssertMsgFailed(("4K PDE reserved flags at %VGv! PdeDst=%#RX64\n",
3311	GCPtr, (uint64_t)PdeDst.u));
3312	cErrors++;
3313	}
3314
3315	const GSTPDE PdeSrc = pPDSrc->a[(iPDDst >> (GST_PD_SHIFT - SHW_PD_SHIFT)) & GST_PD_MASK];
3316	if (!PdeSrc.n.u1Present)
3317	{
3318	AssertMsgFailed(("Guest PDE at %VGv is not present! PdeDst=%#RX64 PdeSrc=%#RX64\n",
3319	GCPtr, (uint64_t)PdeDst.u, (uint64_t)PdeSrc.u));
3320	cErrors++;
3321	continue;
3322	}
3323
3324	if ( !PdeSrc.b.u1Size
3325	\|\| !(cr4 & X86_CR4_PSE))
3326	{
3327	GCPhysGst = PdeSrc.u & GST_PDE_PG_MASK;
3328	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3329	GCPhysGst \|= (iPDDst & 1) * (PAGE_SIZE / 2);
3330	# endif
3331	}
3332	else
3333	{
3334	# if PGM_GST_TYPE == PGM_TYPE_32BIT
3335	if (PdeSrc.u & X86_PDE4M_PG_HIGH_MASK)
3336	{
3337	AssertMsgFailed(("Guest PDE at %VGv is using PSE36 or similar! PdeSrc=%#RX64\n",
3338	GCPtr, (uint64_t)PdeSrc.u));
3339	cErrors++;
3340	continue;
3341	}
3342	# endif
3343	GCPhysGst = PdeSrc.u & GST_PDE_BIG_PG_MASK;
3344	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3345	GCPhysGst \|= GCPtr & RT_BIT(X86_PAGE_2M_SHIFT);
3346	# endif
3347	}
3348
3349	if ( pPoolPage->enmKind
3350	!= (!PdeSrc.b.u1Size \|\| !(cr4 & X86_CR4_PSE) ? BTH_PGMPOOLKIND_PT_FOR_PT : BTH_PGMPOOLKIND_PT_FOR_BIG))
3351	{
3352	AssertMsgFailed(("Invalid shadow page table kind %d at %VGv! PdeSrc=%#RX64\n",
3353	pPoolPage->enmKind, GCPtr, (uint64_t)PdeSrc.u));
3354	cErrors++;
3355	}
3356
3357	PPGMPAGE pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3358	if (!pPhysPage)
3359	{
3360	AssertMsgFailed(("Cannot find guest physical address %VGp in the PDE at %VGv! PdeSrc=%#RX64\n",
3361	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3362	cErrors++;
3363	continue;
3364	}
3365
3366	if (GCPhysGst != pPoolPage->GCPhys)
3367	{
3368	AssertMsgFailed(("GCPhysGst=%VGp != pPage->GCPhys=%VGp at %VGv\n",
3369	GCPhysGst, pPoolPage->GCPhys, GCPtr));
3370	cErrors++;
3371	continue;
3372	}
3373
3374	if ( !PdeSrc.b.u1Size
3375	\|\| !(cr4 & X86_CR4_PSE))
3376	{
3377	/*
3378	* Page Table.
3379	*/
3380	const GSTPT *pPTSrc;
3381	rc = PGM_GCPHYS_2_PTR(pVM, GCPhysGst & ~(RTGCPHYS)(PAGE_SIZE - 1), &pPTSrc);
3382	if (VBOX_FAILURE(rc))
3383	{
3384	AssertMsgFailed(("Cannot map/convert guest physical address %VGp in the PDE at %VGv! PdeSrc=%#RX64\n",
3385	GCPhysGst, GCPtr, (uint64_t)PdeSrc.u));
3386	cErrors++;
3387	continue;
3388	}
3389	if ( (PdeSrc.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/))
3390	!= (PdeDst.u & (X86_PDE_P \| X86_PDE_US \| X86_PDE_RW/* \| X86_PDE_A*/)))
3391	{
3392	/// @todo We get here a lot on out-of-sync CR3 entries. The access handler should zap them to avoid false alarms here!
3393	// (This problem will go away when/if we shadow multiple CR3s.)
3394	AssertMsgFailed(("4K PDE flags mismatch at %VGv! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3395	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3396	cErrors++;
3397	continue;
3398	}
3399	if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
3400	{
3401	AssertMsgFailed(("4K PDEs cannot have PGM_PDFLAGS_TRACK_DIRTY set! GCPtr=%VGv PdeDst=%#RX64\n",
3402	GCPtr, (uint64_t)PdeDst.u));
3403	cErrors++;
3404	continue;
3405	}
3406
3407	/* iterate the page table. */
3408	# if PGM_SHW_TYPE == PGM_TYPE_PAE && PGM_GST_TYPE == PGM_TYPE_32BIT
3409	/* Select the right PDE as we're emulating a 4kb page table with 2 shadow page tables. */
3410	const unsigned offPTSrc = ((GCPtr >> SHW_PD_SHIFT) & 1) * 512;
3411	# else
3412	const unsigned offPTSrc = 0;
3413	# endif
3414	for (unsigned iPT = 0, off = 0;
3415	iPT < ELEMENTS(pPTDst->a);
3416	iPT++, off += PAGE_SIZE)
3417	{
3418	const SHWPTE PteDst = pPTDst->a[iPT];
3419
3420	/* skip not-present entries. */
3421	if (!(PteDst.u & (X86_PTE_P \| PGM_PTFLAGS_TRACK_DIRTY))) /** @todo deal with ALL handlers and CSAM !P pages! */
3422	continue;
3423	Assert(PteDst.n.u1Present);
3424
3425	const GSTPTE PteSrc = pPTSrc->a[iPT + offPTSrc];
3426	if (!PteSrc.n.u1Present)
3427	{
3428	#ifdef IN_RING3
3429	PGMAssertHandlerAndFlagsInSync(pVM);
3430	PGMR3DumpHierarchyGC(pVM, cr3, cr4, (PdeSrc.u & GST_PDE_PG_MASK));
3431	#endif
3432	AssertMsgFailed(("Out of sync (!P) PTE at %VGv! PteSrc=%#RX64 PteDst=%#RX64 pPTSrc=%VGv iPTSrc=%x PdeSrc=%x physpte=%VGp\n",
3433	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u, pPTSrc, iPT + offPTSrc, PdeSrc.au32[0],
3434	(PdeSrc.u & GST_PDE_PG_MASK) + (iPT + offPTSrc)*sizeof(PteSrc)));
3435	cErrors++;
3436	continue;
3437	}
3438
3439	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;
3440	# if 1 /** @todo sync accessed bit properly... */
3441	fIgnoreFlags \|= X86_PTE_A;
3442	# endif
3443
3444	/* match the physical addresses */
3445	HCPhysShw = PteDst.u & SHW_PTE_PG_MASK;
3446	GCPhysGst = PteSrc.u & GST_PTE_PG_MASK;
3447
3448	# ifdef IN_RING3
3449	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
3450	if (VBOX_FAILURE(rc))
3451	{
3452	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3453	{
3454	AssertMsgFailed(("Cannot find guest physical address %VGp at %VGv! PteSrc=%#RX64 PteDst=%#RX64\n",
3455	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3456	cErrors++;
3457	continue;
3458	}
3459	}
3460	else if (HCPhysShw != (HCPhys & SHW_PTE_PG_MASK))
3461	{
3462	AssertMsgFailed(("Out of sync (phys) at %VGv! HCPhysShw=%VHp HCPhys=%VHp GCPhysGst=%VGp PteSrc=%#RX64 PteDst=%#RX64\n",
3463	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3464	cErrors++;
3465	continue;
3466	}
3467	# endif
3468
3469	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3470	if (!pPhysPage)
3471	{
3472	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
3473	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3474	{
3475	AssertMsgFailed(("Cannot find guest physical address %VGp at %VGv! PteSrc=%#RX64 PteDst=%#RX64\n",
3476	GCPhysGst, GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3477	cErrors++;
3478	continue;
3479	}
3480	# endif
3481	if (PteDst.n.u1Write)
3482	{
3483	AssertMsgFailed(("Invalid guest page at %VGv is writable! GCPhysGst=%VGp PteSrc=%#RX64 PteDst=%#RX64\n",
3484	GCPtr + off, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3485	cErrors++;
3486	}
3487	fIgnoreFlags \|= X86_PTE_RW;
3488	}
3489	else if (HCPhysShw != (PGM_PAGE_GET_HCPHYS(pPhysPage) & SHW_PTE_PG_MASK))
3490	{
3491	AssertMsgFailed(("Out of sync (phys) at %VGv! HCPhysShw=%VHp HCPhys=%VHp GCPhysGst=%VGp PteSrc=%#RX64 PteDst=%#RX64\n",
3492	GCPtr + off, HCPhysShw, pPhysPage->HCPhys, GCPhysGst, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3493	cErrors++;
3494	continue;
3495	}
3496
3497	/* flags */
3498	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
3499	{
3500	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
3501	{
3502	if (PteDst.n.u1Write)
3503	{
3504	AssertMsgFailed(("WRITE access flagged at %VGv but the page is writable! HCPhys=%VGv PteSrc=%#RX64 PteDst=%#RX64\n",
3505	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3506	cErrors++;
3507	continue;
3508	}
3509	fIgnoreFlags \|= X86_PTE_RW;
3510	}
3511	else
3512	{
3513	if (PteDst.n.u1Present)
3514	{
3515	AssertMsgFailed(("ALL access flagged at %VGv but the page is present! HCPhys=%VHp PteSrc=%#RX64 PteDst=%#RX64\n",
3516	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3517	cErrors++;
3518	continue;
3519	}
3520	fIgnoreFlags \|= X86_PTE_P;
3521	}
3522	}
3523	else
3524	{
3525	if (!PteSrc.n.u1Dirty && PteSrc.n.u1Write)
3526	{
3527	if (PteDst.n.u1Write)
3528	{
3529	AssertMsgFailed(("!DIRTY page at %VGv is writable! PteSrc=%#RX64 PteDst=%#RX64\n",
3530	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3531	cErrors++;
3532	continue;
3533	}
3534	if (!(PteDst.u & PGM_PTFLAGS_TRACK_DIRTY))
3535	{
3536	AssertMsgFailed(("!DIRTY page at %VGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3537	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3538	cErrors++;
3539	continue;
3540	}
3541	if (PteDst.n.u1Dirty)
3542	{
3543	AssertMsgFailed(("!DIRTY page at %VGv is marked DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3544	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3545	cErrors++;
3546	}
3547	# if 0 /** @todo sync access bit properly... */
3548	if (PteDst.n.u1Accessed != PteSrc.n.u1Accessed)
3549	{
3550	AssertMsgFailed(("!DIRTY page at %VGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
3551	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3552	cErrors++;
3553	}
3554	fIgnoreFlags \|= X86_PTE_RW;
3555	# else
3556	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
3557	# endif
3558	}
3559	else if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
3560	{
3561	/* access bit emulation (not implemented). */
3562	if (PteSrc.n.u1Accessed \|\| PteDst.n.u1Present)
3563	{
3564	AssertMsgFailed(("PGM_PTFLAGS_TRACK_DIRTY set at %VGv but no accessed bit emulation! PteSrc=%#RX64 PteDst=%#RX64\n",
3565	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3566	cErrors++;
3567	continue;
3568	}
3569	if (!PteDst.n.u1Accessed)
3570	{
3571	AssertMsgFailed(("!ACCESSED page at %VGv is has the accessed bit set! PteSrc=%#RX64 PteDst=%#RX64\n",
3572	GCPtr + off, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3573	cErrors++;
3574	}
3575	fIgnoreFlags \|= X86_PTE_P;
3576	}
3577	# ifdef DEBUG_sandervl
3578	fIgnoreFlags \|= X86_PTE_D \| X86_PTE_A;
3579	# endif
3580	}
3581
3582	if ( (PteSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
3583	&& (PteSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags)
3584	)
3585	{
3586	AssertMsgFailed(("Flags mismatch at %VGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PteSrc=%#RX64 PteDst=%#RX64\n",
3587	GCPtr + off, (uint64_t)PteSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
3588	fIgnoreFlags, (uint64_t)PteSrc.u, (uint64_t)PteDst.u));
3589	cErrors++;
3590	continue;
3591	}
3592	} /* foreach PTE */
3593	}
3594	else
3595	{
3596	/*
3597	* Big Page.
3598	*/
3599	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;
3600	if (!PdeSrc.b.u1Dirty && PdeSrc.b.u1Write)
3601	{
3602	if (PdeDst.n.u1Write)
3603	{
3604	AssertMsgFailed(("!DIRTY page at %VGv is writable! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3605	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3606	cErrors++;
3607	continue;
3608	}
3609	if (!(PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY))
3610	{
3611	AssertMsgFailed(("!DIRTY page at %VGv is not marked TRACK_DIRTY! PteSrc=%#RX64 PteDst=%#RX64\n",
3612	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3613	cErrors++;
3614	continue;
3615	}
3616	# if 0 /** @todo sync access bit properly... */
3617	if (PdeDst.n.u1Accessed != PdeSrc.b.u1Accessed)
3618	{
3619	AssertMsgFailed(("!DIRTY page at %VGv is has mismatching accessed bit! PteSrc=%#RX64 PteDst=%#RX64\n",
3620	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3621	cErrors++;
3622	}
3623	fIgnoreFlags \|= X86_PTE_RW;
3624	# else
3625	fIgnoreFlags \|= X86_PTE_RW \| X86_PTE_A;
3626	# endif
3627	}
3628	else if (PdeDst.u & PGM_PDFLAGS_TRACK_DIRTY)
3629	{
3630	/* access bit emulation (not implemented). */
3631	if (PdeSrc.b.u1Accessed \|\| PdeDst.n.u1Present)
3632	{
3633	AssertMsgFailed(("PGM_PDFLAGS_TRACK_DIRTY set at %VGv but no accessed bit emulation! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3634	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3635	cErrors++;
3636	continue;
3637	}
3638	if (!PdeDst.n.u1Accessed)
3639	{
3640	AssertMsgFailed(("!ACCESSED page at %VGv is has the accessed bit set! PdeSrc=%#RX64 PdeDst=%#RX64\n",
3641	GCPtr, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3642	cErrors++;
3643	}
3644	fIgnoreFlags \|= X86_PTE_P;
3645	}
3646
3647	if ((PdeSrc.u & ~fIgnoreFlags) != (PdeDst.u & ~fIgnoreFlags))
3648	{
3649	AssertMsgFailed(("Flags mismatch (B) at %VGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PdeDst=%#RX64\n",
3650	GCPtr, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PdeDst.u & ~fIgnoreFlags,
3651	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PdeDst.u));
3652	cErrors++;
3653	}
3654
3655	/* iterate the page table. */
3656	for (unsigned iPT = 0, off = 0;
3657	iPT < ELEMENTS(pPTDst->a);
3658	iPT++, off += PAGE_SIZE, GCPhysGst += PAGE_SIZE)
3659	{
3660	const SHWPTE PteDst = pPTDst->a[iPT];
3661
3662	if (PteDst.u & PGM_PTFLAGS_TRACK_DIRTY)
3663	{
3664	AssertMsgFailed(("The PTE at %VGv emulating a 2/4M page is marked TRACK_DIRTY! PdeSrc=%#RX64 PteDst=%#RX64\n",
3665	GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3666	cErrors++;
3667	}
3668
3669	/* skip not-present entries. */
3670	if (!PteDst.n.u1Present) /** @todo deal with ALL handlers and CSAM !P pages! */
3671	continue;
3672
3673	fIgnoreFlags = X86_PTE_PAE_PG_MASK \| X86_PTE_AVL_MASK \| X86_PTE_PWT \| X86_PTE_PCD \| X86_PTE_PAT;
3674
3675	/* match the physical addresses */
3676	HCPhysShw = PteDst.u & X86_PTE_PAE_PG_MASK;
3677
3678	# ifdef IN_RING3
3679	rc = PGMPhysGCPhys2HCPhys(pVM, GCPhysGst, &HCPhys);
3680	if (VBOX_FAILURE(rc))
3681	{
3682	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3683	{
3684	AssertMsgFailed(("Cannot find guest physical address %VGp at %VGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
3685	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3686	cErrors++;
3687	}
3688	}
3689	else if (HCPhysShw != (HCPhys & X86_PTE_PAE_PG_MASK))
3690	{
3691	AssertMsgFailed(("Out of sync (phys) at %VGv! HCPhysShw=%VHp HCPhys=%VHp GCPhysGst=%VGp PdeSrc=%#RX64 PteDst=%#RX64\n",
3692	GCPtr + off, HCPhysShw, HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3693	cErrors++;
3694	continue;
3695	}
3696	# endif
3697
3698	pPhysPage = pgmPhysGetPage(pPGM, GCPhysGst);
3699	if (!pPhysPage)
3700	{
3701	# ifdef IN_RING3 /** @todo make MMR3PageDummyHCPhys an 'All' function! */
3702	if (HCPhysShw != MMR3PageDummyHCPhys(pVM))
3703	{
3704	AssertMsgFailed(("Cannot find guest physical address %VGp at %VGv! PdeSrc=%#RX64 PteDst=%#RX64\n",
3705	GCPhysGst, GCPtr + off, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3706	cErrors++;
3707	continue;
3708	}
3709	# endif
3710	if (PteDst.n.u1Write)
3711	{
3712	AssertMsgFailed(("Invalid guest page at %VGv is writable! GCPhysGst=%VGp PdeSrc=%#RX64 PteDst=%#RX64\n",
3713	GCPtr + off, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3714	cErrors++;
3715	}
3716	fIgnoreFlags \|= X86_PTE_RW;
3717	}
3718	else if (HCPhysShw != (pPhysPage->HCPhys & X86_PTE_PAE_PG_MASK))
3719	{
3720	AssertMsgFailed(("Out of sync (phys) at %VGv! HCPhysShw=%VHp HCPhys=%VHp GCPhysGst=%VGp PdeSrc=%#RX64 PteDst=%#RX64\n",
3721	GCPtr + off, HCPhysShw, pPhysPage->HCPhys, GCPhysGst, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3722	cErrors++;
3723	continue;
3724	}
3725
3726	/* flags */
3727	if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPhysPage))
3728	{
3729	if (!PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPhysPage))
3730	{
3731	if (PGM_PAGE_GET_HNDL_PHYS_STATE(pPhysPage) != PGM_PAGE_HNDL_PHYS_STATE_DISABLED)
3732	{
3733	if (PteDst.n.u1Write)
3734	{
3735	AssertMsgFailed(("WRITE access flagged at %VGv but the page is writable! HCPhys=%VGv PdeSrc=%#RX64 PteDst=%#RX64\n",
3736	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3737	cErrors++;
3738	continue;
3739	}
3740	fIgnoreFlags \|= X86_PTE_RW;
3741	}
3742	}
3743	else
3744	{
3745	if (PteDst.n.u1Present)
3746	{
3747	AssertMsgFailed(("ALL access flagged at %VGv but the page is present! HCPhys=%VGv PdeSrc=%#RX64 PteDst=%#RX64\n",
3748	GCPtr + off, pPhysPage->HCPhys, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3749	cErrors++;
3750	continue;
3751	}
3752	fIgnoreFlags \|= X86_PTE_P;
3753	}
3754	}
3755
3756	if ( (PdeSrc.u & ~fIgnoreFlags) != (PteDst.u & ~fIgnoreFlags)
3757	&& (PdeSrc.u & ~(fIgnoreFlags \| X86_PTE_RW)) != (PteDst.u & ~fIgnoreFlags) /* lazy phys handler dereg. */
3758	)
3759	{
3760	AssertMsgFailed(("Flags mismatch (BT) at %VGv! %#RX64 != %#RX64 fIgnoreFlags=%#RX64 PdeSrc=%#RX64 PteDst=%#RX64\n",
3761	GCPtr + off, (uint64_t)PdeSrc.u & ~fIgnoreFlags, (uint64_t)PteDst.u & ~fIgnoreFlags,
3762	fIgnoreFlags, (uint64_t)PdeSrc.u, (uint64_t)PteDst.u));
3763	cErrors++;
3764	continue;
3765	}
3766	} /* foreach PTE */
3767	}
3768	}
3769	/* not present */
3770
3771	} /* forearch PDE */
3772
3773	# ifdef DEBUG
3774	if (cErrors)
3775	LogFlow(("AssertCR3: cErrors=%d\n", cErrors));
3776	# endif
3777
3778	#elif PGM_GST_TYPE == PGM_TYPE_PAE
3779	//# error not implemented
3780
3781
3782	#elif PGM_GST_TYPE == PGM_TYPE_AMD64
3783	//# error not implemented
3784
3785	/#else: guest real and protected mode /
3786	#endif
3787	return cErrors;
3788
3789	#endif /* PGM_SHW_TYPE != PGM_TYPE_NESTED */
3790	}
3791	#endif /* VBOX_STRICT */
3792

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

source: vbox/trunk/src/VBox/VMM/VMMAll/PGMAllBth.h@ 9344

Download in other formats: