PGMAllPool.cpp@ 41801

Last change on this file since 41801 was 41783, checked in by vboxsync, 12 years ago
Doxygen, comment typos.
Property svn:eol-style set to `native` Property svn:keywords set to `Id`
File size: 210.0 KB

Line
1	/* $Id: PGMAllPool.cpp 41783 2012-06-16 19:24:15Z vboxsync $ */
2	/** @file
3	* PGM Shadow Page Pool.
4	*/
5
6	/*
7	* Copyright (C) 2006-2010 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*******************************************************************************
20	* Header Files *
21	*******************************************************************************/
22	#define LOG_GROUP LOG_GROUP_PGM_POOL
23	#include <VBox/vmm/pgm.h>
24	#include <VBox/vmm/mm.h>
25	#include <VBox/vmm/em.h>
26	#include <VBox/vmm/cpum.h>
27	#ifdef IN_RC
28	# include <VBox/vmm/patm.h>
29	#endif
30	#include "PGMInternal.h"
31	#include <VBox/vmm/vm.h>
32	#include "PGMInline.h"
33	#include <VBox/disopcode.h>
34	#include <VBox/vmm/hwacc_vmx.h>
35
36	#include <VBox/log.h>
37	#include <VBox/err.h>
38	#include <iprt/asm.h>
39	#include <iprt/asm-amd64-x86.h>
40	#include <iprt/string.h>
41
42
43	/*******************************************************************************
44	* Internal Functions *
45	*******************************************************************************/
46	RT_C_DECLS_BEGIN
47	DECLINLINE(unsigned) pgmPoolTrackGetShadowEntrySize(PGMPOOLKIND enmKind);
48	DECLINLINE(unsigned) pgmPoolTrackGetGuestEntrySize(PGMPOOLKIND enmKind);
49	static void pgmPoolTrackDeref(PPGMPOOL pPool, PPGMPOOLPAGE pPage);
50	static int pgmPoolTrackAddUser(PPGMPOOL pPool, PPGMPOOLPAGE pPage, uint16_t iUser, uint32_t iUserTable);
51	static void pgmPoolMonitorModifiedRemove(PPGMPOOL pPool, PPGMPOOLPAGE pPage);
52	#ifndef IN_RING3
53	DECLEXPORT(int) pgmPoolAccessHandler(PVM pVM, RTGCUINT uErrorCode, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault, RTGCPHYS GCPhysFault, void *pvUser);
54	#endif
55	#ifdef LOG_ENABLED
56	static const char *pgmPoolPoolKindToStr(uint8_t enmKind);
57	#endif
58	#if 0 /defined(VBOX_STRICT) && defined(PGMPOOL_WITH_OPTIMIZED_DIRTY_PT)/
59	static void pgmPoolTrackCheckPTPaePae(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMSHWPTPAE pShwPT, PCX86PTPAE pGstPT);
60	#endif
61
62	int pgmPoolTrackFlushGCPhysPTsSlow(PVM pVM, PPGMPAGE pPhysPage);
63	PPGMPOOLPHYSEXT pgmPoolTrackPhysExtAlloc(PVM pVM, uint16_t *piPhysExt);
64	void pgmPoolTrackPhysExtFree(PVM pVM, uint16_t iPhysExt);
65	void pgmPoolTrackPhysExtFreeList(PVM pVM, uint16_t iPhysExt);
66
67	RT_C_DECLS_END
68
69
70	/**
71	* Checks if the specified page pool kind is for a 4MB or 2MB guest page.
72	*
73	* @returns true if it's the shadow of a 4MB or 2MB guest page, otherwise false.
74	* @param enmKind The page kind.
75	*/
76	DECLINLINE(bool) pgmPoolIsBigPage(PGMPOOLKIND enmKind)
77	{
78	switch (enmKind)
79	{
80	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
81	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
82	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
83	return true;
84	default:
85	return false;
86	}
87	}
88
89
90	/**
91	* Flushes a chain of pages sharing the same access monitor.
92	*
93	* @returns VBox status code suitable for scheduling.
94	* @param pPool The pool.
95	* @param pPage A page in the chain.
96	* @todo VBOXSTRICTRC
97	*/
98	int pgmPoolMonitorChainFlush(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
99	{
100	LogFlow(("pgmPoolMonitorChainFlush: Flush page %RGp type=%d\n", pPage->GCPhys, pPage->enmKind));
101
102	/*
103	* Find the list head.
104	*/
105	uint16_t idx = pPage->idx;
106	if (pPage->iMonitoredPrev != NIL_PGMPOOL_IDX)
107	{
108	while (pPage->iMonitoredPrev != NIL_PGMPOOL_IDX)
109	{
110	idx = pPage->iMonitoredPrev;
111	Assert(idx != pPage->idx);
112	pPage = &pPool->aPages[idx];
113	}
114	}
115
116	/*
117	* Iterate the list flushing each shadow page.
118	*/
119	int rc = VINF_SUCCESS;
120	for (;;)
121	{
122	idx = pPage->iMonitoredNext;
123	Assert(idx != pPage->idx);
124	if (pPage->idx >= PGMPOOL_IDX_FIRST)
125	{
126	int rc2 = pgmPoolFlushPage(pPool, pPage);
127	AssertRC(rc2);
128	}
129	/* next */
130	if (idx == NIL_PGMPOOL_IDX)
131	break;
132	pPage = &pPool->aPages[idx];
133	}
134	return rc;
135	}
136
137
138	/**
139	* Wrapper for getting the current context pointer to the entry being modified.
140	*
141	* @returns VBox status code suitable for scheduling.
142	* @param pVM Pointer to the VM.
143	* @param pvDst Destination address
144	* @param pvSrc Source guest virtual address.
145	* @param GCPhysSrc The source guest physical address.
146	* @param cb Size of data to read
147	*/
148	DECLINLINE(int) pgmPoolPhysSimpleReadGCPhys(PVM pVM, void *pvDst, CTXTYPE(RTGCPTR, RTHCPTR, RTGCPTR) pvSrc,
149	RTGCPHYS GCPhysSrc, size_t cb)
150	{
151	#if defined(IN_RING3)
152	NOREF(pVM); NOREF(GCPhysSrc);
153	memcpy(pvDst, (RTHCPTR)((uintptr_t)pvSrc & ~(RTHCUINTPTR)(cb - 1)), cb);
154	return VINF_SUCCESS;
155	#else
156	/* @todo in RC we could attempt to use the virtual address, although this can cause many faults (PAE Windows XP guest). */
157	NOREF(pvSrc);
158	return PGMPhysSimpleReadGCPhys(pVM, pvDst, GCPhysSrc & ~(RTGCPHYS)(cb - 1), cb);
159	#endif
160	}
161
162
163	/**
164	* Process shadow entries before they are changed by the guest.
165	*
166	* For PT entries we will clear them. For PD entries, we'll simply check
167	* for mapping conflicts and set the SyncCR3 FF if found.
168	*
169	* @param pVCpu VMCPU handle
170	* @param pPool The pool.
171	* @param pPage The head page.
172	* @param GCPhysFault The guest physical fault address.
173	* @param uAddress In R0 and GC this is the guest context fault address (flat).
174	* In R3 this is the host context 'fault' address.
175	* @param cbWrite Write size; might be zero if the caller knows we're not crossing entry boundaries
176	*/
177	void pgmPoolMonitorChainChanging(PVMCPU pVCpu, PPGMPOOL pPool, PPGMPOOLPAGE pPage, RTGCPHYS GCPhysFault,
178	CTXTYPE(RTGCPTR, RTHCPTR, RTGCPTR) pvAddress, unsigned cbWrite)
179	{
180	AssertMsg(pPage->iMonitoredPrev == NIL_PGMPOOL_IDX, ("%u (idx=%u)\n", pPage->iMonitoredPrev, pPage->idx));
181	const unsigned off = GCPhysFault & PAGE_OFFSET_MASK;
182	PVM pVM = pPool->CTX_SUFF(pVM);
183	NOREF(pVCpu);
184
185	LogFlow(("pgmPoolMonitorChainChanging: %RGv phys=%RGp cbWrite=%d\n", (RTGCPTR)(CTXTYPE(RTGCPTR, uintptr_t, RTGCPTR))pvAddress, GCPhysFault, cbWrite));
186
187	for (;;)
188	{
189	union
190	{
191	void *pv;
192	PX86PT pPT;
193	PPGMSHWPTPAE pPTPae;
194	PX86PD pPD;
195	PX86PDPAE pPDPae;
196	PX86PDPT pPDPT;
197	PX86PML4 pPML4;
198	} uShw;
199
200	LogFlow(("pgmPoolMonitorChainChanging: page idx=%d phys=%RGp (next=%d) kind=%s\n", pPage->idx, pPage->GCPhys, pPage->iMonitoredNext, pgmPoolPoolKindToStr(pPage->enmKind), cbWrite));
201
202	uShw.pv = NULL;
203	switch (pPage->enmKind)
204	{
205	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
206	{
207	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPT));
208	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
209	const unsigned iShw = off / sizeof(X86PTE);
210	LogFlow(("PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT iShw=%x\n", iShw));
211	if (uShw.pPT->a[iShw].n.u1Present)
212	{
213	X86PTE GstPte;
214
215	int rc = pgmPoolPhysSimpleReadGCPhys(pVM, &GstPte, pvAddress, GCPhysFault, sizeof(GstPte));
216	AssertRC(rc);
217	Log4(("pgmPoolMonitorChainChanging 32_32: deref %016RX64 GCPhys %08RX32\n", uShw.pPT->a[iShw].u & X86_PTE_PAE_PG_MASK, GstPte.u & X86_PTE_PG_MASK));
218	pgmPoolTracDerefGCPhysHint(pPool, pPage,
219	uShw.pPT->a[iShw].u & X86_PTE_PAE_PG_MASK,
220	GstPte.u & X86_PTE_PG_MASK,
221	iShw);
222	ASMAtomicWriteU32(&uShw.pPT->a[iShw].u, 0);
223	}
224	break;
225	}
226
227	/* page/2 sized */
228	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
229	{
230	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPT));
231	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
232	if (!((off ^ pPage->GCPhys) & (PAGE_SIZE / 2)))
233	{
234	const unsigned iShw = (off / sizeof(X86PTE)) & (X86_PG_PAE_ENTRIES - 1);
235	LogFlow(("PGMPOOLKIND_PAE_PT_FOR_32BIT_PT iShw=%x\n", iShw));
236	if (PGMSHWPTEPAE_IS_P(uShw.pPTPae->a[iShw]))
237	{
238	X86PTE GstPte;
239	int rc = pgmPoolPhysSimpleReadGCPhys(pVM, &GstPte, pvAddress, GCPhysFault, sizeof(GstPte));
240	AssertRC(rc);
241
242	Log4(("pgmPoolMonitorChainChanging pae_32: deref %016RX64 GCPhys %08RX32\n", uShw.pPT->a[iShw].u & X86_PTE_PAE_PG_MASK, GstPte.u & X86_PTE_PG_MASK));
243	pgmPoolTracDerefGCPhysHint(pPool, pPage,
244	PGMSHWPTEPAE_GET_HCPHYS(uShw.pPTPae->a[iShw]),
245	GstPte.u & X86_PTE_PG_MASK,
246	iShw);
247	PGMSHWPTEPAE_ATOMIC_SET(uShw.pPTPae->a[iShw], 0);
248	}
249	}
250	break;
251	}
252
253	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
254	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
255	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
256	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
257	{
258	unsigned iGst = off / sizeof(X86PDE);
259	unsigned iShwPdpt = iGst / 256;
260	unsigned iShw = (iGst % 256) * 2;
261	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
262
263	LogFlow(("pgmPoolMonitorChainChanging PAE for 32 bits: iGst=%x iShw=%x idx = %d page idx=%d\n", iGst, iShw, iShwPdpt, pPage->enmKind - PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD));
264	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPD));
265	if (iShwPdpt == pPage->enmKind - (unsigned)PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD)
266	{
267	for (unsigned i = 0; i < 2; i++)
268	{
269	# ifndef IN_RING0
270	if ((uShw.pPDPae->a[iShw + i].u & (PGM_PDFLAGS_MAPPING \| X86_PDE_P)) == (PGM_PDFLAGS_MAPPING \| X86_PDE_P))
271	{
272	Assert(pgmMapAreMappingsEnabled(pVM));
273	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
274	LogFlow(("pgmPoolMonitorChainChanging: Detected conflict at iShwPdpt=%#x iShw=%#x!\n", iShwPdpt, iShw+i));
275	break;
276	}
277	# endif /* !IN_RING0 */
278	if (uShw.pPDPae->a[iShw+i].n.u1Present)
279	{
280	LogFlow(("pgmPoolMonitorChainChanging: pae pd iShw=%#x: %RX64 -> freeing it!\n", iShw+i, uShw.pPDPae->a[iShw+i].u));
281	pgmPoolFree(pVM,
282	uShw.pPDPae->a[iShw+i].u & X86_PDE_PAE_PG_MASK,
283	pPage->idx,
284	iShw + i);
285	ASMAtomicWriteU64(&uShw.pPDPae->a[iShw+i].u, 0);
286	}
287
288	/* paranoia / a bit assumptive. */
289	if ( (off & 3)
290	&& (off & 3) + cbWrite > 4)
291	{
292	const unsigned iShw2 = iShw + 2 + i;
293	if (iShw2 < RT_ELEMENTS(uShw.pPDPae->a))
294	{
295	# ifndef IN_RING0
296	if ((uShw.pPDPae->a[iShw2].u & (PGM_PDFLAGS_MAPPING \| X86_PDE_P)) == (PGM_PDFLAGS_MAPPING \| X86_PDE_P))
297	{
298	Assert(pgmMapAreMappingsEnabled(pVM));
299	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
300	LogFlow(("pgmPoolMonitorChainChanging: Detected conflict at iShwPdpt=%#x iShw2=%#x!\n", iShwPdpt, iShw2));
301	break;
302	}
303	# endif /* !IN_RING0 */
304	if (uShw.pPDPae->a[iShw2].n.u1Present)
305	{
306	LogFlow(("pgmPoolMonitorChainChanging: pae pd iShw=%#x: %RX64 -> freeing it!\n", iShw2, uShw.pPDPae->a[iShw2].u));
307	pgmPoolFree(pVM,
308	uShw.pPDPae->a[iShw2].u & X86_PDE_PAE_PG_MASK,
309	pPage->idx,
310	iShw2);
311	ASMAtomicWriteU64(&uShw.pPDPae->a[iShw2].u, 0);
312	}
313	}
314	}
315	}
316	}
317	break;
318	}
319
320	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
321	{
322	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
323	const unsigned iShw = off / sizeof(X86PTEPAE);
324	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPT));
325	if (PGMSHWPTEPAE_IS_P(uShw.pPTPae->a[iShw]))
326	{
327	X86PTEPAE GstPte;
328	int rc = pgmPoolPhysSimpleReadGCPhys(pVM, &GstPte, pvAddress, GCPhysFault, sizeof(GstPte));
329	AssertRC(rc);
330
331	Log4(("pgmPoolMonitorChainChanging pae: deref %016RX64 GCPhys %016RX64\n", PGMSHWPTEPAE_GET_HCPHYS(uShw.pPTPae->a[iShw]), GstPte.u & X86_PTE_PAE_PG_MASK));
332	pgmPoolTracDerefGCPhysHint(pPool, pPage,
333	PGMSHWPTEPAE_GET_HCPHYS(uShw.pPTPae->a[iShw]),
334	GstPte.u & X86_PTE_PAE_PG_MASK,
335	iShw);
336	PGMSHWPTEPAE_ATOMIC_SET(uShw.pPTPae->a[iShw], 0);
337	}
338
339	/* paranoia / a bit assumptive. */
340	if ( (off & 7)
341	&& (off & 7) + cbWrite > sizeof(X86PTEPAE))
342	{
343	const unsigned iShw2 = (off + cbWrite - 1) / sizeof(X86PTEPAE);
344	AssertBreak(iShw2 < RT_ELEMENTS(uShw.pPTPae->a));
345
346	if (PGMSHWPTEPAE_IS_P(uShw.pPTPae->a[iShw2]))
347	{
348	X86PTEPAE GstPte;
349	# ifdef IN_RING3
350	int rc = pgmPoolPhysSimpleReadGCPhys(pVM, &GstPte, (RTHCPTR)((RTHCUINTPTR)pvAddress + sizeof(GstPte)), GCPhysFault + sizeof(GstPte), sizeof(GstPte));
351	# else
352	int rc = pgmPoolPhysSimpleReadGCPhys(pVM, &GstPte, pvAddress + sizeof(GstPte), GCPhysFault + sizeof(GstPte), sizeof(GstPte));
353	# endif
354	AssertRC(rc);
355	Log4(("pgmPoolMonitorChainChanging pae: deref %016RX64 GCPhys %016RX64\n", PGMSHWPTEPAE_GET_HCPHYS(uShw.pPTPae->a[iShw2]), GstPte.u & X86_PTE_PAE_PG_MASK));
356	pgmPoolTracDerefGCPhysHint(pPool, pPage,
357	PGMSHWPTEPAE_GET_HCPHYS(uShw.pPTPae->a[iShw2]),
358	GstPte.u & X86_PTE_PAE_PG_MASK,
359	iShw2);
360	PGMSHWPTEPAE_ATOMIC_SET(uShw.pPTPae->a[iShw2], 0);
361	}
362	}
363	break;
364	}
365
366	case PGMPOOLKIND_32BIT_PD:
367	{
368	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
369	const unsigned iShw = off / sizeof(X86PTE); // ASSUMING 32-bit guest paging!
370
371	LogFlow(("pgmPoolMonitorChainChanging: PGMPOOLKIND_32BIT_PD %x\n", iShw));
372	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPD));
373	# ifndef IN_RING0
374	if (uShw.pPD->a[iShw].u & PGM_PDFLAGS_MAPPING)
375	{
376	Assert(pgmMapAreMappingsEnabled(pVM));
377	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
378	STAM_COUNTER_INC(&(pVCpu->pgm.s.CTX_SUFF(pStats)->StatRZGuestCR3WriteConflict));
379	LogFlow(("pgmPoolMonitorChainChanging: Detected conflict at iShw=%#x!\n", iShw));
380	break;
381	}
382	# endif /* !IN_RING0 */
383	# ifndef IN_RING0
384	else
385	# endif /* !IN_RING0 */
386	{
387	if (uShw.pPD->a[iShw].n.u1Present)
388	{
389	LogFlow(("pgmPoolMonitorChainChanging: 32 bit pd iShw=%#x: %RX64 -> freeing it!\n", iShw, uShw.pPD->a[iShw].u));
390	pgmPoolFree(pVM,
391	uShw.pPD->a[iShw].u & X86_PDE_PAE_PG_MASK,
392	pPage->idx,
393	iShw);
394	ASMAtomicWriteU32(&uShw.pPD->a[iShw].u, 0);
395	}
396	}
397	/* paranoia / a bit assumptive. */
398	if ( (off & 3)
399	&& (off & 3) + cbWrite > sizeof(X86PTE))
400	{
401	const unsigned iShw2 = (off + cbWrite - 1) / sizeof(X86PTE);
402	if ( iShw2 != iShw
403	&& iShw2 < RT_ELEMENTS(uShw.pPD->a))
404	{
405	# ifndef IN_RING0
406	if (uShw.pPD->a[iShw2].u & PGM_PDFLAGS_MAPPING)
407	{
408	Assert(pgmMapAreMappingsEnabled(pVM));
409	STAM_COUNTER_INC(&(pVCpu->pgm.s.CTX_SUFF(pStats)->StatRZGuestCR3WriteConflict));
410	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
411	LogFlow(("pgmPoolMonitorChainChanging: Detected conflict at iShw2=%#x!\n", iShw2));
412	break;
413	}
414	# endif /* !IN_RING0 */
415	if (uShw.pPD->a[iShw2].n.u1Present)
416	{
417	LogFlow(("pgmPoolMonitorChainChanging: 32 bit pd iShw=%#x: %RX64 -> freeing it!\n", iShw2, uShw.pPD->a[iShw2].u));
418	pgmPoolFree(pVM,
419	uShw.pPD->a[iShw2].u & X86_PDE_PAE_PG_MASK,
420	pPage->idx,
421	iShw2);
422	ASMAtomicWriteU32(&uShw.pPD->a[iShw2].u, 0);
423	}
424	}
425	}
426	#if 0 /* useful when running PGMAssertCR3(), a bit too troublesome for general use (TLBs). */
427	if ( uShw.pPD->a[iShw].n.u1Present
428	&& !VMCPU_FF_ISSET(pVCpu, VMCPU_FF_PGM_SYNC_CR3))
429	{
430	LogFlow(("pgmPoolMonitorChainChanging: iShw=%#x: %RX32 -> freeing it!\n", iShw, uShw.pPD->a[iShw].u));
431	# ifdef IN_RC /* TLB load - we're pushing things a bit... */
432	ASMProbeReadByte(pvAddress);
433	# endif
434	pgmPoolFree(pVM, uShw.pPD->a[iShw].u & X86_PDE_PG_MASK, pPage->idx, iShw);
435	ASMAtomicWriteU32(&uShw.pPD->a[iShw].u, 0);
436	}
437	#endif
438	break;
439	}
440
441	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
442	{
443	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
444	const unsigned iShw = off / sizeof(X86PDEPAE);
445	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPD));
446	#ifndef IN_RING0
447	if (uShw.pPDPae->a[iShw].u & PGM_PDFLAGS_MAPPING)
448	{
449	Assert(pgmMapAreMappingsEnabled(pVM));
450	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
451	STAM_COUNTER_INC(&(pVCpu->pgm.s.CTX_SUFF(pStats)->StatRZGuestCR3WriteConflict));
452	LogFlow(("pgmPoolMonitorChainChanging: Detected conflict at iShw=%#x!\n", iShw));
453	break;
454	}
455	#endif /* !IN_RING0 */
456	/*
457	* Causes trouble when the guest uses a PDE to refer to the whole page table level
458	* structure. (Invalidate here; faults later on when it tries to change the page
459	* table entries -> recheck; probably only applies to the RC case.)
460	*/
461	# ifndef IN_RING0
462	else
463	# endif /* !IN_RING0 */
464	{
465	if (uShw.pPDPae->a[iShw].n.u1Present)
466	{
467	LogFlow(("pgmPoolMonitorChainChanging: pae pd iShw=%#x: %RX64 -> freeing it!\n", iShw, uShw.pPDPae->a[iShw].u));
468	pgmPoolFree(pVM,
469	uShw.pPDPae->a[iShw].u & X86_PDE_PAE_PG_MASK,
470	pPage->idx,
471	iShw);
472	ASMAtomicWriteU64(&uShw.pPDPae->a[iShw].u, 0);
473	}
474	}
475	/* paranoia / a bit assumptive. */
476	if ( (off & 7)
477	&& (off & 7) + cbWrite > sizeof(X86PDEPAE))
478	{
479	const unsigned iShw2 = (off + cbWrite - 1) / sizeof(X86PDEPAE);
480	AssertBreak(iShw2 < RT_ELEMENTS(uShw.pPDPae->a));
481
482	#ifndef IN_RING0
483	if ( iShw2 != iShw
484	&& uShw.pPDPae->a[iShw2].u & PGM_PDFLAGS_MAPPING)
485	{
486	Assert(pgmMapAreMappingsEnabled(pVM));
487	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
488	STAM_COUNTER_INC(&(pVCpu->pgm.s.CTX_SUFF(pStats)->StatRZGuestCR3WriteConflict));
489	LogFlow(("pgmPoolMonitorChainChanging: Detected conflict at iShw2=%#x!\n", iShw2));
490	break;
491	}
492	#endif /* !IN_RING0 */
493	# ifndef IN_RING0
494	else
495	# endif /* !IN_RING0 */
496	if (uShw.pPDPae->a[iShw2].n.u1Present)
497	{
498	LogFlow(("pgmPoolMonitorChainChanging: pae pd iShw2=%#x: %RX64 -> freeing it!\n", iShw2, uShw.pPDPae->a[iShw2].u));
499	pgmPoolFree(pVM,
500	uShw.pPDPae->a[iShw2].u & X86_PDE_PAE_PG_MASK,
501	pPage->idx,
502	iShw2);
503	ASMAtomicWriteU64(&uShw.pPDPae->a[iShw2].u, 0);
504	}
505	}
506	break;
507	}
508
509	case PGMPOOLKIND_PAE_PDPT:
510	{
511	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPDPT));
512	/*
513	* Hopefully this doesn't happen very often:
514	* - touching unused parts of the page
515	* - messing with the bits of pd pointers without changing the physical address
516	*/
517	/* PDPT roots are not page aligned; 32 byte only! */
518	const unsigned offPdpt = GCPhysFault - pPage->GCPhys;
519
520	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
521	const unsigned iShw = offPdpt / sizeof(X86PDPE);
522	if (iShw < X86_PG_PAE_PDPE_ENTRIES) /* don't use RT_ELEMENTS(uShw.pPDPT->a), because that's for long mode only */
523	{
524	# ifndef IN_RING0
525	if (uShw.pPDPT->a[iShw].u & PGM_PLXFLAGS_MAPPING)
526	{
527	Assert(pgmMapAreMappingsEnabled(pVM));
528	STAM_COUNTER_INC(&(pVCpu->pgm.s.CTX_SUFF(pStats)->StatRZGuestCR3WriteConflict));
529	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
530	LogFlow(("pgmPoolMonitorChainChanging: Detected pdpt conflict at iShw=%#x!\n", iShw));
531	break;
532	}
533	# endif /* !IN_RING0 */
534	# ifndef IN_RING0
535	else
536	# endif /* !IN_RING0 */
537	if (uShw.pPDPT->a[iShw].n.u1Present)
538	{
539	LogFlow(("pgmPoolMonitorChainChanging: pae pdpt iShw=%#x: %RX64 -> freeing it!\n", iShw, uShw.pPDPT->a[iShw].u));
540	pgmPoolFree(pVM,
541	uShw.pPDPT->a[iShw].u & X86_PDPE_PG_MASK,
542	pPage->idx,
543	iShw);
544	ASMAtomicWriteU64(&uShw.pPDPT->a[iShw].u, 0);
545	}
546
547	/* paranoia / a bit assumptive. */
548	if ( (offPdpt & 7)
549	&& (offPdpt & 7) + cbWrite > sizeof(X86PDPE))
550	{
551	const unsigned iShw2 = (offPdpt + cbWrite - 1) / sizeof(X86PDPE);
552	if ( iShw2 != iShw
553	&& iShw2 < X86_PG_PAE_PDPE_ENTRIES)
554	{
555	# ifndef IN_RING0
556	if (uShw.pPDPT->a[iShw2].u & PGM_PLXFLAGS_MAPPING)
557	{
558	Assert(pgmMapAreMappingsEnabled(pVM));
559	STAM_COUNTER_INC(&(pVCpu->pgm.s.CTX_SUFF(pStats)->StatRZGuestCR3WriteConflict));
560	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
561	LogFlow(("pgmPoolMonitorChainChanging: Detected conflict at iShw2=%#x!\n", iShw2));
562	break;
563	}
564	# endif /* !IN_RING0 */
565	# ifndef IN_RING0
566	else
567	# endif /* !IN_RING0 */
568	if (uShw.pPDPT->a[iShw2].n.u1Present)
569	{
570	LogFlow(("pgmPoolMonitorChainChanging: pae pdpt iShw=%#x: %RX64 -> freeing it!\n", iShw2, uShw.pPDPT->a[iShw2].u));
571	pgmPoolFree(pVM,
572	uShw.pPDPT->a[iShw2].u & X86_PDPE_PG_MASK,
573	pPage->idx,
574	iShw2);
575	ASMAtomicWriteU64(&uShw.pPDPT->a[iShw2].u, 0);
576	}
577	}
578	}
579	}
580	break;
581	}
582
583	#ifndef IN_RC
584	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
585	{
586	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPD));
587	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
588	const unsigned iShw = off / sizeof(X86PDEPAE);
589	Assert(!(uShw.pPDPae->a[iShw].u & PGM_PDFLAGS_MAPPING));
590	if (uShw.pPDPae->a[iShw].n.u1Present)
591	{
592	LogFlow(("pgmPoolMonitorChainChanging: pae pd iShw=%#x: %RX64 -> freeing it!\n", iShw, uShw.pPDPae->a[iShw].u));
593	pgmPoolFree(pVM,
594	uShw.pPDPae->a[iShw].u & X86_PDE_PAE_PG_MASK,
595	pPage->idx,
596	iShw);
597	ASMAtomicWriteU64(&uShw.pPDPae->a[iShw].u, 0);
598	}
599	/* paranoia / a bit assumptive. */
600	if ( (off & 7)
601	&& (off & 7) + cbWrite > sizeof(X86PDEPAE))
602	{
603	const unsigned iShw2 = (off + cbWrite - 1) / sizeof(X86PDEPAE);
604	AssertBreak(iShw2 < RT_ELEMENTS(uShw.pPDPae->a));
605
606	Assert(!(uShw.pPDPae->a[iShw2].u & PGM_PDFLAGS_MAPPING));
607	if (uShw.pPDPae->a[iShw2].n.u1Present)
608	{
609	LogFlow(("pgmPoolMonitorChainChanging: pae pd iShw2=%#x: %RX64 -> freeing it!\n", iShw2, uShw.pPDPae->a[iShw2].u));
610	pgmPoolFree(pVM,
611	uShw.pPDPae->a[iShw2].u & X86_PDE_PAE_PG_MASK,
612	pPage->idx,
613	iShw2);
614	ASMAtomicWriteU64(&uShw.pPDPae->a[iShw2].u, 0);
615	}
616	}
617	break;
618	}
619
620	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
621	{
622	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPDPT));
623	/*
624	* Hopefully this doesn't happen very often:
625	* - messing with the bits of pd pointers without changing the physical address
626	*/
627	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
628	const unsigned iShw = off / sizeof(X86PDPE);
629	if (uShw.pPDPT->a[iShw].n.u1Present)
630	{
631	LogFlow(("pgmPoolMonitorChainChanging: pdpt iShw=%#x: %RX64 -> freeing it!\n", iShw, uShw.pPDPT->a[iShw].u));
632	pgmPoolFree(pVM, uShw.pPDPT->a[iShw].u & X86_PDPE_PG_MASK, pPage->idx, iShw);
633	ASMAtomicWriteU64(&uShw.pPDPT->a[iShw].u, 0);
634	}
635	/* paranoia / a bit assumptive. */
636	if ( (off & 7)
637	&& (off & 7) + cbWrite > sizeof(X86PDPE))
638	{
639	const unsigned iShw2 = (off + cbWrite - 1) / sizeof(X86PDPE);
640	if (uShw.pPDPT->a[iShw2].n.u1Present)
641	{
642	LogFlow(("pgmPoolMonitorChainChanging: pdpt iShw2=%#x: %RX64 -> freeing it!\n", iShw2, uShw.pPDPT->a[iShw2].u));
643	pgmPoolFree(pVM, uShw.pPDPT->a[iShw2].u & X86_PDPE_PG_MASK, pPage->idx, iShw2);
644	ASMAtomicWriteU64(&uShw.pPDPT->a[iShw2].u, 0);
645	}
646	}
647	break;
648	}
649
650	case PGMPOOLKIND_64BIT_PML4:
651	{
652	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FaultPML4));
653	/*
654	* Hopefully this doesn't happen very often:
655	* - messing with the bits of pd pointers without changing the physical address
656	*/
657	uShw.pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
658	const unsigned iShw = off / sizeof(X86PDPE);
659	if (uShw.pPML4->a[iShw].n.u1Present)
660	{
661	LogFlow(("pgmPoolMonitorChainChanging: pml4 iShw=%#x: %RX64 -> freeing it!\n", iShw, uShw.pPML4->a[iShw].u));
662	pgmPoolFree(pVM, uShw.pPML4->a[iShw].u & X86_PML4E_PG_MASK, pPage->idx, iShw);
663	ASMAtomicWriteU64(&uShw.pPML4->a[iShw].u, 0);
664	}
665	/* paranoia / a bit assumptive. */
666	if ( (off & 7)
667	&& (off & 7) + cbWrite > sizeof(X86PDPE))
668	{
669	const unsigned iShw2 = (off + cbWrite - 1) / sizeof(X86PML4E);
670	if (uShw.pPML4->a[iShw2].n.u1Present)
671	{
672	LogFlow(("pgmPoolMonitorChainChanging: pml4 iShw2=%#x: %RX64 -> freeing it!\n", iShw2, uShw.pPML4->a[iShw2].u));
673	pgmPoolFree(pVM, uShw.pPML4->a[iShw2].u & X86_PML4E_PG_MASK, pPage->idx, iShw2);
674	ASMAtomicWriteU64(&uShw.pPML4->a[iShw2].u, 0);
675	}
676	}
677	break;
678	}
679	#endif /* IN_RING0 */
680
681	default:
682	AssertFatalMsgFailed(("enmKind=%d\n", pPage->enmKind));
683	}
684	PGM_DYNMAP_UNUSED_HINT_VM(pVM, uShw.pv);
685
686	/* next */
687	if (pPage->iMonitoredNext == NIL_PGMPOOL_IDX)
688	return;
689	pPage = &pPool->aPages[pPage->iMonitoredNext];
690	}
691	}
692
693	# ifndef IN_RING3
694
695	/**
696	* Checks if a access could be a fork operation in progress.
697	*
698	* Meaning, that the guest is setting up the parent process for Copy-On-Write.
699	*
700	* @returns true if it's likely that we're forking, otherwise false.
701	* @param pPool The pool.
702	* @param pDis The disassembled instruction.
703	* @param offFault The access offset.
704	*/
705	DECLINLINE(bool) pgmPoolMonitorIsForking(PPGMPOOL pPool, PDISCPUSTATE pDis, unsigned offFault)
706	{
707	/*
708	* i386 linux is using btr to clear X86_PTE_RW.
709	* The functions involved are (2.6.16 source inspection):
710	* clear_bit
711	* ptep_set_wrprotect
712	* copy_one_pte
713	* copy_pte_range
714	* copy_pmd_range
715	* copy_pud_range
716	* copy_page_range
717	* dup_mmap
718	* dup_mm
719	* copy_mm
720	* copy_process
721	* do_fork
722	*/
723	if ( pDis->pCurInstr->uOpcode == OP_BTR
724	&& !(offFault & 4)
725	/** @todo Validate that the bit index is X86_PTE_RW. */
726	)
727	{
728	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,Fork));
729	return true;
730	}
731	return false;
732	}
733
734
735	/**
736	* Determine whether the page is likely to have been reused.
737	*
738	* @returns true if we consider the page as being reused for a different purpose.
739	* @returns false if we consider it to still be a paging page.
740	* @param pVM Pointer to the VM.
741	* @param pVCpu VMCPU Handle.
742	* @param pRegFrame Trap register frame.
743	* @param pDis The disassembly info for the faulting instruction.
744	* @param pvFault The fault address.
745	*
746	* @remark The REP prefix check is left to the caller because of STOSD/W.
747	*/
748	DECLINLINE(bool) pgmPoolMonitorIsReused(PVM pVM, PVMCPU pVCpu, PCPUMCTXCORE pRegFrame, PDISCPUSTATE pDis, RTGCPTR pvFault)
749	{
750	#ifndef IN_RC
751	/** @todo could make this general, faulting close to rsp should be a safe reuse heuristic. */
752	if ( HWACCMHasPendingIrq(pVM)
753	&& (pRegFrame->rsp - pvFault) < 32)
754	{
755	/* Fault caused by stack writes while trying to inject an interrupt event. */
756	Log(("pgmPoolMonitorIsReused: reused %RGv for interrupt stack (rsp=%RGv).\n", pvFault, pRegFrame->rsp));
757	return true;
758	}
759	#else
760	NOREF(pVM); NOREF(pvFault);
761	#endif
762
763	LogFlow(("Reused instr %RGv %d at %RGv param1.fUse=%llx param1.reg=%d\n", pRegFrame->rip, pDis->pCurInstr->uOpcode, pvFault, pDis->Param1.fUse, pDis->Param1.Base.idxGenReg));
764
765	/* Non-supervisor mode write means it's used for something else. */
766	if (CPUMGetGuestCPL(pVCpu, pRegFrame) != 0)
767	return true;
768
769	switch (pDis->pCurInstr->uOpcode)
770	{
771	/* call implies the actual push of the return address faulted */
772	case OP_CALL:
773	Log4(("pgmPoolMonitorIsReused: CALL\n"));
774	return true;
775	case OP_PUSH:
776	Log4(("pgmPoolMonitorIsReused: PUSH\n"));
777	return true;
778	case OP_PUSHF:
779	Log4(("pgmPoolMonitorIsReused: PUSHF\n"));
780	return true;
781	case OP_PUSHA:
782	Log4(("pgmPoolMonitorIsReused: PUSHA\n"));
783	return true;
784	case OP_FXSAVE:
785	Log4(("pgmPoolMonitorIsReused: FXSAVE\n"));
786	return true;
787	case OP_MOVNTI: /* solaris - block_zero_no_xmm */
788	Log4(("pgmPoolMonitorIsReused: MOVNTI\n"));
789	return true;
790	case OP_MOVNTDQ: /* solaris - hwblkclr & hwblkpagecopy */
791	Log4(("pgmPoolMonitorIsReused: MOVNTDQ\n"));
792	return true;
793	case OP_MOVSWD:
794	case OP_STOSWD:
795	if ( pDis->fPrefix == (DISPREFIX_REP\|DISPREFIX_REX)
796	&& pRegFrame->rcx >= 0x40
797	)
798	{
799	Assert(pDis->uCpuMode == DISCPUMODE_64BIT);
800
801	Log(("pgmPoolMonitorIsReused: OP_STOSQ\n"));
802	return true;
803	}
804	return false;
805	}
806	if ( ( (pDis->Param1.fUse & DISUSE_REG_GEN32)
807	\|\| (pDis->Param1.fUse & DISUSE_REG_GEN64))
808	&& (pDis->Param1.Base.idxGenReg == DISGREG_ESP))
809	{
810	Log4(("pgmPoolMonitorIsReused: ESP\n"));
811	return true;
812	}
813
814	return false;
815	}
816
817
818	/**
819	* Flushes the page being accessed.
820	*
821	* @returns VBox status code suitable for scheduling.
822	* @param pVM Pointer to the VM.
823	* @param pVCpu The VMCPU handle.
824	* @param pPool The pool.
825	* @param pPage The pool page (head).
826	* @param pDis The disassembly of the write instruction.
827	* @param pRegFrame The trap register frame.
828	* @param GCPhysFault The fault address as guest physical address.
829	* @param pvFault The fault address.
830	* @todo VBOXSTRICTRC
831	*/
832	static int pgmPoolAccessHandlerFlush(PVM pVM, PVMCPU pVCpu, PPGMPOOL pPool, PPGMPOOLPAGE pPage, PDISCPUSTATE pDis,
833	PCPUMCTXCORE pRegFrame, RTGCPHYS GCPhysFault, RTGCPTR pvFault)
834	{
835	NOREF(GCPhysFault);
836
837	/*
838	* First, do the flushing.
839	*/
840	int rc = pgmPoolMonitorChainFlush(pPool, pPage);
841
842	/*
843	* Emulate the instruction (xp/w2k problem, requires pc/cr2/sp detection).
844	* Must do this in raw mode (!); XP boot will fail otherwise.
845	*/
846	VBOXSTRICTRC rc2 = EMInterpretInstructionDisasState(pVCpu, pDis, pRegFrame, pvFault, EMCODETYPE_ALL);
847	if (RT_SUCCESS(rc2))
848	AssertMsg(rc2 == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rc2))); /* ASSUMES no complicated stuff here. */
849	else if (rc2 == VERR_EM_INTERPRETER)
850	{
851	#ifdef IN_RC
852	if (PATMIsPatchGCAddr(pVM, pRegFrame->eip))
853	{
854	LogFlow(("pgmPoolAccessHandlerPTWorker: Interpretation failed for patch code %04x:%RGv, ignoring.\n",
855	pRegFrame->cs, (RTGCPTR)pRegFrame->eip));
856	rc = VINF_SUCCESS;
857	STAM_COUNTER_INC(&pPool->StatMonitorRZIntrFailPatch2);
858	}
859	else
860	#endif
861	{
862	rc = VINF_EM_RAW_EMULATE_INSTR;
863	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,EmulateInstr));
864	}
865	}
866	else
867	rc = VBOXSTRICTRC_VAL(rc2);
868
869	LogFlow(("pgmPoolAccessHandlerPT: returns %Rrc (flushed)\n", rc));
870	return rc;
871	}
872
873
874	/**
875	* Handles the STOSD write accesses.
876	*
877	* @returns VBox status code suitable for scheduling.
878	* @param pVM Pointer to the VM.
879	* @param pPool The pool.
880	* @param pPage The pool page (head).
881	* @param pDis The disassembly of the write instruction.
882	* @param pRegFrame The trap register frame.
883	* @param GCPhysFault The fault address as guest physical address.
884	* @param pvFault The fault address.
885	*/
886	DECLINLINE(int) pgmPoolAccessHandlerSTOSD(PVM pVM, PPGMPOOL pPool, PPGMPOOLPAGE pPage, PDISCPUSTATE pDis,
887	PCPUMCTXCORE pRegFrame, RTGCPHYS GCPhysFault, RTGCPTR pvFault)
888	{
889	unsigned uIncrement = pDis->Param1.cb;
890	NOREF(pVM);
891
892	Assert(pDis->uCpuMode == DISCPUMODE_32BIT \|\| pDis->uCpuMode == DISCPUMODE_64BIT);
893	Assert(pRegFrame->rcx <= 0x20);
894
895	#ifdef VBOX_STRICT
896	if (pDis->uOpMode == DISCPUMODE_32BIT)
897	Assert(uIncrement == 4);
898	else
899	Assert(uIncrement == 8);
900	#endif
901
902	Log3(("pgmPoolAccessHandlerSTOSD\n"));
903
904	/*
905	* Increment the modification counter and insert it into the list
906	* of modified pages the first time.
907	*/
908	if (!pPage->cModifications++)
909	pgmPoolMonitorModifiedInsert(pPool, pPage);
910
911	/*
912	* Execute REP STOSD.
913	*
914	* This ASSUMES that we're not invoked by Trap0e on in a out-of-sync
915	* write situation, meaning that it's safe to write here.
916	*/
917	PVMCPU pVCpu = VMMGetCpu(pPool->CTX_SUFF(pVM));
918	RTGCUINTPTR pu32 = (RTGCUINTPTR)pvFault;
919	while (pRegFrame->rcx)
920	{
921	#if defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) \|\| defined(IN_RC)
922	uint32_t iPrevSubset = PGMRZDynMapPushAutoSubset(pVCpu);
923	pgmPoolMonitorChainChanging(pVCpu, pPool, pPage, GCPhysFault, (RTGCPTR)pu32, uIncrement);
924	PGMRZDynMapPopAutoSubset(pVCpu, iPrevSubset);
925	#else
926	pgmPoolMonitorChainChanging(pVCpu, pPool, pPage, GCPhysFault, (RTGCPTR)pu32, uIncrement);
927	#endif
928	#ifdef IN_RC
929	(uint32_t )(uintptr_t)pu32 = pRegFrame->eax;
930	#else
931	PGMPhysSimpleWriteGCPhys(pVM, GCPhysFault, &pRegFrame->rax, uIncrement);
932	#endif
933	pu32 += uIncrement;
934	GCPhysFault += uIncrement;
935	pRegFrame->rdi += uIncrement;
936	pRegFrame->rcx--;
937	}
938	pRegFrame->rip += pDis->cbInstr;
939
940	LogFlow(("pgmPoolAccessHandlerSTOSD: returns\n"));
941	return VINF_SUCCESS;
942	}
943
944
945	/**
946	* Handles the simple write accesses.
947	*
948	* @returns VBox status code suitable for scheduling.
949	* @param pVM Pointer to the VM.
950	* @param pVCpu The VMCPU handle.
951	* @param pPool The pool.
952	* @param pPage The pool page (head).
953	* @param pDis The disassembly of the write instruction.
954	* @param pRegFrame The trap register frame.
955	* @param GCPhysFault The fault address as guest physical address.
956	* @param pvFault The fault address.
957	* @param pfReused Reused state (in/out)
958	*/
959	DECLINLINE(int) pgmPoolAccessHandlerSimple(PVM pVM, PVMCPU pVCpu, PPGMPOOL pPool, PPGMPOOLPAGE pPage, PDISCPUSTATE pDis,
960	PCPUMCTXCORE pRegFrame, RTGCPHYS GCPhysFault, RTGCPTR pvFault, bool *pfReused)
961	{
962	Log3(("pgmPoolAccessHandlerSimple\n"));
963	NOREF(pfReused); /* initialized by caller */
964
965	/*
966	* Increment the modification counter and insert it into the list
967	* of modified pages the first time.
968	*/
969	if (!pPage->cModifications++)
970	pgmPoolMonitorModifiedInsert(pPool, pPage);
971
972	/*
973	* Clear all the pages. ASSUMES that pvFault is readable.
974	*/
975	#if defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) \|\| defined(IN_RC)
976	uint32_t iPrevSubset = PGMRZDynMapPushAutoSubset(pVCpu);
977	pgmPoolMonitorChainChanging(pVCpu, pPool, pPage, GCPhysFault, pvFault, DISGetParamSize(pDis, &pDis->Param1));
978	PGMRZDynMapPopAutoSubset(pVCpu, iPrevSubset);
979	#else
980	pgmPoolMonitorChainChanging(pVCpu, pPool, pPage, GCPhysFault, pvFault, DISGetParamSize(pDis, &pDis->Param1));
981	#endif
982
983	/*
984	* Interpret the instruction.
985	*/
986	VBOXSTRICTRC rc = EMInterpretInstructionDisasState(pVCpu, pDis, pRegFrame, pvFault, EMCODETYPE_ALL);
987	if (RT_SUCCESS(rc))
988	AssertMsg(rc == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rc))); /* ASSUMES no complicated stuff here. */
989	else if (rc == VERR_EM_INTERPRETER)
990	{
991	LogFlow(("pgmPoolAccessHandlerPTWorker: Interpretation failed for %04x:%RGv - opcode=%d\n",
992	pRegFrame->cs, (RTGCPTR)pRegFrame->rip, pDis->pCurInstr->uOpcode));
993	rc = VINF_EM_RAW_EMULATE_INSTR;
994	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,EmulateInstr));
995	}
996
997	#if 0 /* experimental code */
998	if (rc == VINF_SUCCESS)
999	{
1000	switch (pPage->enmKind)
1001	{
1002	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
1003	{
1004	X86PTEPAE GstPte;
1005	int rc = pgmPoolPhysSimpleReadGCPhys(pVM, &GstPte, pvFault, GCPhysFault, sizeof(GstPte));
1006	AssertRC(rc);
1007
1008	/* Check the new value written by the guest. If present and with a bogus physical address, then
1009	* it's fairly safe to assume the guest is reusing the PT.
1010	*/
1011	if (GstPte.n.u1Present)
1012	{
1013	RTHCPHYS HCPhys = -1;
1014	int rc = PGMPhysGCPhys2HCPhys(pVM, GstPte.u & X86_PTE_PAE_PG_MASK, &HCPhys);
1015	if (rc != VINF_SUCCESS)
1016	{
1017	*pfReused = true;
1018	STAM_COUNTER_INC(&pPool->StatForceFlushReused);
1019	}
1020	}
1021	break;
1022	}
1023	}
1024	}
1025	#endif
1026
1027	LogFlow(("pgmPoolAccessHandlerSimple: returns %Rrc\n", VBOXSTRICTRC_VAL(rc)));
1028	return VBOXSTRICTRC_VAL(rc);
1029	}
1030
1031
1032	/**
1033	* \#PF Handler callback for PT write accesses.
1034	*
1035	* @returns VBox status code (appropriate for GC return).
1036	* @param pVM Pointer to the VM.
1037	* @param uErrorCode CPU Error code.
1038	* @param pRegFrame Trap register frame.
1039	* NULL on DMA and other non CPU access.
1040	* @param pvFault The fault address (cr2).
1041	* @param GCPhysFault The GC physical address corresponding to pvFault.
1042	* @param pvUser User argument.
1043	*/
1044	DECLEXPORT(int) pgmPoolAccessHandler(PVM pVM, RTGCUINT uErrorCode, PCPUMCTXCORE pRegFrame, RTGCPTR pvFault,
1045	RTGCPHYS GCPhysFault, void *pvUser)
1046	{
1047	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pPool)->CTX_SUFF_Z(StatMonitor), a);
1048	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1049	PPGMPOOLPAGE pPage = (PPGMPOOLPAGE)pvUser;
1050	PVMCPU pVCpu = VMMGetCpu(pVM);
1051	unsigned cMaxModifications;
1052	bool fForcedFlush = false;
1053	NOREF(uErrorCode);
1054
1055	LogFlow(("pgmPoolAccessHandler: pvFault=%RGv pPage=%p:{.idx=%d} GCPhysFault=%RGp\n", pvFault, pPage, pPage->idx, GCPhysFault));
1056
1057	pgmLock(pVM);
1058	if (PHYS_PAGE_ADDRESS(GCPhysFault) != PHYS_PAGE_ADDRESS(pPage->GCPhys))
1059	{
1060	/* Pool page changed while we were waiting for the lock; ignore. */
1061	Log(("CPU%d: pgmPoolAccessHandler pgm pool page for %RGp changed (to %RGp) while waiting!\n", pVCpu->idCpu, PHYS_PAGE_ADDRESS(GCPhysFault), PHYS_PAGE_ADDRESS(pPage->GCPhys)));
1062	STAM_PROFILE_STOP_EX(&pVM->pgm.s.CTX_SUFF(pPool)->CTX_SUFF_Z(StatMonitor), &pPool->CTX_MID_Z(StatMonitor,Handled), a);
1063	pgmUnlock(pVM);
1064	return VINF_SUCCESS;
1065	}
1066	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
1067	if (pPage->fDirty)
1068	{
1069	Assert(VMCPU_FF_ISSET(pVCpu, VMCPU_FF_TLB_FLUSH));
1070	pgmUnlock(pVM);
1071	return VINF_SUCCESS; /* SMP guest case where we were blocking on the pgm lock while the same page was being marked dirty. */
1072	}
1073	#endif
1074
1075	#if 0 /* test code defined(VBOX_STRICT) && defined(PGMPOOL_WITH_OPTIMIZED_DIRTY_PT) */
1076	if (pPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_PAE_PT)
1077	{
1078	void *pvShw = PGMPOOL_PAGE_2_PTR(pPool->CTX_SUFF(pVM), pPage);
1079	void *pvGst;
1080	int rc = PGM_GCPHYS_2_PTR(pPool->CTX_SUFF(pVM), pPage->GCPhys, &pvGst); AssertReleaseRC(rc);
1081	pgmPoolTrackCheckPTPaePae(pPool, pPage, (PPGMSHWPTPAE)pvShw, (PCX86PTPAE)pvGst);
1082	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvGst);
1083	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvShw);
1084	}
1085	#endif
1086
1087	/*
1088	* Disassemble the faulting instruction.
1089	*/
1090	PDISCPUSTATE pDis = &pVCpu->pgm.s.DisState;
1091	int rc = EMInterpretDisasOne(pVM, pVCpu, pRegFrame, pDis, NULL);
1092	if (RT_UNLIKELY(rc != VINF_SUCCESS))
1093	{
1094	AssertMsg(rc == VERR_PAGE_NOT_PRESENT \|\| rc == VERR_PAGE_TABLE_NOT_PRESENT, ("Unexpected rc %d\n", rc));
1095	pgmUnlock(pVM);
1096	return rc;
1097	}
1098
1099	Assert(pPage->enmKind != PGMPOOLKIND_FREE);
1100
1101	/*
1102	* We should ALWAYS have the list head as user parameter. This
1103	* is because we use that page to record the changes.
1104	*/
1105	Assert(pPage->iMonitoredPrev == NIL_PGMPOOL_IDX);
1106
1107	#ifdef IN_RING0
1108	/* Maximum nr of modifications depends on the page type. */
1109	if ( pPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_PAE_PT
1110	\|\| pPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_32BIT_PT)
1111	cMaxModifications = 4;
1112	else
1113	cMaxModifications = 24;
1114	#else
1115	cMaxModifications = 48;
1116	#endif
1117
1118	/*
1119	* Incremental page table updates should weigh more than random ones.
1120	* (Only applies when started from offset 0)
1121	*/
1122	pVCpu->pgm.s.cPoolAccessHandler++;
1123	if ( pPage->GCPtrLastAccessHandlerRip >= pRegFrame->rip - 0x40 /* observed loops in Windows 7 x64 */
1124	&& pPage->GCPtrLastAccessHandlerRip < pRegFrame->rip + 0x40
1125	&& pvFault == (pPage->GCPtrLastAccessHandlerFault + pDis->Param1.cb)
1126	&& pVCpu->pgm.s.cPoolAccessHandler == pPage->cLastAccessHandler + 1)
1127	{
1128	Log(("Possible page reuse cMods=%d -> %d (locked=%d type=%s)\n", pPage->cModifications, pPage->cModifications * 2, pgmPoolIsPageLocked(pPage), pgmPoolPoolKindToStr(pPage->enmKind)));
1129	Assert(pPage->cModifications < 32000);
1130	pPage->cModifications = pPage->cModifications * 2;
1131	pPage->GCPtrLastAccessHandlerFault = pvFault;
1132	pPage->cLastAccessHandler = pVCpu->pgm.s.cPoolAccessHandler;
1133	if (pPage->cModifications >= cMaxModifications)
1134	{
1135	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FlushReinit));
1136	fForcedFlush = true;
1137	}
1138	}
1139
1140	if (pPage->cModifications >= cMaxModifications)
1141	Log(("Mod overflow %RGv cMods=%d (locked=%d type=%s)\n", pvFault, pPage->cModifications, pgmPoolIsPageLocked(pPage), pgmPoolPoolKindToStr(pPage->enmKind)));
1142
1143	/*
1144	* Check if it's worth dealing with.
1145	*/
1146	bool fReused = false;
1147	bool fNotReusedNotForking = false;
1148	if ( ( pPage->cModifications < cMaxModifications /** @todo #define / /* @todo need to check that it's not mapping EIP. / /* @todo adjust this! */
1149	\|\| pgmPoolIsPageLocked(pPage)
1150	)
1151	&& !(fReused = pgmPoolMonitorIsReused(pVM, pVCpu, pRegFrame, pDis, pvFault))
1152	&& !pgmPoolMonitorIsForking(pPool, pDis, GCPhysFault & PAGE_OFFSET_MASK))
1153	{
1154	/*
1155	* Simple instructions, no REP prefix.
1156	*/
1157	if (!(pDis->fPrefix & (DISPREFIX_REP \| DISPREFIX_REPNE)))
1158	{
1159	rc = pgmPoolAccessHandlerSimple(pVM, pVCpu, pPool, pPage, pDis, pRegFrame, GCPhysFault, pvFault, &fReused);
1160	if (fReused)
1161	goto flushPage;
1162
1163	/* A mov instruction to change the first page table entry will be remembered so we can detect
1164	* full page table changes early on. This will reduce the amount of unnecessary traps we'll take.
1165	*/
1166	if ( rc == VINF_SUCCESS
1167	&& !pPage->cLocked /* only applies to unlocked pages as we can't free locked ones (e.g. cr3 root). */
1168	&& pDis->pCurInstr->uOpcode == OP_MOV
1169	&& (pvFault & PAGE_OFFSET_MASK) == 0)
1170	{
1171	pPage->GCPtrLastAccessHandlerFault = pvFault;
1172	pPage->cLastAccessHandler = pVCpu->pgm.s.cPoolAccessHandler;
1173	pPage->GCPtrLastAccessHandlerRip = pRegFrame->rip;
1174	/* Make sure we don't kick out a page too quickly. */
1175	if (pPage->cModifications > 8)
1176	pPage->cModifications = 2;
1177	}
1178	else if (pPage->GCPtrLastAccessHandlerFault == pvFault)
1179	{
1180	/* ignore the 2nd write to this page table entry. */
1181	pPage->cLastAccessHandler = pVCpu->pgm.s.cPoolAccessHandler;
1182	}
1183	else
1184	{
1185	pPage->GCPtrLastAccessHandlerFault = NIL_RTGCPTR;
1186	pPage->GCPtrLastAccessHandlerRip = 0;
1187	}
1188
1189	STAM_PROFILE_STOP_EX(&pVM->pgm.s.CTX_SUFF(pPool)->CTX_SUFF_Z(StatMonitor), &pPool->CTX_MID_Z(StatMonitor,Handled), a);
1190	pgmUnlock(pVM);
1191	return rc;
1192	}
1193
1194	/*
1195	* Windows is frequently doing small memset() operations (netio test 4k+).
1196	* We have to deal with these or we'll kill the cache and performance.
1197	*/
1198	if ( pDis->pCurInstr->uOpcode == OP_STOSWD
1199	&& !pRegFrame->eflags.Bits.u1DF
1200	&& pDis->uOpMode == pDis->uCpuMode
1201	&& pDis->uAddrMode == pDis->uCpuMode)
1202	{
1203	bool fValidStosd = false;
1204
1205	if ( pDis->uCpuMode == DISCPUMODE_32BIT
1206	&& pDis->fPrefix == DISPREFIX_REP
1207	&& pRegFrame->ecx <= 0x20
1208	&& pRegFrame->ecx * 4 <= PAGE_SIZE - ((uintptr_t)pvFault & PAGE_OFFSET_MASK)
1209	&& !((uintptr_t)pvFault & 3)
1210	&& (pRegFrame->eax == 0 \|\| pRegFrame->eax == 0x80) /* the two values observed. */
1211	)
1212	{
1213	fValidStosd = true;
1214	pRegFrame->rcx &= 0xffffffff; /* paranoia */
1215	}
1216	else
1217	if ( pDis->uCpuMode == DISCPUMODE_64BIT
1218	&& pDis->fPrefix == (DISPREFIX_REP \| DISPREFIX_REX)
1219	&& pRegFrame->rcx <= 0x20
1220	&& pRegFrame->rcx * 8 <= PAGE_SIZE - ((uintptr_t)pvFault & PAGE_OFFSET_MASK)
1221	&& !((uintptr_t)pvFault & 7)
1222	&& (pRegFrame->rax == 0 \|\| pRegFrame->rax == 0x80) /* the two values observed. */
1223	)
1224	{
1225	fValidStosd = true;
1226	}
1227
1228	if (fValidStosd)
1229	{
1230	rc = pgmPoolAccessHandlerSTOSD(pVM, pPool, pPage, pDis, pRegFrame, GCPhysFault, pvFault);
1231	STAM_PROFILE_STOP_EX(&pVM->pgm.s.CTX_SUFF(pPool)->CTX_SUFF_Z(StatMonitor), &pPool->CTX_MID_Z(StatMonitor,RepStosd), a);
1232	pgmUnlock(pVM);
1233	return rc;
1234	}
1235	}
1236
1237	/* REP prefix, don't bother. */
1238	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,RepPrefix));
1239	Log4(("pgmPoolAccessHandler: eax=%#x ecx=%#x edi=%#x esi=%#x rip=%RGv opcode=%d prefix=%#x\n",
1240	pRegFrame->eax, pRegFrame->ecx, pRegFrame->edi, pRegFrame->esi, (RTGCPTR)pRegFrame->rip, pDis->pCurInstr->uOpcode, pDis->fPrefix));
1241	fNotReusedNotForking = true;
1242	}
1243
1244	#if defined(PGMPOOL_WITH_OPTIMIZED_DIRTY_PT) && defined(IN_RING0)
1245	/* E.g. Windows 7 x64 initializes page tables and touches some pages in the table during the process. This
1246	* leads to pgm pool trashing and an excessive amount of write faults due to page monitoring.
1247	*/
1248	if ( pPage->cModifications >= cMaxModifications
1249	&& !fForcedFlush
1250	&& (pPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_PAE_PT \|\| pPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_32BIT_PT)
1251	&& ( fNotReusedNotForking
1252	\|\| ( !pgmPoolMonitorIsReused(pVM, pVCpu, pRegFrame, pDis, pvFault)
1253	&& !pgmPoolMonitorIsForking(pPool, pDis, GCPhysFault & PAGE_OFFSET_MASK))
1254	)
1255	)
1256	{
1257	Assert(!pgmPoolIsPageLocked(pPage));
1258	Assert(pPage->fDirty == false);
1259
1260	/* Flush any monitored duplicates as we will disable write protection. */
1261	if ( pPage->iMonitoredNext != NIL_PGMPOOL_IDX
1262	\|\| pPage->iMonitoredPrev != NIL_PGMPOOL_IDX)
1263	{
1264	PPGMPOOLPAGE pPageHead = pPage;
1265
1266	/* Find the monitor head. */
1267	while (pPageHead->iMonitoredPrev != NIL_PGMPOOL_IDX)
1268	pPageHead = &pPool->aPages[pPageHead->iMonitoredPrev];
1269
1270	while (pPageHead)
1271	{
1272	unsigned idxNext = pPageHead->iMonitoredNext;
1273
1274	if (pPageHead != pPage)
1275	{
1276	STAM_COUNTER_INC(&pPool->StatDirtyPageDupFlush);
1277	Log(("Flush duplicate page idx=%d GCPhys=%RGp type=%s\n", pPageHead->idx, pPageHead->GCPhys, pgmPoolPoolKindToStr(pPageHead->enmKind)));
1278	int rc2 = pgmPoolFlushPage(pPool, pPageHead);
1279	AssertRC(rc2);
1280	}
1281
1282	if (idxNext == NIL_PGMPOOL_IDX)
1283	break;
1284
1285	pPageHead = &pPool->aPages[idxNext];
1286	}
1287	}
1288
1289	/* The flushing above might fail for locked pages, so double check. */
1290	if ( pPage->iMonitoredNext == NIL_PGMPOOL_IDX
1291	&& pPage->iMonitoredPrev == NIL_PGMPOOL_IDX)
1292	{
1293	pgmPoolAddDirtyPage(pVM, pPool, pPage);
1294
1295	/* Temporarily allow write access to the page table again. */
1296	rc = PGMHandlerPhysicalPageTempOff(pVM, pPage->GCPhys & PAGE_BASE_GC_MASK, pPage->GCPhys & PAGE_BASE_GC_MASK);
1297	if (rc == VINF_SUCCESS)
1298	{
1299	rc = PGMShwMakePageWritable(pVCpu, pvFault, PGM_MK_PG_IS_WRITE_FAULT);
1300	AssertMsg(rc == VINF_SUCCESS
1301	/* In the SMP case the page table might be removed while we wait for the PGM lock in the trap handler. */
1302	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
1303	\|\| rc == VERR_PAGE_NOT_PRESENT,
1304	("PGMShwModifyPage -> GCPtr=%RGv rc=%d\n", pvFault, rc));
1305	# ifdef VBOX_STRICT
1306	pPage->GCPtrDirtyFault = pvFault;
1307	# endif
1308
1309	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pPool)->CTX_SUFF_Z(StatMonitor), a);
1310	pgmUnlock(pVM);
1311	return rc;
1312	}
1313	}
1314	}
1315	#endif /* PGMPOOL_WITH_OPTIMIZED_DIRTY_PT */
1316
1317	STAM_COUNTER_INC(&pPool->CTX_MID_Z(StatMonitor,FlushModOverflow));
1318	flushPage:
1319	/*
1320	* Not worth it, so flush it.
1321	*
1322	* If we considered it to be reused, don't go back to ring-3
1323	* to emulate failed instructions since we usually cannot
1324	* interpret then. This may be a bit risky, in which case
1325	* the reuse detection must be fixed.
1326	*/
1327	rc = pgmPoolAccessHandlerFlush(pVM, pVCpu, pPool, pPage, pDis, pRegFrame, GCPhysFault, pvFault);
1328	if ( rc == VINF_EM_RAW_EMULATE_INSTR
1329	&& fReused)
1330	{
1331	/* Make sure that the current instruction still has shadow page backing, otherwise we'll end up in a loop. */
1332	if (PGMShwGetPage(pVCpu, pRegFrame->rip, NULL, NULL) == VINF_SUCCESS)
1333	rc = VINF_SUCCESS; /* safe to restart the instruction. */
1334	}
1335	STAM_PROFILE_STOP_EX(&pVM->pgm.s.CTX_SUFF(pPool)->CTX_SUFF_Z(StatMonitor), &pPool->CTX_MID_Z(StatMonitor,FlushPage), a);
1336	pgmUnlock(pVM);
1337	return rc;
1338	}
1339
1340	# endif /* !IN_RING3 */
1341
1342	# ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
1343
1344	# if defined(VBOX_STRICT) && !defined(IN_RING3)
1345
1346	/**
1347	* Check references to guest physical memory in a PAE / PAE page table.
1348	*
1349	* @param pPool The pool.
1350	* @param pPage The page.
1351	* @param pShwPT The shadow page table (mapping of the page).
1352	* @param pGstPT The guest page table.
1353	*/
1354	static void pgmPoolTrackCheckPTPaePae(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMSHWPTPAE pShwPT, PCX86PTPAE pGstPT)
1355	{
1356	unsigned cErrors = 0;
1357	int LastRc = -1; /* initialized to shut up gcc */
1358	unsigned LastPTE = ~0U; /* initialized to shut up gcc */
1359	RTHCPHYS LastHCPhys = NIL_RTHCPHYS; /* initialized to shut up gcc */
1360	PVM pVM = pPool->CTX_SUFF(pVM);
1361
1362	#ifdef VBOX_STRICT
1363	for (unsigned i = 0; i < RT_MIN(RT_ELEMENTS(pShwPT->a), pPage->iFirstPresent); i++)
1364	AssertMsg(!PGMSHWPTEPAE_IS_P(pShwPT->a[i]), ("Unexpected PTE: idx=%d %RX64 (first=%d)\n", i, PGMSHWPTEPAE_GET_LOG(pShwPT->a[i]), pPage->iFirstPresent));
1365	#endif
1366	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++)
1367	{
1368	if (PGMSHWPTEPAE_IS_P(pShwPT->a[i]))
1369	{
1370	RTHCPHYS HCPhys = NIL_RTHCPHYS;
1371	int rc = PGMPhysGCPhys2HCPhys(pVM, pGstPT->a[i].u & X86_PTE_PAE_PG_MASK, &HCPhys);
1372	if ( rc != VINF_SUCCESS
1373	\|\| PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]) != HCPhys)
1374	{
1375	Log(("rc=%d idx=%d guest %RX64 shw=%RX64 vs %RHp\n", rc, i, pGstPT->a[i].u, PGMSHWPTEPAE_GET_LOG(pShwPT->a[i]), HCPhys));
1376	LastPTE = i;
1377	LastRc = rc;
1378	LastHCPhys = HCPhys;
1379	cErrors++;
1380
1381	RTHCPHYS HCPhysPT = NIL_RTHCPHYS;
1382	rc = PGMPhysGCPhys2HCPhys(pVM, pPage->GCPhys, &HCPhysPT);
1383	AssertRC(rc);
1384
1385	for (unsigned iPage = 0; iPage < pPool->cCurPages; iPage++)
1386	{
1387	PPGMPOOLPAGE pTempPage = &pPool->aPages[iPage];
1388
1389	if (pTempPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_PAE_PT)
1390	{
1391	PPGMSHWPTPAE pShwPT2 = (PPGMSHWPTPAE)PGMPOOL_PAGE_2_PTR(pVM, pTempPage);
1392
1393	for (unsigned j = 0; j < RT_ELEMENTS(pShwPT->a); j++)
1394	{
1395	if ( PGMSHWPTEPAE_IS_P_RW(pShwPT2->a[j])
1396	&& PGMSHWPTEPAE_GET_HCPHYS(pShwPT2->a[j]) == HCPhysPT)
1397	{
1398	Log(("GCPhys=%RGp idx=%d %RX64 vs %RX64\n", pTempPage->GCPhys, j, PGMSHWPTEPAE_GET_LOG(pShwPT->a[j]), PGMSHWPTEPAE_GET_LOG(pShwPT2->a[j])));
1399	}
1400	}
1401
1402	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pShwPT2);
1403	}
1404	}
1405	}
1406	}
1407	}
1408	AssertMsg(!cErrors, ("cErrors=%d: last rc=%d idx=%d guest %RX64 shw=%RX64 vs %RHp\n", cErrors, LastRc, LastPTE, pGstPT->a[LastPTE].u, PGMSHWPTEPAE_GET_LOG(pShwPT->a[LastPTE]), LastHCPhys));
1409	}
1410
1411
1412	/**
1413	* Check references to guest physical memory in a PAE / 32-bit page table.
1414	*
1415	* @param pPool The pool.
1416	* @param pPage The page.
1417	* @param pShwPT The shadow page table (mapping of the page).
1418	* @param pGstPT The guest page table.
1419	*/
1420	static void pgmPoolTrackCheckPTPae32Bit(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMSHWPTPAE pShwPT, PCX86PT pGstPT)
1421	{
1422	unsigned cErrors = 0;
1423	int LastRc = -1; /* initialized to shut up gcc */
1424	unsigned LastPTE = ~0U; /* initialized to shut up gcc */
1425	RTHCPHYS LastHCPhys = NIL_RTHCPHYS; /* initialized to shut up gcc */
1426	PVM pVM = pPool->CTX_SUFF(pVM);
1427
1428	#ifdef VBOX_STRICT
1429	for (unsigned i = 0; i < RT_MIN(RT_ELEMENTS(pShwPT->a), pPage->iFirstPresent); i++)
1430	AssertMsg(!PGMSHWPTEPAE_IS_P(pShwPT->a[i]), ("Unexpected PTE: idx=%d %RX64 (first=%d)\n", i, PGMSHWPTEPAE_GET_LOG(pShwPT->a[i]), pPage->iFirstPresent));
1431	#endif
1432	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++)
1433	{
1434	if (PGMSHWPTEPAE_IS_P(pShwPT->a[i]))
1435	{
1436	RTHCPHYS HCPhys = NIL_RTHCPHYS;
1437	int rc = PGMPhysGCPhys2HCPhys(pVM, pGstPT->a[i].u & X86_PTE_PG_MASK, &HCPhys);
1438	if ( rc != VINF_SUCCESS
1439	\|\| PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]) != HCPhys)
1440	{
1441	Log(("rc=%d idx=%d guest %x shw=%RX64 vs %RHp\n", rc, i, pGstPT->a[i].u, PGMSHWPTEPAE_GET_LOG(pShwPT->a[i]), HCPhys));
1442	LastPTE = i;
1443	LastRc = rc;
1444	LastHCPhys = HCPhys;
1445	cErrors++;
1446
1447	RTHCPHYS HCPhysPT = NIL_RTHCPHYS;
1448	rc = PGMPhysGCPhys2HCPhys(pVM, pPage->GCPhys, &HCPhysPT);
1449	AssertRC(rc);
1450
1451	for (unsigned iPage = 0; iPage < pPool->cCurPages; iPage++)
1452	{
1453	PPGMPOOLPAGE pTempPage = &pPool->aPages[iPage];
1454
1455	if (pTempPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_32BIT_PT)
1456	{
1457	PPGMSHWPTPAE pShwPT2 = (PPGMSHWPTPAE)PGMPOOL_PAGE_2_PTR(pVM, pTempPage);
1458
1459	for (unsigned j = 0; j < RT_ELEMENTS(pShwPT->a); j++)
1460	{
1461	if ( PGMSHWPTEPAE_IS_P_RW(pShwPT2->a[j])
1462	&& PGMSHWPTEPAE_GET_HCPHYS(pShwPT2->a[j]) == HCPhysPT)
1463	{
1464	Log(("GCPhys=%RGp idx=%d %RX64 vs %RX64\n", pTempPage->GCPhys, j, PGMSHWPTEPAE_GET_LOG(pShwPT->a[j]), PGMSHWPTEPAE_GET_LOG(pShwPT2->a[j])));
1465	}
1466	}
1467
1468	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pShwPT2);
1469	}
1470	}
1471	}
1472	}
1473	}
1474	AssertMsg(!cErrors, ("cErrors=%d: last rc=%d idx=%d guest %x shw=%RX64 vs %RHp\n", cErrors, LastRc, LastPTE, pGstPT->a[LastPTE].u, PGMSHWPTEPAE_GET_LOG(pShwPT->a[LastPTE]), LastHCPhys));
1475	}
1476
1477	# endif /* VBOX_STRICT && !IN_RING3 */
1478
1479	/**
1480	* Clear references to guest physical memory in a PAE / PAE page table.
1481	*
1482	* @returns nr of changed PTEs
1483	* @param pPool The pool.
1484	* @param pPage The page.
1485	* @param pShwPT The shadow page table (mapping of the page).
1486	* @param pGstPT The guest page table.
1487	* @param pOldGstPT The old cached guest page table.
1488	* @param fAllowRemoval Bail out as soon as we encounter an invalid PTE
1489	* @param pfFlush Flush reused page table (out)
1490	*/
1491	DECLINLINE(unsigned) pgmPoolTrackFlushPTPaePae(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMSHWPTPAE pShwPT, PCX86PTPAE pGstPT,
1492	PCX86PTPAE pOldGstPT, bool fAllowRemoval, bool *pfFlush)
1493	{
1494	unsigned cChanged = 0;
1495
1496	#ifdef VBOX_STRICT
1497	for (unsigned i = 0; i < RT_MIN(RT_ELEMENTS(pShwPT->a), pPage->iFirstPresent); i++)
1498	AssertMsg(!PGMSHWPTEPAE_IS_P(pShwPT->a[i]), ("Unexpected PTE: idx=%d %RX64 (first=%d)\n", i, PGMSHWPTEPAE_GET_LOG(pShwPT->a[i]), pPage->iFirstPresent));
1499	#endif
1500	*pfFlush = false;
1501
1502	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++)
1503	{
1504	/* Check the new value written by the guest. If present and with a bogus physical address, then
1505	* it's fairly safe to assume the guest is reusing the PT.
1506	*/
1507	if ( fAllowRemoval
1508	&& pGstPT->a[i].n.u1Present)
1509	{
1510	if (!PGMPhysIsGCPhysValid(pPool->CTX_SUFF(pVM), pGstPT->a[i].u & X86_PTE_PAE_PG_MASK))
1511	{
1512	*pfFlush = true;
1513	return ++cChanged;
1514	}
1515	}
1516	if (PGMSHWPTEPAE_IS_P(pShwPT->a[i]))
1517	{
1518	/* If the old cached PTE is identical, then there's no need to flush the shadow copy. */
1519	if ((pGstPT->a[i].u & X86_PTE_PAE_PG_MASK) == (pOldGstPT->a[i].u & X86_PTE_PAE_PG_MASK))
1520	{
1521	#ifdef VBOX_STRICT
1522	RTHCPHYS HCPhys = NIL_RTGCPHYS;
1523	int rc = PGMPhysGCPhys2HCPhys(pPool->CTX_SUFF(pVM), pGstPT->a[i].u & X86_PTE_PAE_PG_MASK, &HCPhys);
1524	AssertMsg(rc == VINF_SUCCESS && PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]) == HCPhys, ("rc=%d guest %RX64 old %RX64 shw=%RX64 vs %RHp\n", rc, pGstPT->a[i].u, pOldGstPT->a[i].u, PGMSHWPTEPAE_GET_LOG(pShwPT->a[i]), HCPhys));
1525	#endif
1526	uint64_t uHostAttr = PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & (X86_PTE_P \| X86_PTE_US \| X86_PTE_A \| X86_PTE_D \| X86_PTE_G \| X86_PTE_PAE_NX);
1527	bool fHostRW = !!(PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & X86_PTE_RW);
1528	uint64_t uGuestAttr = pGstPT->a[i].u & (X86_PTE_P \| X86_PTE_US \| X86_PTE_A \| X86_PTE_D \| X86_PTE_G \| X86_PTE_PAE_NX);
1529	bool fGuestRW = !!(pGstPT->a[i].u & X86_PTE_RW);
1530
1531	if ( uHostAttr == uGuestAttr
1532	&& fHostRW <= fGuestRW)
1533	continue;
1534	}
1535	cChanged++;
1536	/* Something was changed, so flush it. */
1537	Log4(("pgmPoolTrackDerefPTPaePae: i=%d pte=%RX64 hint=%RX64\n",
1538	i, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), pOldGstPT->a[i].u & X86_PTE_PAE_PG_MASK));
1539	pgmPoolTracDerefGCPhysHint(pPool, pPage, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), pOldGstPT->a[i].u & X86_PTE_PAE_PG_MASK, i);
1540	PGMSHWPTEPAE_ATOMIC_SET(pShwPT->a[i], 0);
1541	}
1542	}
1543	return cChanged;
1544	}
1545
1546
1547	/**
1548	* Clear references to guest physical memory in a PAE / PAE page table.
1549	*
1550	* @returns nr of changed PTEs
1551	* @param pPool The pool.
1552	* @param pPage The page.
1553	* @param pShwPT The shadow page table (mapping of the page).
1554	* @param pGstPT The guest page table.
1555	* @param pOldGstPT The old cached guest page table.
1556	* @param fAllowRemoval Bail out as soon as we encounter an invalid PTE
1557	* @param pfFlush Flush reused page table (out)
1558	*/
1559	DECLINLINE(unsigned) pgmPoolTrackFlushPTPae32Bit(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMSHWPTPAE pShwPT, PCX86PT pGstPT,
1560	PCX86PT pOldGstPT, bool fAllowRemoval, bool *pfFlush)
1561	{
1562	unsigned cChanged = 0;
1563
1564	#ifdef VBOX_STRICT
1565	for (unsigned i = 0; i < RT_MIN(RT_ELEMENTS(pShwPT->a), pPage->iFirstPresent); i++)
1566	AssertMsg(!PGMSHWPTEPAE_IS_P(pShwPT->a[i]), ("Unexpected PTE: idx=%d %RX64 (first=%d)\n", i, PGMSHWPTEPAE_GET_LOG(pShwPT->a[i]), pPage->iFirstPresent));
1567	#endif
1568	*pfFlush = false;
1569
1570	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++)
1571	{
1572	/* Check the new value written by the guest. If present and with a bogus physical address, then
1573	* it's fairly safe to assume the guest is reusing the PT.
1574	*/
1575	if ( fAllowRemoval
1576	&& pGstPT->a[i].n.u1Present)
1577	{
1578	if (!PGMPhysIsGCPhysValid(pPool->CTX_SUFF(pVM), pGstPT->a[i].u & X86_PTE_PG_MASK))
1579	{
1580	*pfFlush = true;
1581	return ++cChanged;
1582	}
1583	}
1584	if (PGMSHWPTEPAE_IS_P(pShwPT->a[i]))
1585	{
1586	/* If the old cached PTE is identical, then there's no need to flush the shadow copy. */
1587	if ((pGstPT->a[i].u & X86_PTE_PG_MASK) == (pOldGstPT->a[i].u & X86_PTE_PG_MASK))
1588	{
1589	#ifdef VBOX_STRICT
1590	RTHCPHYS HCPhys = NIL_RTGCPHYS;
1591	int rc = PGMPhysGCPhys2HCPhys(pPool->CTX_SUFF(pVM), pGstPT->a[i].u & X86_PTE_PG_MASK, &HCPhys);
1592	AssertMsg(rc == VINF_SUCCESS && PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]) == HCPhys, ("rc=%d guest %x old %x shw=%RX64 vs %RHp\n", rc, pGstPT->a[i].u, pOldGstPT->a[i].u, PGMSHWPTEPAE_GET_LOG(pShwPT->a[i]), HCPhys));
1593	#endif
1594	uint64_t uHostAttr = PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & (X86_PTE_P \| X86_PTE_US \| X86_PTE_A \| X86_PTE_D \| X86_PTE_G);
1595	bool fHostRW = !!(PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & X86_PTE_RW);
1596	uint64_t uGuestAttr = pGstPT->a[i].u & (X86_PTE_P \| X86_PTE_US \| X86_PTE_A \| X86_PTE_D \| X86_PTE_G);
1597	bool fGuestRW = !!(pGstPT->a[i].u & X86_PTE_RW);
1598
1599	if ( uHostAttr == uGuestAttr
1600	&& fHostRW <= fGuestRW)
1601	continue;
1602	}
1603	cChanged++;
1604	/* Something was changed, so flush it. */
1605	Log4(("pgmPoolTrackDerefPTPaePae: i=%d pte=%RX64 hint=%x\n",
1606	i, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), pOldGstPT->a[i].u & X86_PTE_PG_MASK));
1607	pgmPoolTracDerefGCPhysHint(pPool, pPage, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), pOldGstPT->a[i].u & X86_PTE_PG_MASK, i);
1608	PGMSHWPTEPAE_ATOMIC_SET(pShwPT->a[i], 0);
1609	}
1610	}
1611	return cChanged;
1612	}
1613
1614
1615	/**
1616	* Flush a dirty page
1617	*
1618	* @param pVM Pointer to the VM.
1619	* @param pPool The pool.
1620	* @param idxSlot Dirty array slot index
1621	* @param fAllowRemoval Allow a reused page table to be removed
1622	*/
1623	static void pgmPoolFlushDirtyPage(PVM pVM, PPGMPOOL pPool, unsigned idxSlot, bool fAllowRemoval = false)
1624	{
1625	PPGMPOOLPAGE pPage;
1626	unsigned idxPage;
1627
1628	Assert(idxSlot < RT_ELEMENTS(pPool->aDirtyPages));
1629	if (pPool->aDirtyPages[idxSlot].uIdx == NIL_PGMPOOL_IDX)
1630	return;
1631
1632	idxPage = pPool->aDirtyPages[idxSlot].uIdx;
1633	AssertRelease(idxPage != NIL_PGMPOOL_IDX);
1634	pPage = &pPool->aPages[idxPage];
1635	Assert(pPage->idx == idxPage);
1636	Assert(pPage->iMonitoredNext == NIL_PGMPOOL_IDX && pPage->iMonitoredPrev == NIL_PGMPOOL_IDX);
1637
1638	AssertMsg(pPage->fDirty, ("Page %RGp (slot=%d) not marked dirty!", pPage->GCPhys, idxSlot));
1639	Log(("Flush dirty page %RGp cMods=%d\n", pPage->GCPhys, pPage->cModifications));
1640
1641	#if defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) \|\| defined(IN_RC)
1642	PVMCPU pVCpu = VMMGetCpu(pVM);
1643	uint32_t iPrevSubset = PGMRZDynMapPushAutoSubset(pVCpu);
1644	#endif
1645
1646	/* First write protect the page again to catch all write accesses. (before checking for changes -> SMP) */
1647	int rc = PGMHandlerPhysicalReset(pVM, pPage->GCPhys & PAGE_BASE_GC_MASK);
1648	Assert(rc == VINF_SUCCESS);
1649	pPage->fDirty = false;
1650
1651	#ifdef VBOX_STRICT
1652	uint64_t fFlags = 0;
1653	RTHCPHYS HCPhys;
1654	rc = PGMShwGetPage(VMMGetCpu(pVM), pPage->GCPtrDirtyFault, &fFlags, &HCPhys);
1655	AssertMsg( ( rc == VINF_SUCCESS
1656	&& (!(fFlags & X86_PTE_RW) \|\| HCPhys != pPage->Core.Key))
1657	/* In the SMP case the page table might be removed while we wait for the PGM lock in the trap handler. */
1658	\|\| rc == VERR_PAGE_TABLE_NOT_PRESENT
1659	\|\| rc == VERR_PAGE_NOT_PRESENT,
1660	("PGMShwGetPage -> GCPtr=%RGv rc=%d flags=%RX64\n", pPage->GCPtrDirtyFault, rc, fFlags));
1661	#endif
1662
1663	/* Flush those PTEs that have changed. */
1664	STAM_PROFILE_START(&pPool->StatTrackDeref,a);
1665	void *pvShw = PGMPOOL_PAGE_2_PTR(pVM, pPage);
1666	void *pvGst;
1667	rc = PGM_GCPHYS_2_PTR_EX(pVM, pPage->GCPhys, &pvGst); AssertReleaseRC(rc);
1668	bool fFlush;
1669	unsigned cChanges;
1670
1671	if (pPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_PAE_PT)
1672	cChanges = pgmPoolTrackFlushPTPaePae(pPool, pPage, (PPGMSHWPTPAE)pvShw, (PCX86PTPAE)pvGst,
1673	(PCX86PTPAE)&pPool->aDirtyPages[idxSlot].aPage[0], fAllowRemoval, &fFlush);
1674	else
1675	cChanges = pgmPoolTrackFlushPTPae32Bit(pPool, pPage, (PPGMSHWPTPAE)pvShw, (PCX86PT)pvGst,
1676	(PCX86PT)&pPool->aDirtyPages[idxSlot].aPage[0], fAllowRemoval, &fFlush);
1677
1678	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvGst);
1679	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvShw);
1680	STAM_PROFILE_STOP(&pPool->StatTrackDeref,a);
1681	/* Note: we might want to consider keeping the dirty page active in case there were many changes. */
1682
1683	/* This page is likely to be modified again, so reduce the nr of modifications just a bit here. */
1684	Assert(pPage->cModifications);
1685	if (cChanges < 4)
1686	pPage->cModifications = 1; /* must use > 0 here */
1687	else
1688	pPage->cModifications = RT_MAX(1, pPage->cModifications / 2);
1689
1690	STAM_COUNTER_INC(&pPool->StatResetDirtyPages);
1691	if (pPool->cDirtyPages == RT_ELEMENTS(pPool->aDirtyPages))
1692	pPool->idxFreeDirtyPage = idxSlot;
1693
1694	pPool->cDirtyPages--;
1695	pPool->aDirtyPages[idxSlot].uIdx = NIL_PGMPOOL_IDX;
1696	Assert(pPool->cDirtyPages <= RT_ELEMENTS(pPool->aDirtyPages));
1697	if (fFlush)
1698	{
1699	Assert(fAllowRemoval);
1700	Log(("Flush reused page table!\n"));
1701	pgmPoolFlushPage(pPool, pPage);
1702	STAM_COUNTER_INC(&pPool->StatForceFlushReused);
1703	}
1704	else
1705	Log(("Removed dirty page %RGp cMods=%d cChanges=%d\n", pPage->GCPhys, pPage->cModifications, cChanges));
1706
1707	#if defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) \|\| defined(IN_RC)
1708	PGMRZDynMapPopAutoSubset(pVCpu, iPrevSubset);
1709	#endif
1710	}
1711
1712
1713	# ifndef IN_RING3
1714	/**
1715	* Add a new dirty page
1716	*
1717	* @param pVM Pointer to the VM.
1718	* @param pPool The pool.
1719	* @param pPage The page.
1720	*/
1721	void pgmPoolAddDirtyPage(PVM pVM, PPGMPOOL pPool, PPGMPOOLPAGE pPage)
1722	{
1723	unsigned idxFree;
1724
1725	PGM_LOCK_ASSERT_OWNER(pVM);
1726	AssertCompile(RT_ELEMENTS(pPool->aDirtyPages) == 8 \|\| RT_ELEMENTS(pPool->aDirtyPages) == 16);
1727	Assert(!pPage->fDirty);
1728
1729	idxFree = pPool->idxFreeDirtyPage;
1730	Assert(idxFree < RT_ELEMENTS(pPool->aDirtyPages));
1731	Assert(pPage->iMonitoredNext == NIL_PGMPOOL_IDX && pPage->iMonitoredPrev == NIL_PGMPOOL_IDX);
1732
1733	if (pPool->cDirtyPages >= RT_ELEMENTS(pPool->aDirtyPages))
1734	{
1735	STAM_COUNTER_INC(&pPool->StatDirtyPageOverFlowFlush);
1736	pgmPoolFlushDirtyPage(pVM, pPool, idxFree, true /* allow removal of reused page tables*/);
1737	}
1738	Assert(pPool->cDirtyPages < RT_ELEMENTS(pPool->aDirtyPages));
1739	AssertMsg(pPool->aDirtyPages[idxFree].uIdx == NIL_PGMPOOL_IDX, ("idxFree=%d cDirtyPages=%d\n", idxFree, pPool->cDirtyPages));
1740
1741	Log(("Add dirty page %RGp (slot=%d)\n", pPage->GCPhys, idxFree));
1742
1743	/*
1744	* Make a copy of the guest page table as we require valid GCPhys addresses
1745	* when removing references to physical pages.
1746	* (The HCPhys linear lookup is extremely expensive!)
1747	*/
1748	void *pvGst;
1749	int rc = PGM_GCPHYS_2_PTR_EX(pVM, pPage->GCPhys, &pvGst); AssertReleaseRC(rc);
1750	memcpy(&pPool->aDirtyPages[idxFree].aPage[0], pvGst, (pPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_PAE_PT) ? PAGE_SIZE : PAGE_SIZE/2);
1751	# ifdef VBOX_STRICT
1752	void *pvShw = PGMPOOL_PAGE_2_PTR(pVM, pPage);
1753	if (pPage->enmKind == PGMPOOLKIND_PAE_PT_FOR_PAE_PT)
1754	pgmPoolTrackCheckPTPaePae(pPool, pPage, (PPGMSHWPTPAE)pvShw, (PCX86PTPAE)pvGst);
1755	else
1756	pgmPoolTrackCheckPTPae32Bit(pPool, pPage, (PPGMSHWPTPAE)pvShw, (PCX86PT)pvGst);
1757	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvShw);
1758	# endif
1759	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvGst);
1760
1761	STAM_COUNTER_INC(&pPool->StatDirtyPage);
1762	pPage->fDirty = true;
1763	pPage->idxDirtyEntry = (uint8_t)idxFree; Assert(pPage->idxDirtyEntry == idxFree);
1764	pPool->aDirtyPages[idxFree].uIdx = pPage->idx;
1765	pPool->cDirtyPages++;
1766
1767	pPool->idxFreeDirtyPage = (pPool->idxFreeDirtyPage + 1) & (RT_ELEMENTS(pPool->aDirtyPages) - 1);
1768	if ( pPool->cDirtyPages < RT_ELEMENTS(pPool->aDirtyPages)
1769	&& pPool->aDirtyPages[pPool->idxFreeDirtyPage].uIdx != NIL_PGMPOOL_IDX)
1770	{
1771	unsigned i;
1772	for (i = 1; i < RT_ELEMENTS(pPool->aDirtyPages); i++)
1773	{
1774	idxFree = (pPool->idxFreeDirtyPage + i) & (RT_ELEMENTS(pPool->aDirtyPages) - 1);
1775	if (pPool->aDirtyPages[idxFree].uIdx == NIL_PGMPOOL_IDX)
1776	{
1777	pPool->idxFreeDirtyPage = idxFree;
1778	break;
1779	}
1780	}
1781	Assert(i != RT_ELEMENTS(pPool->aDirtyPages));
1782	}
1783
1784	Assert(pPool->cDirtyPages == RT_ELEMENTS(pPool->aDirtyPages) \|\| pPool->aDirtyPages[pPool->idxFreeDirtyPage].uIdx == NIL_PGMPOOL_IDX);
1785	return;
1786	}
1787	# endif /* !IN_RING3 */
1788
1789
1790	/**
1791	* Check if the specified page is dirty (not write monitored)
1792	*
1793	* @return dirty or not
1794	* @param pVM Pointer to the VM.
1795	* @param GCPhys Guest physical address
1796	*/
1797	bool pgmPoolIsDirtyPage(PVM pVM, RTGCPHYS GCPhys)
1798	{
1799	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1800	PGM_LOCK_ASSERT_OWNER(pVM);
1801	if (!pPool->cDirtyPages)
1802	return false;
1803
1804	GCPhys = GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK;
1805
1806	for (unsigned i = 0; i < RT_ELEMENTS(pPool->aDirtyPages); i++)
1807	{
1808	if (pPool->aDirtyPages[i].uIdx != NIL_PGMPOOL_IDX)
1809	{
1810	PPGMPOOLPAGE pPage;
1811	unsigned idxPage = pPool->aDirtyPages[i].uIdx;
1812
1813	pPage = &pPool->aPages[idxPage];
1814	if (pPage->GCPhys == GCPhys)
1815	return true;
1816	}
1817	}
1818	return false;
1819	}
1820
1821
1822	/**
1823	* Reset all dirty pages by reinstating page monitoring.
1824	*
1825	* @param pVM Pointer to the VM.
1826	*/
1827	void pgmPoolResetDirtyPages(PVM pVM)
1828	{
1829	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1830	PGM_LOCK_ASSERT_OWNER(pVM);
1831	Assert(pPool->cDirtyPages <= RT_ELEMENTS(pPool->aDirtyPages));
1832
1833	if (!pPool->cDirtyPages)
1834	return;
1835
1836	Log(("pgmPoolResetDirtyPages\n"));
1837	for (unsigned i = 0; i < RT_ELEMENTS(pPool->aDirtyPages); i++)
1838	pgmPoolFlushDirtyPage(pVM, pPool, i, true /* allow removal of reused page tables*/);
1839
1840	pPool->idxFreeDirtyPage = 0;
1841	if ( pPool->cDirtyPages != RT_ELEMENTS(pPool->aDirtyPages)
1842	&& pPool->aDirtyPages[pPool->idxFreeDirtyPage].uIdx != NIL_PGMPOOL_IDX)
1843	{
1844	unsigned i;
1845	for (i = 1; i < RT_ELEMENTS(pPool->aDirtyPages); i++)
1846	{
1847	if (pPool->aDirtyPages[i].uIdx == NIL_PGMPOOL_IDX)
1848	{
1849	pPool->idxFreeDirtyPage = i;
1850	break;
1851	}
1852	}
1853	AssertMsg(i != RT_ELEMENTS(pPool->aDirtyPages), ("cDirtyPages %d", pPool->cDirtyPages));
1854	}
1855
1856	Assert(pPool->aDirtyPages[pPool->idxFreeDirtyPage].uIdx == NIL_PGMPOOL_IDX \|\| pPool->cDirtyPages == RT_ELEMENTS(pPool->aDirtyPages));
1857	return;
1858	}
1859
1860
1861	/**
1862	* Invalidate the PT entry for the specified page
1863	*
1864	* @param pVM Pointer to the VM.
1865	* @param GCPtrPage Guest page to invalidate
1866	*/
1867	void pgmPoolResetDirtyPage(PVM pVM, RTGCPTR GCPtrPage)
1868	{
1869	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1870	PGM_LOCK_ASSERT_OWNER(pVM);
1871	Assert(pPool->cDirtyPages <= RT_ELEMENTS(pPool->aDirtyPages));
1872
1873	if (!pPool->cDirtyPages)
1874	return;
1875
1876	Log(("pgmPoolResetDirtyPage %RGv\n", GCPtrPage));
1877	for (unsigned i = 0; i < RT_ELEMENTS(pPool->aDirtyPages); i++)
1878	{
1879	}
1880	}
1881
1882
1883	/**
1884	* Reset all dirty pages by reinstating page monitoring.
1885	*
1886	* @param pVM Pointer to the VM.
1887	* @param GCPhysPT Physical address of the page table
1888	*/
1889	void pgmPoolInvalidateDirtyPage(PVM pVM, RTGCPHYS GCPhysPT)
1890	{
1891	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
1892	PGM_LOCK_ASSERT_OWNER(pVM);
1893	Assert(pPool->cDirtyPages <= RT_ELEMENTS(pPool->aDirtyPages));
1894	unsigned idxDirtyPage = RT_ELEMENTS(pPool->aDirtyPages);
1895
1896	if (!pPool->cDirtyPages)
1897	return;
1898
1899	GCPhysPT = GCPhysPT & ~(RTGCPHYS)PAGE_OFFSET_MASK;
1900
1901	for (unsigned i = 0; i < RT_ELEMENTS(pPool->aDirtyPages); i++)
1902	{
1903	if (pPool->aDirtyPages[i].uIdx != NIL_PGMPOOL_IDX)
1904	{
1905	unsigned idxPage = pPool->aDirtyPages[i].uIdx;
1906
1907	PPGMPOOLPAGE pPage = &pPool->aPages[idxPage];
1908	if (pPage->GCPhys == GCPhysPT)
1909	{
1910	idxDirtyPage = i;
1911	break;
1912	}
1913	}
1914	}
1915
1916	if (idxDirtyPage != RT_ELEMENTS(pPool->aDirtyPages))
1917	{
1918	pgmPoolFlushDirtyPage(pVM, pPool, idxDirtyPage, true /* allow removal of reused page tables*/);
1919	if ( pPool->cDirtyPages != RT_ELEMENTS(pPool->aDirtyPages)
1920	&& pPool->aDirtyPages[pPool->idxFreeDirtyPage].uIdx != NIL_PGMPOOL_IDX)
1921	{
1922	unsigned i;
1923	for (i = 0; i < RT_ELEMENTS(pPool->aDirtyPages); i++)
1924	{
1925	if (pPool->aDirtyPages[i].uIdx == NIL_PGMPOOL_IDX)
1926	{
1927	pPool->idxFreeDirtyPage = i;
1928	break;
1929	}
1930	}
1931	AssertMsg(i != RT_ELEMENTS(pPool->aDirtyPages), ("cDirtyPages %d", pPool->cDirtyPages));
1932	}
1933	}
1934	}
1935
1936	# endif /* PGMPOOL_WITH_OPTIMIZED_DIRTY_PT */
1937
1938	/**
1939	* Inserts a page into the GCPhys hash table.
1940	*
1941	* @param pPool The pool.
1942	* @param pPage The page.
1943	*/
1944	DECLINLINE(void) pgmPoolHashInsert(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
1945	{
1946	Log3(("pgmPoolHashInsert: %RGp\n", pPage->GCPhys));
1947	Assert(pPage->GCPhys != NIL_RTGCPHYS); Assert(pPage->iNext == NIL_PGMPOOL_IDX);
1948	uint16_t iHash = PGMPOOL_HASH(pPage->GCPhys);
1949	pPage->iNext = pPool->aiHash[iHash];
1950	pPool->aiHash[iHash] = pPage->idx;
1951	}
1952
1953
1954	/**
1955	* Removes a page from the GCPhys hash table.
1956	*
1957	* @param pPool The pool.
1958	* @param pPage The page.
1959	*/
1960	DECLINLINE(void) pgmPoolHashRemove(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
1961	{
1962	Log3(("pgmPoolHashRemove: %RGp\n", pPage->GCPhys));
1963	uint16_t iHash = PGMPOOL_HASH(pPage->GCPhys);
1964	if (pPool->aiHash[iHash] == pPage->idx)
1965	pPool->aiHash[iHash] = pPage->iNext;
1966	else
1967	{
1968	uint16_t iPrev = pPool->aiHash[iHash];
1969	for (;;)
1970	{
1971	const int16_t i = pPool->aPages[iPrev].iNext;
1972	if (i == pPage->idx)
1973	{
1974	pPool->aPages[iPrev].iNext = pPage->iNext;
1975	break;
1976	}
1977	if (i == NIL_PGMPOOL_IDX)
1978	{
1979	AssertReleaseMsgFailed(("GCPhys=%RGp idx=%d\n", pPage->GCPhys, pPage->idx));
1980	break;
1981	}
1982	iPrev = i;
1983	}
1984	}
1985	pPage->iNext = NIL_PGMPOOL_IDX;
1986	}
1987
1988
1989	/**
1990	* Frees up one cache page.
1991	*
1992	* @returns VBox status code.
1993	* @retval VINF_SUCCESS on success.
1994	* @param pPool The pool.
1995	* @param iUser The user index.
1996	*/
1997	static int pgmPoolCacheFreeOne(PPGMPOOL pPool, uint16_t iUser)
1998	{
1999	#ifndef IN_RC
2000	const PVM pVM = pPool->CTX_SUFF(pVM);
2001	#endif
2002	Assert(pPool->iAgeHead != pPool->iAgeTail); /* We shouldn't be here if there < 2 cached entries! */
2003	STAM_COUNTER_INC(&pPool->StatCacheFreeUpOne);
2004
2005	/*
2006	* Select one page from the tail of the age list.
2007	*/
2008	PPGMPOOLPAGE pPage;
2009	for (unsigned iLoop = 0; ; iLoop++)
2010	{
2011	uint16_t iToFree = pPool->iAgeTail;
2012	if (iToFree == iUser)
2013	iToFree = pPool->aPages[iToFree].iAgePrev;
2014	/* This is the alternative to the SyncCR3 pgmPoolCacheUsed calls.
2015	if (pPool->aPages[iToFree].iUserHead != NIL_PGMPOOL_USER_INDEX)
2016	{
2017	uint16_t i = pPool->aPages[iToFree].iAgePrev;
2018	for (unsigned j = 0; j < 10 && i != NIL_PGMPOOL_USER_INDEX; j++, i = pPool->aPages[i].iAgePrev)
2019	{
2020	if (pPool->aPages[iToFree].iUserHead == NIL_PGMPOOL_USER_INDEX)
2021	continue;
2022	iToFree = i;
2023	break;
2024	}
2025	}
2026	*/
2027	Assert(iToFree != iUser);
2028	AssertRelease(iToFree != NIL_PGMPOOL_IDX);
2029	pPage = &pPool->aPages[iToFree];
2030
2031	/*
2032	* Reject any attempts at flushing the currently active shadow CR3 mapping.
2033	* Call pgmPoolCacheUsed to move the page to the head of the age list.
2034	*/
2035	if (!pgmPoolIsPageLocked(pPage))
2036	break;
2037	LogFlow(("pgmPoolCacheFreeOne: refuse CR3 mapping\n"));
2038	pgmPoolCacheUsed(pPool, pPage);
2039	AssertLogRelReturn(iLoop < 8192, VERR_PGM_POOL_TOO_MANY_LOOPS);
2040	}
2041
2042	/*
2043	* Found a usable page, flush it and return.
2044	*/
2045	int rc = pgmPoolFlushPage(pPool, pPage);
2046	/* This flush was initiated by us and not the guest, so explicitly flush the TLB. */
2047	/* todo: find out why this is necessary; pgmPoolFlushPage should trigger a flush if one is really needed. */
2048	if (rc == VINF_SUCCESS)
2049	PGM_INVL_ALL_VCPU_TLBS(pVM);
2050	return rc;
2051	}
2052
2053
2054	/**
2055	* Checks if a kind mismatch is really a page being reused
2056	* or if it's just normal remappings.
2057	*
2058	* @returns true if reused and the cached page (enmKind1) should be flushed
2059	* @returns false if not reused.
2060	* @param enmKind1 The kind of the cached page.
2061	* @param enmKind2 The kind of the requested page.
2062	*/
2063	static bool pgmPoolCacheReusedByKind(PGMPOOLKIND enmKind1, PGMPOOLKIND enmKind2)
2064	{
2065	switch (enmKind1)
2066	{
2067	/*
2068	* Never reuse them. There is no remapping in non-paging mode.
2069	*/
2070	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
2071	case PGMPOOLKIND_32BIT_PD_PHYS:
2072	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
2073	case PGMPOOLKIND_PAE_PD_PHYS:
2074	case PGMPOOLKIND_PAE_PDPT_PHYS:
2075	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
2076	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
2077	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
2078	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
2079	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
2080	case PGMPOOLKIND_PAE_PDPT_FOR_32BIT: /* never reuse them for other types */
2081	return false;
2082
2083	/*
2084	* It's perfectly fine to reuse these, except for PAE and non-paging stuff.
2085	*/
2086	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
2087	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
2088	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
2089	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
2090	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
2091	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
2092	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
2093	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
2094	case PGMPOOLKIND_32BIT_PD:
2095	case PGMPOOLKIND_PAE_PDPT:
2096	switch (enmKind2)
2097	{
2098	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
2099	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
2100	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
2101	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
2102	case PGMPOOLKIND_64BIT_PML4:
2103	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
2104	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
2105	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
2106	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
2107	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
2108	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
2109	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
2110	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
2111	return true;
2112	default:
2113	return false;
2114	}
2115
2116	/*
2117	* It's perfectly fine to reuse these, except for PAE and non-paging stuff.
2118	*/
2119	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
2120	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
2121	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
2122	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
2123	case PGMPOOLKIND_64BIT_PML4:
2124	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
2125	switch (enmKind2)
2126	{
2127	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
2128	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
2129	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
2130	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
2131	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
2132	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
2133	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
2134	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
2135	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
2136	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
2137	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
2138	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
2139	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
2140	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
2141	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
2142	return true;
2143	default:
2144	return false;
2145	}
2146
2147	/*
2148	* These cannot be flushed, and it's common to reuse the PDs as PTs.
2149	*/
2150	case PGMPOOLKIND_ROOT_NESTED:
2151	return false;
2152
2153	default:
2154	AssertFatalMsgFailed(("enmKind1=%d\n", enmKind1));
2155	}
2156	}
2157
2158
2159	/**
2160	* Attempts to satisfy a pgmPoolAlloc request from the cache.
2161	*
2162	* @returns VBox status code.
2163	* @retval VINF_PGM_CACHED_PAGE on success.
2164	* @retval VERR_FILE_NOT_FOUND if not found.
2165	* @param pPool The pool.
2166	* @param GCPhys The GC physical address of the page we're gonna shadow.
2167	* @param enmKind The kind of mapping.
2168	* @param enmAccess Access type for the mapping (only relevant for big pages)
2169	* @param fA20Enabled Whether the CPU has the A20 gate enabled.
2170	* @param iUser The shadow page pool index of the user table.
2171	* @param iUserTable The index into the user table (shadowed).
2172	* @param ppPage Where to store the pointer to the page.
2173	*/
2174	static int pgmPoolCacheAlloc(PPGMPOOL pPool, RTGCPHYS GCPhys, PGMPOOLKIND enmKind, PGMPOOLACCESS enmAccess, bool fA20Enabled,
2175	uint16_t iUser, uint32_t iUserTable, PPPGMPOOLPAGE ppPage)
2176	{
2177	/*
2178	* Look up the GCPhys in the hash.
2179	*/
2180	unsigned i = pPool->aiHash[PGMPOOL_HASH(GCPhys)];
2181	Log3(("pgmPoolCacheAlloc: %RGp kind %s iUser=%d iUserTable=%x SLOT=%d\n", GCPhys, pgmPoolPoolKindToStr(enmKind), iUser, iUserTable, i));
2182	if (i != NIL_PGMPOOL_IDX)
2183	{
2184	do
2185	{
2186	PPGMPOOLPAGE pPage = &pPool->aPages[i];
2187	Log4(("pgmPoolCacheAlloc: slot %d found page %RGp\n", i, pPage->GCPhys));
2188	if (pPage->GCPhys == GCPhys)
2189	{
2190	if ( (PGMPOOLKIND)pPage->enmKind == enmKind
2191	&& (PGMPOOLACCESS)pPage->enmAccess == enmAccess
2192	&& pPage->fA20Enabled == fA20Enabled)
2193	{
2194	/* Put it at the start of the use list to make sure pgmPoolTrackAddUser
2195	* doesn't flush it in case there are no more free use records.
2196	*/
2197	pgmPoolCacheUsed(pPool, pPage);
2198
2199	int rc = pgmPoolTrackAddUser(pPool, pPage, iUser, iUserTable);
2200	if (RT_SUCCESS(rc))
2201	{
2202	Assert((PGMPOOLKIND)pPage->enmKind == enmKind);
2203	*ppPage = pPage;
2204	if (pPage->cModifications)
2205	pPage->cModifications = 1; /* reset counter (can't use 0, or else it will be reinserted in the modified list) */
2206	STAM_COUNTER_INC(&pPool->StatCacheHits);
2207	return VINF_PGM_CACHED_PAGE;
2208	}
2209	return rc;
2210	}
2211
2212	if ((PGMPOOLKIND)pPage->enmKind != enmKind)
2213	{
2214	/*
2215	* The kind is different. In some cases we should now flush the page
2216	* as it has been reused, but in most cases this is normal remapping
2217	* of PDs as PT or big pages using the GCPhys field in a slightly
2218	* different way than the other kinds.
2219	*/
2220	if (pgmPoolCacheReusedByKind((PGMPOOLKIND)pPage->enmKind, enmKind))
2221	{
2222	STAM_COUNTER_INC(&pPool->StatCacheKindMismatches);
2223	pgmPoolFlushPage(pPool, pPage);
2224	break;
2225	}
2226	}
2227	}
2228
2229	/* next */
2230	i = pPage->iNext;
2231	} while (i != NIL_PGMPOOL_IDX);
2232	}
2233
2234	Log3(("pgmPoolCacheAlloc: Missed GCPhys=%RGp enmKind=%s\n", GCPhys, pgmPoolPoolKindToStr(enmKind)));
2235	STAM_COUNTER_INC(&pPool->StatCacheMisses);
2236	return VERR_FILE_NOT_FOUND;
2237	}
2238
2239
2240	/**
2241	* Inserts a page into the cache.
2242	*
2243	* @param pPool The pool.
2244	* @param pPage The cached page.
2245	* @param fCanBeCached Set if the page is fit for caching from the caller's point of view.
2246	*/
2247	static void pgmPoolCacheInsert(PPGMPOOL pPool, PPGMPOOLPAGE pPage, bool fCanBeCached)
2248	{
2249	/*
2250	* Insert into the GCPhys hash if the page is fit for that.
2251	*/
2252	Assert(!pPage->fCached);
2253	if (fCanBeCached)
2254	{
2255	pPage->fCached = true;
2256	pgmPoolHashInsert(pPool, pPage);
2257	Log3(("pgmPoolCacheInsert: Caching %p:{.Core=%RHp, .idx=%d, .enmKind=%s, GCPhys=%RGp}\n",
2258	pPage, pPage->Core.Key, pPage->idx, pgmPoolPoolKindToStr(pPage->enmKind), pPage->GCPhys));
2259	STAM_COUNTER_INC(&pPool->StatCacheCacheable);
2260	}
2261	else
2262	{
2263	Log3(("pgmPoolCacheInsert: Not caching %p:{.Core=%RHp, .idx=%d, .enmKind=%s, GCPhys=%RGp}\n",
2264	pPage, pPage->Core.Key, pPage->idx, pgmPoolPoolKindToStr(pPage->enmKind), pPage->GCPhys));
2265	STAM_COUNTER_INC(&pPool->StatCacheUncacheable);
2266	}
2267
2268	/*
2269	* Insert at the head of the age list.
2270	*/
2271	pPage->iAgePrev = NIL_PGMPOOL_IDX;
2272	pPage->iAgeNext = pPool->iAgeHead;
2273	if (pPool->iAgeHead != NIL_PGMPOOL_IDX)
2274	pPool->aPages[pPool->iAgeHead].iAgePrev = pPage->idx;
2275	else
2276	pPool->iAgeTail = pPage->idx;
2277	pPool->iAgeHead = pPage->idx;
2278	}
2279
2280
2281	/**
2282	* Flushes a cached page.
2283	*
2284	* @param pPool The pool.
2285	* @param pPage The cached page.
2286	*/
2287	static void pgmPoolCacheFlushPage(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
2288	{
2289	Log3(("pgmPoolCacheFlushPage: %RGp\n", pPage->GCPhys));
2290
2291	/*
2292	* Remove the page from the hash.
2293	*/
2294	if (pPage->fCached)
2295	{
2296	pPage->fCached = false;
2297	pgmPoolHashRemove(pPool, pPage);
2298	}
2299	else
2300	Assert(pPage->iNext == NIL_PGMPOOL_IDX);
2301
2302	/*
2303	* Remove it from the age list.
2304	*/
2305	if (pPage->iAgeNext != NIL_PGMPOOL_IDX)
2306	pPool->aPages[pPage->iAgeNext].iAgePrev = pPage->iAgePrev;
2307	else
2308	pPool->iAgeTail = pPage->iAgePrev;
2309	if (pPage->iAgePrev != NIL_PGMPOOL_IDX)
2310	pPool->aPages[pPage->iAgePrev].iAgeNext = pPage->iAgeNext;
2311	else
2312	pPool->iAgeHead = pPage->iAgeNext;
2313	pPage->iAgeNext = NIL_PGMPOOL_IDX;
2314	pPage->iAgePrev = NIL_PGMPOOL_IDX;
2315	}
2316
2317
2318	/**
2319	* Looks for pages sharing the monitor.
2320	*
2321	* @returns Pointer to the head page.
2322	* @returns NULL if not found.
2323	* @param pPool The Pool
2324	* @param pNewPage The page which is going to be monitored.
2325	*/
2326	static PPGMPOOLPAGE pgmPoolMonitorGetPageByGCPhys(PPGMPOOL pPool, PPGMPOOLPAGE pNewPage)
2327	{
2328	/*
2329	* Look up the GCPhys in the hash.
2330	*/
2331	RTGCPHYS GCPhys = pNewPage->GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK;
2332	unsigned i = pPool->aiHash[PGMPOOL_HASH(GCPhys)];
2333	if (i == NIL_PGMPOOL_IDX)
2334	return NULL;
2335	do
2336	{
2337	PPGMPOOLPAGE pPage = &pPool->aPages[i];
2338	if ( pPage->GCPhys - GCPhys < PAGE_SIZE
2339	&& pPage != pNewPage)
2340	{
2341	switch (pPage->enmKind)
2342	{
2343	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
2344	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
2345	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
2346	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
2347	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
2348	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
2349	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
2350	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
2351	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
2352	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
2353	case PGMPOOLKIND_64BIT_PML4:
2354	case PGMPOOLKIND_32BIT_PD:
2355	case PGMPOOLKIND_PAE_PDPT:
2356	{
2357	/* find the head */
2358	while (pPage->iMonitoredPrev != NIL_PGMPOOL_IDX)
2359	{
2360	Assert(pPage->iMonitoredPrev != pPage->idx);
2361	pPage = &pPool->aPages[pPage->iMonitoredPrev];
2362	}
2363	return pPage;
2364	}
2365
2366	/* ignore, no monitoring. */
2367	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
2368	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
2369	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
2370	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
2371	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
2372	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
2373	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
2374	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
2375	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
2376	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
2377	case PGMPOOLKIND_ROOT_NESTED:
2378	case PGMPOOLKIND_PAE_PD_PHYS:
2379	case PGMPOOLKIND_PAE_PDPT_PHYS:
2380	case PGMPOOLKIND_32BIT_PD_PHYS:
2381	case PGMPOOLKIND_PAE_PDPT_FOR_32BIT:
2382	break;
2383	default:
2384	AssertFatalMsgFailed(("enmKind=%d idx=%d\n", pPage->enmKind, pPage->idx));
2385	}
2386	}
2387
2388	/* next */
2389	i = pPage->iNext;
2390	} while (i != NIL_PGMPOOL_IDX);
2391	return NULL;
2392	}
2393
2394
2395	/**
2396	* Enabled write monitoring of a guest page.
2397	*
2398	* @returns VBox status code.
2399	* @retval VINF_SUCCESS on success.
2400	* @param pPool The pool.
2401	* @param pPage The cached page.
2402	*/
2403	static int pgmPoolMonitorInsert(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
2404	{
2405	LogFlow(("pgmPoolMonitorInsert %RGp\n", pPage->GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK));
2406
2407	/*
2408	* Filter out the relevant kinds.
2409	*/
2410	switch (pPage->enmKind)
2411	{
2412	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
2413	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
2414	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
2415	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
2416	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
2417	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
2418	case PGMPOOLKIND_64BIT_PML4:
2419	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
2420	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
2421	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
2422	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
2423	case PGMPOOLKIND_32BIT_PD:
2424	case PGMPOOLKIND_PAE_PDPT:
2425	break;
2426
2427	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
2428	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
2429	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
2430	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
2431	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
2432	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
2433	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
2434	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
2435	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
2436	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
2437	case PGMPOOLKIND_ROOT_NESTED:
2438	/* Nothing to monitor here. */
2439	return VINF_SUCCESS;
2440
2441	case PGMPOOLKIND_32BIT_PD_PHYS:
2442	case PGMPOOLKIND_PAE_PDPT_PHYS:
2443	case PGMPOOLKIND_PAE_PD_PHYS:
2444	case PGMPOOLKIND_PAE_PDPT_FOR_32BIT:
2445	/* Nothing to monitor here. */
2446	return VINF_SUCCESS;
2447	default:
2448	AssertFatalMsgFailed(("This can't happen! enmKind=%d\n", pPage->enmKind));
2449	}
2450
2451	/*
2452	* Install handler.
2453	*/
2454	int rc;
2455	PPGMPOOLPAGE pPageHead = pgmPoolMonitorGetPageByGCPhys(pPool, pPage);
2456	if (pPageHead)
2457	{
2458	Assert(pPageHead != pPage); Assert(pPageHead->iMonitoredNext != pPage->idx);
2459	Assert(pPageHead->iMonitoredPrev != pPage->idx);
2460
2461	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
2462	if (pPageHead->fDirty)
2463	pgmPoolFlushDirtyPage(pPool->CTX_SUFF(pVM), pPool, pPageHead->idxDirtyEntry, false /* do not remove */);
2464	#endif
2465
2466	pPage->iMonitoredPrev = pPageHead->idx;
2467	pPage->iMonitoredNext = pPageHead->iMonitoredNext;
2468	if (pPageHead->iMonitoredNext != NIL_PGMPOOL_IDX)
2469	pPool->aPages[pPageHead->iMonitoredNext].iMonitoredPrev = pPage->idx;
2470	pPageHead->iMonitoredNext = pPage->idx;
2471	rc = VINF_SUCCESS;
2472	}
2473	else
2474	{
2475	Assert(pPage->iMonitoredNext == NIL_PGMPOOL_IDX); Assert(pPage->iMonitoredPrev == NIL_PGMPOOL_IDX);
2476	PVM pVM = pPool->CTX_SUFF(pVM);
2477	const RTGCPHYS GCPhysPage = pPage->GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK;
2478	rc = PGMHandlerPhysicalRegisterEx(pVM, PGMPHYSHANDLERTYPE_PHYSICAL_WRITE,
2479	GCPhysPage, GCPhysPage + PAGE_OFFSET_MASK,
2480	pPool->pfnAccessHandlerR3, MMHyperCCToR3(pVM, pPage),
2481	pPool->pfnAccessHandlerR0, MMHyperCCToR0(pVM, pPage),
2482	pPool->pfnAccessHandlerRC, MMHyperCCToRC(pVM, pPage),
2483	pPool->pszAccessHandler);
2484	/** @todo we should probably deal with out-of-memory conditions here, but for now increasing
2485	* the heap size should suffice. */
2486	AssertFatalMsgRC(rc, ("PGMHandlerPhysicalRegisterEx %RGp failed with %Rrc\n", GCPhysPage, rc));
2487	PVMCPU pVCpu = VMMGetCpu(pVM);
2488	AssertFatalMsg(!(pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL) \|\| VMCPU_FF_ISSET(pVCpu, VMCPU_FF_PGM_SYNC_CR3), ("fSyncFlags=%x syncff=%d\n", pVCpu->pgm.s.fSyncFlags, VMCPU_FF_ISSET(pVCpu, VMCPU_FF_PGM_SYNC_CR3)));
2489	}
2490	pPage->fMonitored = true;
2491	return rc;
2492	}
2493
2494
2495	/**
2496	* Disables write monitoring of a guest page.
2497	*
2498	* @returns VBox status code.
2499	* @retval VINF_SUCCESS on success.
2500	* @param pPool The pool.
2501	* @param pPage The cached page.
2502	*/
2503	static int pgmPoolMonitorFlush(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
2504	{
2505	/*
2506	* Filter out the relevant kinds.
2507	*/
2508	switch (pPage->enmKind)
2509	{
2510	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
2511	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
2512	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
2513	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
2514	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
2515	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
2516	case PGMPOOLKIND_64BIT_PML4:
2517	case PGMPOOLKIND_32BIT_PD:
2518	case PGMPOOLKIND_PAE_PDPT:
2519	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
2520	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
2521	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
2522	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
2523	break;
2524
2525	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
2526	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
2527	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
2528	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
2529	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
2530	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
2531	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
2532	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
2533	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
2534	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
2535	case PGMPOOLKIND_ROOT_NESTED:
2536	case PGMPOOLKIND_PAE_PD_PHYS:
2537	case PGMPOOLKIND_PAE_PDPT_PHYS:
2538	case PGMPOOLKIND_32BIT_PD_PHYS:
2539	/* Nothing to monitor here. */
2540	Assert(!pPage->fMonitored);
2541	return VINF_SUCCESS;
2542
2543	default:
2544	AssertFatalMsgFailed(("This can't happen! enmKind=%d\n", pPage->enmKind));
2545	}
2546	Assert(pPage->fMonitored);
2547
2548	/*
2549	* Remove the page from the monitored list or uninstall it if last.
2550	*/
2551	const PVM pVM = pPool->CTX_SUFF(pVM);
2552	int rc;
2553	if ( pPage->iMonitoredNext != NIL_PGMPOOL_IDX
2554	\|\| pPage->iMonitoredPrev != NIL_PGMPOOL_IDX)
2555	{
2556	if (pPage->iMonitoredPrev == NIL_PGMPOOL_IDX)
2557	{
2558	PPGMPOOLPAGE pNewHead = &pPool->aPages[pPage->iMonitoredNext];
2559	pNewHead->iMonitoredPrev = NIL_PGMPOOL_IDX;
2560	rc = PGMHandlerPhysicalChangeCallbacks(pVM, pPage->GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK,
2561	pPool->pfnAccessHandlerR3, MMHyperCCToR3(pVM, pNewHead),
2562	pPool->pfnAccessHandlerR0, MMHyperCCToR0(pVM, pNewHead),
2563	pPool->pfnAccessHandlerRC, MMHyperCCToRC(pVM, pNewHead),
2564	pPool->pszAccessHandler);
2565	AssertFatalRCSuccess(rc);
2566	pPage->iMonitoredNext = NIL_PGMPOOL_IDX;
2567	}
2568	else
2569	{
2570	pPool->aPages[pPage->iMonitoredPrev].iMonitoredNext = pPage->iMonitoredNext;
2571	if (pPage->iMonitoredNext != NIL_PGMPOOL_IDX)
2572	{
2573	pPool->aPages[pPage->iMonitoredNext].iMonitoredPrev = pPage->iMonitoredPrev;
2574	pPage->iMonitoredNext = NIL_PGMPOOL_IDX;
2575	}
2576	pPage->iMonitoredPrev = NIL_PGMPOOL_IDX;
2577	rc = VINF_SUCCESS;
2578	}
2579	}
2580	else
2581	{
2582	rc = PGMHandlerPhysicalDeregister(pVM, pPage->GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK);
2583	AssertFatalRC(rc);
2584	PVMCPU pVCpu = VMMGetCpu(pVM);
2585	AssertFatalMsg(!(pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL) \|\| VMCPU_FF_ISSET(pVCpu, VMCPU_FF_PGM_SYNC_CR3),
2586	("%#x %#x\n", pVCpu->pgm.s.fSyncFlags, pVM->fGlobalForcedActions));
2587	}
2588	pPage->fMonitored = false;
2589
2590	/*
2591	* Remove it from the list of modified pages (if in it).
2592	*/
2593	pgmPoolMonitorModifiedRemove(pPool, pPage);
2594
2595	return rc;
2596	}
2597
2598
2599	/**
2600	* Inserts the page into the list of modified pages.
2601	*
2602	* @param pPool The pool.
2603	* @param pPage The page.
2604	*/
2605	void pgmPoolMonitorModifiedInsert(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
2606	{
2607	Log3(("pgmPoolMonitorModifiedInsert: idx=%d\n", pPage->idx));
2608	AssertMsg( pPage->iModifiedNext == NIL_PGMPOOL_IDX
2609	&& pPage->iModifiedPrev == NIL_PGMPOOL_IDX
2610	&& pPool->iModifiedHead != pPage->idx,
2611	("Next=%d Prev=%d idx=%d cModifications=%d Head=%d cModifiedPages=%d\n",
2612	pPage->iModifiedNext, pPage->iModifiedPrev, pPage->idx, pPage->cModifications,
2613	pPool->iModifiedHead, pPool->cModifiedPages));
2614
2615	pPage->iModifiedNext = pPool->iModifiedHead;
2616	if (pPool->iModifiedHead != NIL_PGMPOOL_IDX)
2617	pPool->aPages[pPool->iModifiedHead].iModifiedPrev = pPage->idx;
2618	pPool->iModifiedHead = pPage->idx;
2619	pPool->cModifiedPages++;
2620	#ifdef VBOX_WITH_STATISTICS
2621	if (pPool->cModifiedPages > pPool->cModifiedPagesHigh)
2622	pPool->cModifiedPagesHigh = pPool->cModifiedPages;
2623	#endif
2624	}
2625
2626
2627	/**
2628	* Removes the page from the list of modified pages and resets the
2629	* modification counter.
2630	*
2631	* @param pPool The pool.
2632	* @param pPage The page which is believed to be in the list of modified pages.
2633	*/
2634	static void pgmPoolMonitorModifiedRemove(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
2635	{
2636	Log3(("pgmPoolMonitorModifiedRemove: idx=%d cModifications=%d\n", pPage->idx, pPage->cModifications));
2637	if (pPool->iModifiedHead == pPage->idx)
2638	{
2639	Assert(pPage->iModifiedPrev == NIL_PGMPOOL_IDX);
2640	pPool->iModifiedHead = pPage->iModifiedNext;
2641	if (pPage->iModifiedNext != NIL_PGMPOOL_IDX)
2642	{
2643	pPool->aPages[pPage->iModifiedNext].iModifiedPrev = NIL_PGMPOOL_IDX;
2644	pPage->iModifiedNext = NIL_PGMPOOL_IDX;
2645	}
2646	pPool->cModifiedPages--;
2647	}
2648	else if (pPage->iModifiedPrev != NIL_PGMPOOL_IDX)
2649	{
2650	pPool->aPages[pPage->iModifiedPrev].iModifiedNext = pPage->iModifiedNext;
2651	if (pPage->iModifiedNext != NIL_PGMPOOL_IDX)
2652	{
2653	pPool->aPages[pPage->iModifiedNext].iModifiedPrev = pPage->iModifiedPrev;
2654	pPage->iModifiedNext = NIL_PGMPOOL_IDX;
2655	}
2656	pPage->iModifiedPrev = NIL_PGMPOOL_IDX;
2657	pPool->cModifiedPages--;
2658	}
2659	else
2660	Assert(pPage->iModifiedPrev == NIL_PGMPOOL_IDX);
2661	pPage->cModifications = 0;
2662	}
2663
2664
2665	/**
2666	* Zaps the list of modified pages, resetting their modification counters in the process.
2667	*
2668	* @param pVM Pointer to the VM.
2669	*/
2670	static void pgmPoolMonitorModifiedClearAll(PVM pVM)
2671	{
2672	pgmLock(pVM);
2673	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
2674	LogFlow(("pgmPoolMonitorModifiedClearAll: cModifiedPages=%d\n", pPool->cModifiedPages));
2675
2676	unsigned cPages = 0; NOREF(cPages);
2677
2678	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
2679	pgmPoolResetDirtyPages(pVM);
2680	#endif
2681
2682	uint16_t idx = pPool->iModifiedHead;
2683	pPool->iModifiedHead = NIL_PGMPOOL_IDX;
2684	while (idx != NIL_PGMPOOL_IDX)
2685	{
2686	PPGMPOOLPAGE pPage = &pPool->aPages[idx];
2687	idx = pPage->iModifiedNext;
2688	pPage->iModifiedNext = NIL_PGMPOOL_IDX;
2689	pPage->iModifiedPrev = NIL_PGMPOOL_IDX;
2690	pPage->cModifications = 0;
2691	Assert(++cPages);
2692	}
2693	AssertMsg(cPages == pPool->cModifiedPages, ("%d != %d\n", cPages, pPool->cModifiedPages));
2694	pPool->cModifiedPages = 0;
2695	pgmUnlock(pVM);
2696	}
2697
2698
2699	/**
2700	* Handle SyncCR3 pool tasks
2701	*
2702	* @returns VBox status code.
2703	* @retval VINF_SUCCESS if successfully added.
2704	* @retval VINF_PGM_SYNC_CR3 is it needs to be deferred to ring 3 (GC only)
2705	* @param pVCpu The VMCPU handle.
2706	* @remark Should only be used when monitoring is available, thus placed in
2707	* the PGMPOOL_WITH_MONITORING #ifdef.
2708	*/
2709	int pgmPoolSyncCR3(PVMCPU pVCpu)
2710	{
2711	PVM pVM = pVCpu->CTX_SUFF(pVM);
2712	LogFlow(("pgmPoolSyncCR3 fSyncFlags=%x\n", pVCpu->pgm.s.fSyncFlags));
2713
2714	/*
2715	* When monitoring shadowed pages, we reset the modification counters on CR3 sync.
2716	* Occasionally we will have to clear all the shadow page tables because we wanted
2717	* to monitor a page which was mapped by too many shadowed page tables. This operation
2718	* sometimes referred to as a 'lightweight flush'.
2719	*/
2720	# ifdef IN_RING3 /* Don't flush in ring-0 or raw mode, it's taking too long. */
2721	if (pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL)
2722	pgmR3PoolClearAll(pVM, false /fFlushRemTlb/);
2723	# else /* !IN_RING3 */
2724	if (pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL)
2725	{
2726	Log(("SyncCR3: PGM_SYNC_CLEAR_PGM_POOL is set -> VINF_PGM_SYNC_CR3\n"));
2727	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3); /** @todo no need to do global sync, right? */
2728
2729	/* Make sure all other VCPUs return to ring 3. */
2730	if (pVM->cCpus > 1)
2731	{
2732	VM_FF_SET(pVM, VM_FF_PGM_POOL_FLUSH_PENDING);
2733	PGM_INVL_ALL_VCPU_TLBS(pVM);
2734	}
2735	return VINF_PGM_SYNC_CR3;
2736	}
2737	# endif /* !IN_RING3 */
2738	else
2739	{
2740	pgmPoolMonitorModifiedClearAll(pVM);
2741
2742	/* pgmPoolMonitorModifiedClearAll can cause a pgm pool flush (dirty page clearing), so make sure we handle this! */
2743	if (pVCpu->pgm.s.fSyncFlags & PGM_SYNC_CLEAR_PGM_POOL)
2744	{
2745	Log(("pgmPoolMonitorModifiedClearAll caused a pgm flush -> call pgmPoolSyncCR3 again!\n"));
2746	return pgmPoolSyncCR3(pVCpu);
2747	}
2748	}
2749	return VINF_SUCCESS;
2750	}
2751
2752
2753	/**
2754	* Frees up at least one user entry.
2755	*
2756	* @returns VBox status code.
2757	* @retval VINF_SUCCESS if successfully added.
2758	* @retval VERR_PGM_POOL_FLUSHED if the pool was flushed.
2759	* @param pPool The pool.
2760	* @param iUser The user index.
2761	*/
2762	static int pgmPoolTrackFreeOneUser(PPGMPOOL pPool, uint16_t iUser)
2763	{
2764	STAM_COUNTER_INC(&pPool->StatTrackFreeUpOneUser);
2765	/*
2766	* Just free cached pages in a braindead fashion.
2767	*/
2768	/** @todo walk the age list backwards and free the first with usage. */
2769	int rc = VINF_SUCCESS;
2770	do
2771	{
2772	int rc2 = pgmPoolCacheFreeOne(pPool, iUser);
2773	if (RT_FAILURE(rc2) && rc == VINF_SUCCESS)
2774	rc = rc2;
2775	} while (pPool->iUserFreeHead == NIL_PGMPOOL_USER_INDEX);
2776	return rc;
2777	}
2778
2779
2780	/**
2781	* Inserts a page into the cache.
2782	*
2783	* This will create user node for the page, insert it into the GCPhys
2784	* hash, and insert it into the age list.
2785	*
2786	* @returns VBox status code.
2787	* @retval VINF_SUCCESS if successfully added.
2788	* @retval VERR_PGM_POOL_FLUSHED if the pool was flushed.
2789	* @param pPool The pool.
2790	* @param pPage The cached page.
2791	* @param GCPhys The GC physical address of the page we're gonna shadow.
2792	* @param iUser The user index.
2793	* @param iUserTable The user table index.
2794	*/
2795	DECLINLINE(int) pgmPoolTrackInsert(PPGMPOOL pPool, PPGMPOOLPAGE pPage, RTGCPHYS GCPhys, uint16_t iUser, uint32_t iUserTable)
2796	{
2797	int rc = VINF_SUCCESS;
2798	PPGMPOOLUSER paUsers = pPool->CTX_SUFF(paUsers);
2799
2800	LogFlow(("pgmPoolTrackInsert GCPhys=%RGp iUser=%d iUserTable=%x\n", GCPhys, iUser, iUserTable));
2801
2802	#ifdef VBOX_STRICT
2803	/*
2804	* Check that the entry doesn't already exists.
2805	*/
2806	if (pPage->iUserHead != NIL_PGMPOOL_USER_INDEX)
2807	{
2808	uint16_t i = pPage->iUserHead;
2809	do
2810	{
2811	Assert(i < pPool->cMaxUsers);
2812	AssertMsg(paUsers[i].iUser != iUser \|\| paUsers[i].iUserTable != iUserTable, ("%x %x vs new %x %x\n", paUsers[i].iUser, paUsers[i].iUserTable, iUser, iUserTable));
2813	i = paUsers[i].iNext;
2814	} while (i != NIL_PGMPOOL_USER_INDEX);
2815	}
2816	#endif
2817
2818	/*
2819	* Find free a user node.
2820	*/
2821	uint16_t i = pPool->iUserFreeHead;
2822	if (i == NIL_PGMPOOL_USER_INDEX)
2823	{
2824	rc = pgmPoolTrackFreeOneUser(pPool, iUser);
2825	if (RT_FAILURE(rc))
2826	return rc;
2827	i = pPool->iUserFreeHead;
2828	}
2829
2830	/*
2831	* Unlink the user node from the free list,
2832	* initialize and insert it into the user list.
2833	*/
2834	pPool->iUserFreeHead = paUsers[i].iNext;
2835	paUsers[i].iNext = NIL_PGMPOOL_USER_INDEX;
2836	paUsers[i].iUser = iUser;
2837	paUsers[i].iUserTable = iUserTable;
2838	pPage->iUserHead = i;
2839
2840	/*
2841	* Insert into cache and enable monitoring of the guest page if enabled.
2842	*
2843	* Until we implement caching of all levels, including the CR3 one, we'll
2844	* have to make sure we don't try monitor & cache any recursive reuse of
2845	* a monitored CR3 page. Because all windows versions are doing this we'll
2846	* have to be able to do combined access monitoring, CR3 + PT and
2847	* PD + PT (guest PAE).
2848	*
2849	* Update:
2850	* We're now cooperating with the CR3 monitor if an uncachable page is found.
2851	*/
2852	const bool fCanBeMonitored = true;
2853	pgmPoolCacheInsert(pPool, pPage, fCanBeMonitored); /* This can be expanded. */
2854	if (fCanBeMonitored)
2855	{
2856	rc = pgmPoolMonitorInsert(pPool, pPage);
2857	AssertRC(rc);
2858	}
2859	return rc;
2860	}
2861
2862
2863	/**
2864	* Adds a user reference to a page.
2865	*
2866	* This will move the page to the head of the
2867	*
2868	* @returns VBox status code.
2869	* @retval VINF_SUCCESS if successfully added.
2870	* @retval VERR_PGM_POOL_FLUSHED if the pool was flushed.
2871	* @param pPool The pool.
2872	* @param pPage The cached page.
2873	* @param iUser The user index.
2874	* @param iUserTable The user table.
2875	*/
2876	static int pgmPoolTrackAddUser(PPGMPOOL pPool, PPGMPOOLPAGE pPage, uint16_t iUser, uint32_t iUserTable)
2877	{
2878	PPGMPOOLUSER paUsers = pPool->CTX_SUFF(paUsers);
2879
2880	Log3(("pgmPoolTrackAddUser GCPhys = %RGp iUser %x iUserTable %x\n", pPage->GCPhys, iUser, iUserTable));
2881
2882	# ifdef VBOX_STRICT
2883	/*
2884	* Check that the entry doesn't already exists. We only allow multiple
2885	* users of top-level paging structures (SHW_POOL_ROOT_IDX).
2886	*/
2887	if (pPage->iUserHead != NIL_PGMPOOL_USER_INDEX)
2888	{
2889	uint16_t i = pPage->iUserHead;
2890	do
2891	{
2892	Assert(i < pPool->cMaxUsers);
2893	AssertMsg(iUser != PGMPOOL_IDX_PD \|\| iUser != PGMPOOL_IDX_PDPT \|\| iUser != PGMPOOL_IDX_NESTED_ROOT \|\| iUser != PGMPOOL_IDX_AMD64_CR3 \|\|
2894	paUsers[i].iUser != iUser \|\| paUsers[i].iUserTable != iUserTable, ("%x %x vs new %x %x\n", paUsers[i].iUser, paUsers[i].iUserTable, iUser, iUserTable));
2895	i = paUsers[i].iNext;
2896	} while (i != NIL_PGMPOOL_USER_INDEX);
2897	}
2898	# endif
2899
2900	/*
2901	* Allocate a user node.
2902	*/
2903	uint16_t i = pPool->iUserFreeHead;
2904	if (i == NIL_PGMPOOL_USER_INDEX)
2905	{
2906	int rc = pgmPoolTrackFreeOneUser(pPool, iUser);
2907	if (RT_FAILURE(rc))
2908	return rc;
2909	i = pPool->iUserFreeHead;
2910	}
2911	pPool->iUserFreeHead = paUsers[i].iNext;
2912
2913	/*
2914	* Initialize the user node and insert it.
2915	*/
2916	paUsers[i].iNext = pPage->iUserHead;
2917	paUsers[i].iUser = iUser;
2918	paUsers[i].iUserTable = iUserTable;
2919	pPage->iUserHead = i;
2920
2921	# ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
2922	if (pPage->fDirty)
2923	pgmPoolFlushDirtyPage(pPool->CTX_SUFF(pVM), pPool, pPage->idxDirtyEntry, false /* do not remove */);
2924	# endif
2925
2926	/*
2927	* Tell the cache to update its replacement stats for this page.
2928	*/
2929	pgmPoolCacheUsed(pPool, pPage);
2930	return VINF_SUCCESS;
2931	}
2932
2933
2934	/**
2935	* Frees a user record associated with a page.
2936	*
2937	* This does not clear the entry in the user table, it simply replaces the
2938	* user record to the chain of free records.
2939	*
2940	* @param pPool The pool.
2941	* @param HCPhys The HC physical address of the shadow page.
2942	* @param iUser The shadow page pool index of the user table.
2943	* @param iUserTable The index into the user table (shadowed).
2944	*/
2945	static void pgmPoolTrackFreeUser(PPGMPOOL pPool, PPGMPOOLPAGE pPage, uint16_t iUser, uint32_t iUserTable)
2946	{
2947	/*
2948	* Unlink and free the specified user entry.
2949	*/
2950	PPGMPOOLUSER paUsers = pPool->CTX_SUFF(paUsers);
2951
2952	Log3(("pgmPoolTrackFreeUser %RGp %x %x\n", pPage->GCPhys, iUser, iUserTable));
2953	/* Special: For PAE and 32-bit paging, there is usually no more than one user. */
2954	uint16_t i = pPage->iUserHead;
2955	if ( i != NIL_PGMPOOL_USER_INDEX
2956	&& paUsers[i].iUser == iUser
2957	&& paUsers[i].iUserTable == iUserTable)
2958	{
2959	pPage->iUserHead = paUsers[i].iNext;
2960
2961	paUsers[i].iUser = NIL_PGMPOOL_IDX;
2962	paUsers[i].iNext = pPool->iUserFreeHead;
2963	pPool->iUserFreeHead = i;
2964	return;
2965	}
2966
2967	/* General: Linear search. */
2968	uint16_t iPrev = NIL_PGMPOOL_USER_INDEX;
2969	while (i != NIL_PGMPOOL_USER_INDEX)
2970	{
2971	if ( paUsers[i].iUser == iUser
2972	&& paUsers[i].iUserTable == iUserTable)
2973	{
2974	if (iPrev != NIL_PGMPOOL_USER_INDEX)
2975	paUsers[iPrev].iNext = paUsers[i].iNext;
2976	else
2977	pPage->iUserHead = paUsers[i].iNext;
2978
2979	paUsers[i].iUser = NIL_PGMPOOL_IDX;
2980	paUsers[i].iNext = pPool->iUserFreeHead;
2981	pPool->iUserFreeHead = i;
2982	return;
2983	}
2984	iPrev = i;
2985	i = paUsers[i].iNext;
2986	}
2987
2988	/* Fatal: didn't find it */
2989	AssertFatalMsgFailed(("Didn't find the user entry! iUser=%d iUserTable=%#x GCPhys=%RGp\n",
2990	iUser, iUserTable, pPage->GCPhys));
2991	}
2992
2993
2994	/**
2995	* Gets the entry size of a shadow table.
2996	*
2997	* @param enmKind The kind of page.
2998	*
2999	* @returns The size of the entry in bytes. That is, 4 or 8.
3000	* @returns If the kind is not for a table, an assertion is raised and 0 is
3001	* returned.
3002	*/
3003	DECLINLINE(unsigned) pgmPoolTrackGetShadowEntrySize(PGMPOOLKIND enmKind)
3004	{
3005	switch (enmKind)
3006	{
3007	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
3008	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
3009	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
3010	case PGMPOOLKIND_32BIT_PD:
3011	case PGMPOOLKIND_32BIT_PD_PHYS:
3012	return 4;
3013
3014	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
3015	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
3016	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
3017	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
3018	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
3019	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
3020	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
3021	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
3022	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
3023	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
3024	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
3025	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
3026	case PGMPOOLKIND_64BIT_PML4:
3027	case PGMPOOLKIND_PAE_PDPT:
3028	case PGMPOOLKIND_ROOT_NESTED:
3029	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
3030	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
3031	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
3032	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
3033	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
3034	case PGMPOOLKIND_PAE_PD_PHYS:
3035	case PGMPOOLKIND_PAE_PDPT_PHYS:
3036	return 8;
3037
3038	default:
3039	AssertFatalMsgFailed(("enmKind=%d\n", enmKind));
3040	}
3041	}
3042
3043
3044	/**
3045	* Gets the entry size of a guest table.
3046	*
3047	* @param enmKind The kind of page.
3048	*
3049	* @returns The size of the entry in bytes. That is, 0, 4 or 8.
3050	* @returns If the kind is not for a table, an assertion is raised and 0 is
3051	* returned.
3052	*/
3053	DECLINLINE(unsigned) pgmPoolTrackGetGuestEntrySize(PGMPOOLKIND enmKind)
3054	{
3055	switch (enmKind)
3056	{
3057	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
3058	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
3059	case PGMPOOLKIND_32BIT_PD:
3060	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
3061	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
3062	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
3063	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
3064	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
3065	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
3066	return 4;
3067
3068	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
3069	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
3070	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
3071	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
3072	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
3073	case PGMPOOLKIND_64BIT_PML4:
3074	case PGMPOOLKIND_PAE_PDPT:
3075	return 8;
3076
3077	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
3078	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
3079	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
3080	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
3081	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
3082	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
3083	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
3084	case PGMPOOLKIND_ROOT_NESTED:
3085	case PGMPOOLKIND_PAE_PD_PHYS:
3086	case PGMPOOLKIND_PAE_PDPT_PHYS:
3087	case PGMPOOLKIND_32BIT_PD_PHYS:
3088	/** @todo can we return 0? (nobody is calling this...) */
3089	AssertFailed();
3090	return 0;
3091
3092	default:
3093	AssertFatalMsgFailed(("enmKind=%d\n", enmKind));
3094	}
3095	}
3096
3097
3098	/**
3099	* Checks one shadow page table entry for a mapping of a physical page.
3100	*
3101	* @returns true / false indicating removal of all relevant PTEs
3102	*
3103	* @param pVM Pointer to the VM.
3104	* @param pPhysPage The guest page in question.
3105	* @param fFlushPTEs Flush PTEs or allow them to be updated (e.g. in case of an RW bit change)
3106	* @param iShw The shadow page table.
3107	* @param iPte Page table entry or NIL_PGMPOOL_PHYSEXT_IDX_PTE if unknown
3108	*/
3109	static bool pgmPoolTrackFlushGCPhysPTInt(PVM pVM, PCPGMPAGE pPhysPage, bool fFlushPTEs, uint16_t iShw, uint16_t iPte)
3110	{
3111	LogFlow(("pgmPoolTrackFlushGCPhysPTInt: pPhysPage=%RHp iShw=%d iPte=%d\n", PGM_PAGE_GET_HCPHYS(pPhysPage), iShw, iPte));
3112	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3113	bool fRet = false;
3114
3115	/*
3116	* Assert sanity.
3117	*/
3118	Assert(iPte != NIL_PGMPOOL_PHYSEXT_IDX_PTE);
3119	AssertFatalMsg(iShw < pPool->cCurPages && iShw != NIL_PGMPOOL_IDX, ("iShw=%d\n", iShw));
3120	PPGMPOOLPAGE pPage = &pPool->aPages[iShw];
3121
3122	/*
3123	* Then, clear the actual mappings to the page in the shadow PT.
3124	*/
3125	switch (pPage->enmKind)
3126	{
3127	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
3128	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
3129	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
3130	{
3131	const uint32_t u32 = PGM_PAGE_GET_HCPHYS(pPhysPage) \| X86_PTE_P;
3132	PX86PT pPT = (PX86PT)PGMPOOL_PAGE_2_PTR(pVM, pPage);
3133	uint32_t u32AndMask = 0;
3134	uint32_t u32OrMask = 0;
3135
3136	if (!fFlushPTEs)
3137	{
3138	switch (PGM_PAGE_GET_HNDL_PHYS_STATE(pPhysPage))
3139	{
3140	case PGM_PAGE_HNDL_PHYS_STATE_NONE: /** No handler installed. */
3141	case PGM_PAGE_HNDL_PHYS_STATE_DISABLED: /** Monitoring is temporarily disabled. */
3142	u32OrMask = X86_PTE_RW;
3143	u32AndMask = UINT32_MAX;
3144	fRet = true;
3145	STAM_COUNTER_INC(&pPool->StatTrackFlushEntryKeep);
3146	break;
3147
3148	case PGM_PAGE_HNDL_PHYS_STATE_WRITE: /** Write access is monitored. */
3149	u32OrMask = 0;
3150	u32AndMask = ~X86_PTE_RW;
3151	fRet = true;
3152	STAM_COUNTER_INC(&pPool->StatTrackFlushEntryKeep);
3153	break;
3154	default:
3155	/* (shouldn't be here, will assert below) */
3156	STAM_COUNTER_INC(&pPool->StatTrackFlushEntry);
3157	break;
3158	}
3159	}
3160	else
3161	STAM_COUNTER_INC(&pPool->StatTrackFlushEntry);
3162
3163	/* Update the counter if we're removing references. */
3164	if (!u32AndMask)
3165	{
3166	Assert(pPage->cPresent );
3167	Assert(pPool->cPresent);
3168	pPage->cPresent--;
3169	pPool->cPresent--;
3170	}
3171
3172	if ((pPT->a[iPte].u & (X86_PTE_PG_MASK \| X86_PTE_P)) == u32)
3173	{
3174	X86PTE Pte;
3175
3176	Log4(("pgmPoolTrackFlushGCPhysPTs: i=%d pte=%RX32\n", iPte, pPT->a[iPte]));
3177	Pte.u = (pPT->a[iPte].u & u32AndMask) \| u32OrMask;
3178	if (Pte.u & PGM_PTFLAGS_TRACK_DIRTY)
3179	Pte.n.u1Write = 0; /* need to disallow writes when dirty bit tracking is still active. */
3180
3181	ASMAtomicWriteU32(&pPT->a[iPte].u, Pte.u);
3182	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPT);
3183	return fRet;
3184	}
3185	#ifdef LOG_ENABLED
3186	Log(("iFirstPresent=%d cPresent=%d\n", pPage->iFirstPresent, pPage->cPresent));
3187	for (unsigned i = 0, cFound = 0; i < RT_ELEMENTS(pPT->a); i++)
3188	if ((pPT->a[i].u & (X86_PTE_PG_MASK \| X86_PTE_P)) == u32)
3189	{
3190	Log(("i=%d cFound=%d\n", i, ++cFound));
3191	}
3192	#endif
3193	AssertFatalMsgFailed(("iFirstPresent=%d cPresent=%d u32=%RX32 poolkind=%x\n", pPage->iFirstPresent, pPage->cPresent, u32, pPage->enmKind));
3194	/PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPT);/
3195	break;
3196	}
3197
3198	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
3199	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
3200	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
3201	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
3202	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
3203	case PGMPOOLKIND_EPT_PT_FOR_PHYS: /* physical mask the same as PAE; RW bit as well; be careful! */
3204	{
3205	const uint64_t u64 = PGM_PAGE_GET_HCPHYS(pPhysPage) \| X86_PTE_P;
3206	PPGMSHWPTPAE pPT = (PPGMSHWPTPAE)PGMPOOL_PAGE_2_PTR(pVM, pPage);
3207	uint64_t u64OrMask = 0;
3208	uint64_t u64AndMask = 0;
3209
3210	if (!fFlushPTEs)
3211	{
3212	switch (PGM_PAGE_GET_HNDL_PHYS_STATE(pPhysPage))
3213	{
3214	case PGM_PAGE_HNDL_PHYS_STATE_NONE: /* No handler installed. */
3215	case PGM_PAGE_HNDL_PHYS_STATE_DISABLED: /* Monitoring is temporarily disabled. */
3216	u64OrMask = X86_PTE_RW;
3217	u64AndMask = UINT64_MAX;
3218	fRet = true;
3219	STAM_COUNTER_INC(&pPool->StatTrackFlushEntryKeep);
3220	break;
3221
3222	case PGM_PAGE_HNDL_PHYS_STATE_WRITE: /* Write access is monitored. */
3223	u64OrMask = 0;
3224	u64AndMask = ~(uint64_t)X86_PTE_RW;
3225	fRet = true;
3226	STAM_COUNTER_INC(&pPool->StatTrackFlushEntryKeep);
3227	break;
3228
3229	default:
3230	/* (shouldn't be here, will assert below) */
3231	STAM_COUNTER_INC(&pPool->StatTrackFlushEntry);
3232	break;
3233	}
3234	}
3235	else
3236	STAM_COUNTER_INC(&pPool->StatTrackFlushEntry);
3237
3238	/* Update the counter if we're removing references. */
3239	if (!u64AndMask)
3240	{
3241	Assert(pPage->cPresent);
3242	Assert(pPool->cPresent);
3243	pPage->cPresent--;
3244	pPool->cPresent--;
3245	}
3246
3247	if ((PGMSHWPTEPAE_GET_U(pPT->a[iPte]) & (X86_PTE_PAE_PG_MASK \| X86_PTE_P \| X86_PTE_PAE_MBZ_MASK_NX)) == u64)
3248	{
3249	X86PTEPAE Pte;
3250
3251	Log4(("pgmPoolTrackFlushGCPhysPTs: i=%d pte=%RX64\n", iPte, PGMSHWPTEPAE_GET_LOG(pPT->a[iPte])));
3252	Pte.u = (PGMSHWPTEPAE_GET_U(pPT->a[iPte]) & u64AndMask) \| u64OrMask;
3253	if (Pte.u & PGM_PTFLAGS_TRACK_DIRTY)
3254	Pte.n.u1Write = 0; /* need to disallow writes when dirty bit tracking is still active. */
3255
3256	PGMSHWPTEPAE_ATOMIC_SET(pPT->a[iPte], Pte.u);
3257	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPT);
3258	return fRet;
3259	}
3260	#ifdef LOG_ENABLED
3261	Log(("iFirstPresent=%d cPresent=%d\n", pPage->iFirstPresent, pPage->cPresent));
3262	Log(("Found %RX64 expected %RX64\n", PGMSHWPTEPAE_GET_U(pPT->a[iPte]) & (X86_PTE_PAE_PG_MASK \| X86_PTE_P \| X86_PTE_PAE_MBZ_MASK_NX), u64));
3263	for (unsigned i = 0, cFound = 0; i < RT_ELEMENTS(pPT->a); i++)
3264	if ((PGMSHWPTEPAE_GET_U(pPT->a[i]) & (X86_PTE_PAE_PG_MASK \| X86_PTE_P \| X86_PTE_PAE_MBZ_MASK_NX)) == u64)
3265	Log(("i=%d cFound=%d\n", i, ++cFound));
3266	#endif
3267	AssertFatalMsgFailed(("iFirstPresent=%d cPresent=%d u64=%RX64 poolkind=%x iPte=%d PT=%RX64\n", pPage->iFirstPresent, pPage->cPresent, u64, pPage->enmKind, iPte, PGMSHWPTEPAE_GET_LOG(pPT->a[iPte])));
3268	/PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPT);/
3269	break;
3270	}
3271
3272	#ifdef PGM_WITH_LARGE_PAGES
3273	/* Large page case only. */
3274	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
3275	{
3276	Assert(pVM->pgm.s.fNestedPaging);
3277
3278	const uint64_t u64 = PGM_PAGE_GET_HCPHYS(pPhysPage) \| X86_PDE4M_P \| X86_PDE4M_PS;
3279	PEPTPD pPD = (PEPTPD)PGMPOOL_PAGE_2_PTR(pVM, pPage);
3280
3281	if ((pPD->a[iPte].u & (EPT_PDE2M_PG_MASK \| X86_PDE4M_P \| X86_PDE4M_PS)) == u64)
3282	{
3283	Log4(("pgmPoolTrackFlushGCPhysPTs: i=%d pde=%RX64\n", iPte, pPD->a[iPte]));
3284	STAM_COUNTER_INC(&pPool->StatTrackFlushEntry);
3285	pPD->a[iPte].u = 0;
3286	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPD);
3287
3288	/* Update the counter as we're removing references. */
3289	Assert(pPage->cPresent);
3290	Assert(pPool->cPresent);
3291	pPage->cPresent--;
3292	pPool->cPresent--;
3293
3294	return fRet;
3295	}
3296	# ifdef LOG_ENABLED
3297	Log(("iFirstPresent=%d cPresent=%d\n", pPage->iFirstPresent, pPage->cPresent));
3298	for (unsigned i = 0, cFound = 0; i < RT_ELEMENTS(pPD->a); i++)
3299	if ((pPD->a[i].u & (EPT_PDE2M_PG_MASK \| X86_PDE4M_P \| X86_PDE4M_PS)) == u64)
3300	Log(("i=%d cFound=%d\n", i, ++cFound));
3301	# endif
3302	AssertFatalMsgFailed(("iFirstPresent=%d cPresent=%d\n", pPage->iFirstPresent, pPage->cPresent));
3303	/PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPD);/
3304	break;
3305	}
3306
3307	/* AMD-V nested paging / /* @todo merge with EPT as we only check the parts that are identical. */
3308	case PGMPOOLKIND_PAE_PD_PHYS:
3309	{
3310	Assert(pVM->pgm.s.fNestedPaging);
3311
3312	const uint64_t u64 = PGM_PAGE_GET_HCPHYS(pPhysPage) \| X86_PDE4M_P \| X86_PDE4M_PS;
3313	PX86PD pPD = (PX86PD)PGMPOOL_PAGE_2_PTR(pVM, pPage);
3314
3315	if ((pPD->a[iPte].u & (X86_PDE2M_PAE_PG_MASK \| X86_PDE4M_P \| X86_PDE4M_PS)) == u64)
3316	{
3317	Log4(("pgmPoolTrackFlushGCPhysPTs: i=%d pde=%RX64\n", iPte, pPD->a[iPte]));
3318	STAM_COUNTER_INC(&pPool->StatTrackFlushEntry);
3319	pPD->a[iPte].u = 0;
3320	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPD);
3321
3322	/* Update the counter as we're removing references. */
3323	Assert(pPage->cPresent);
3324	Assert(pPool->cPresent);
3325	pPage->cPresent--;
3326	pPool->cPresent--;
3327	return fRet;
3328	}
3329	# ifdef LOG_ENABLED
3330	Log(("iFirstPresent=%d cPresent=%d\n", pPage->iFirstPresent, pPage->cPresent));
3331	for (unsigned i = 0, cFound = 0; i < RT_ELEMENTS(pPD->a); i++)
3332	if ((pPD->a[i].u & (X86_PDE2M_PAE_PG_MASK \| X86_PDE4M_P \| X86_PDE4M_PS)) == u64)
3333	Log(("i=%d cFound=%d\n", i, ++cFound));
3334	# endif
3335	AssertFatalMsgFailed(("iFirstPresent=%d cPresent=%d\n", pPage->iFirstPresent, pPage->cPresent));
3336	/PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPD);/
3337	break;
3338	}
3339	#endif /* PGM_WITH_LARGE_PAGES */
3340
3341	default:
3342	AssertFatalMsgFailed(("enmKind=%d iShw=%d\n", pPage->enmKind, iShw));
3343	}
3344
3345	/* not reached. */
3346	#ifndef _MSC_VER
3347	return fRet;
3348	#endif
3349	}
3350
3351
3352	/**
3353	* Scans one shadow page table for mappings of a physical page.
3354	*
3355	* @param pVM Pointer to the VM.
3356	* @param pPhysPage The guest page in question.
3357	* @param fFlushPTEs Flush PTEs or allow them to be updated (e.g. in case of an RW bit change)
3358	* @param iShw The shadow page table.
3359	*/
3360	static void pgmPoolTrackFlushGCPhysPT(PVM pVM, PPGMPAGE pPhysPage, bool fFlushPTEs, uint16_t iShw)
3361	{
3362	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool); NOREF(pPool);
3363
3364	/* We should only come here with when there's only one reference to this physical page. */
3365	Assert(PGMPOOL_TD_GET_CREFS(PGM_PAGE_GET_TRACKING(pPhysPage)) == 1);
3366
3367	Log2(("pgmPoolTrackFlushGCPhysPT: pPhysPage=%RHp iShw=%d\n", PGM_PAGE_GET_HCPHYS(pPhysPage), iShw));
3368	STAM_PROFILE_START(&pPool->StatTrackFlushGCPhysPT, f);
3369	bool fKeptPTEs = pgmPoolTrackFlushGCPhysPTInt(pVM, pPhysPage, fFlushPTEs, iShw, PGM_PAGE_GET_PTE_INDEX(pPhysPage));
3370	if (!fKeptPTEs)
3371	PGM_PAGE_SET_TRACKING(pVM, pPhysPage, 0);
3372	STAM_PROFILE_STOP(&pPool->StatTrackFlushGCPhysPT, f);
3373	}
3374
3375
3376	/**
3377	* Flushes a list of shadow page tables mapping the same physical page.
3378	*
3379	* @param pVM Pointer to the VM.
3380	* @param pPhysPage The guest page in question.
3381	* @param fFlushPTEs Flush PTEs or allow them to be updated (e.g. in case of an RW bit change)
3382	* @param iPhysExt The physical cross reference extent list to flush.
3383	*/
3384	static void pgmPoolTrackFlushGCPhysPTs(PVM pVM, PPGMPAGE pPhysPage, bool fFlushPTEs, uint16_t iPhysExt)
3385	{
3386	PGM_LOCK_ASSERT_OWNER(pVM);
3387	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3388	bool fKeepList = false;
3389
3390	STAM_PROFILE_START(&pPool->StatTrackFlushGCPhysPTs, f);
3391	Log2(("pgmPoolTrackFlushGCPhysPTs: pPhysPage=%RHp iPhysExt\n", PGM_PAGE_GET_HCPHYS(pPhysPage), iPhysExt));
3392
3393	const uint16_t iPhysExtStart = iPhysExt;
3394	PPGMPOOLPHYSEXT pPhysExt;
3395	do
3396	{
3397	Assert(iPhysExt < pPool->cMaxPhysExts);
3398	pPhysExt = &pPool->CTX_SUFF(paPhysExts)[iPhysExt];
3399	for (unsigned i = 0; i < RT_ELEMENTS(pPhysExt->aidx); i++)
3400	{
3401	if (pPhysExt->aidx[i] != NIL_PGMPOOL_IDX)
3402	{
3403	bool fKeptPTEs = pgmPoolTrackFlushGCPhysPTInt(pVM, pPhysPage, fFlushPTEs, pPhysExt->aidx[i], pPhysExt->apte[i]);
3404	if (!fKeptPTEs)
3405	{
3406	pPhysExt->aidx[i] = NIL_PGMPOOL_IDX;
3407	pPhysExt->apte[i] = NIL_PGMPOOL_PHYSEXT_IDX_PTE;
3408	}
3409	else
3410	fKeepList = true;
3411	}
3412	}
3413	/* next */
3414	iPhysExt = pPhysExt->iNext;
3415	} while (iPhysExt != NIL_PGMPOOL_PHYSEXT_INDEX);
3416
3417	if (!fKeepList)
3418	{
3419	/* insert the list into the free list and clear the ram range entry. */
3420	pPhysExt->iNext = pPool->iPhysExtFreeHead;
3421	pPool->iPhysExtFreeHead = iPhysExtStart;
3422	/* Invalidate the tracking data. */
3423	PGM_PAGE_SET_TRACKING(pVM, pPhysPage, 0);
3424	}
3425
3426	STAM_PROFILE_STOP(&pPool->StatTrackFlushGCPhysPTs, f);
3427	}
3428
3429
3430	/**
3431	* Flushes all shadow page table mappings of the given guest page.
3432	*
3433	* This is typically called when the host page backing the guest one has been
3434	* replaced or when the page protection was changed due to a guest access
3435	* caught by the monitoring.
3436	*
3437	* @returns VBox status code.
3438	* @retval VINF_SUCCESS if all references has been successfully cleared.
3439	* @retval VINF_PGM_SYNC_CR3 if we're better off with a CR3 sync and a page
3440	* pool cleaning. FF and sync flags are set.
3441	*
3442	* @param pVM Pointer to the VM.
3443	* @param GCPhysPage GC physical address of the page in question
3444	* @param pPhysPage The guest page in question.
3445	* @param fFlushPTEs Flush PTEs or allow them to be updated (e.g. in case of an RW bit change)
3446	* @param pfFlushTLBs This is set to @a true if the shadow TLBs should be
3447	* flushed, it is NOT touched if this isn't necessary.
3448	* The caller MUST initialized this to @a false.
3449	*/
3450	int pgmPoolTrackUpdateGCPhys(PVM pVM, RTGCPHYS GCPhysPage, PPGMPAGE pPhysPage, bool fFlushPTEs, bool *pfFlushTLBs)
3451	{
3452	PVMCPU pVCpu = VMMGetCpu(pVM);
3453	pgmLock(pVM);
3454	int rc = VINF_SUCCESS;
3455
3456	#ifdef PGM_WITH_LARGE_PAGES
3457	/* Is this page part of a large page? */
3458	if (PGM_PAGE_GET_PDE_TYPE(pPhysPage) == PGM_PAGE_PDE_TYPE_PDE)
3459	{
3460	RTGCPHYS GCPhysBase = GCPhysPage & X86_PDE2M_PAE_PG_MASK;
3461	GCPhysPage &= X86_PDE_PAE_PG_MASK;
3462
3463	/* Fetch the large page base. */
3464	PPGMPAGE pLargePage;
3465	if (GCPhysBase != GCPhysPage)
3466	{
3467	pLargePage = pgmPhysGetPage(pVM, GCPhysBase);
3468	AssertFatal(pLargePage);
3469	}
3470	else
3471	pLargePage = pPhysPage;
3472
3473	Log(("pgmPoolTrackUpdateGCPhys: update large page PDE for %RGp (%RGp)\n", GCPhysBase, GCPhysPage));
3474
3475	if (PGM_PAGE_GET_PDE_TYPE(pLargePage) == PGM_PAGE_PDE_TYPE_PDE)
3476	{
3477	/* Mark the large page as disabled as we need to break it up to change a single page in the 2 MB range. */
3478	PGM_PAGE_SET_PDE_TYPE(pVM, pLargePage, PGM_PAGE_PDE_TYPE_PDE_DISABLED);
3479	pVM->pgm.s.cLargePagesDisabled++;
3480
3481	/* Update the base as that only that one has a reference and there's only one PDE to clear. */
3482	rc = pgmPoolTrackUpdateGCPhys(pVM, GCPhysBase, pLargePage, fFlushPTEs, pfFlushTLBs);
3483
3484	*pfFlushTLBs = true;
3485	pgmUnlock(pVM);
3486	return rc;
3487	}
3488	}
3489	#else
3490	NOREF(GCPhysPage);
3491	#endif /* PGM_WITH_LARGE_PAGES */
3492
3493	const uint16_t u16 = PGM_PAGE_GET_TRACKING(pPhysPage);
3494	if (u16)
3495	{
3496	/*
3497	* The zero page is currently screwing up the tracking and we'll
3498	* have to flush the whole shebang. Unless VBOX_WITH_NEW_LAZY_PAGE_ALLOC
3499	* is defined, zero pages won't normally be mapped. Some kind of solution
3500	* will be needed for this problem of course, but it will have to wait...
3501	*/
3502	if ( PGM_PAGE_IS_ZERO(pPhysPage)
3503	\|\| PGM_PAGE_IS_BALLOONED(pPhysPage))
3504	rc = VINF_PGM_GCPHYS_ALIASED;
3505	else
3506	{
3507	# if defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) \|\| defined(IN_RC) /** @todo we can drop this now. */
3508	/* Start a subset here because pgmPoolTrackFlushGCPhysPTsSlow and
3509	pgmPoolTrackFlushGCPhysPTs will/may kill the pool otherwise. */
3510	uint32_t iPrevSubset = PGMRZDynMapPushAutoSubset(pVCpu);
3511	# endif
3512
3513	if (PGMPOOL_TD_GET_CREFS(u16) != PGMPOOL_TD_CREFS_PHYSEXT)
3514	{
3515	Assert(PGMPOOL_TD_GET_CREFS(u16) == 1);
3516	pgmPoolTrackFlushGCPhysPT(pVM,
3517	pPhysPage,
3518	fFlushPTEs,
3519	PGMPOOL_TD_GET_IDX(u16));
3520	}
3521	else if (u16 != PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, PGMPOOL_TD_IDX_OVERFLOWED))
3522	pgmPoolTrackFlushGCPhysPTs(pVM, pPhysPage, fFlushPTEs, PGMPOOL_TD_GET_IDX(u16));
3523	else
3524	rc = pgmPoolTrackFlushGCPhysPTsSlow(pVM, pPhysPage);
3525	*pfFlushTLBs = true;
3526
3527	# if defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) \|\| defined(IN_RC)
3528	PGMRZDynMapPopAutoSubset(pVCpu, iPrevSubset);
3529	# endif
3530	}
3531	}
3532
3533	if (rc == VINF_PGM_GCPHYS_ALIASED)
3534	{
3535	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3536	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3537	rc = VINF_PGM_SYNC_CR3;
3538	}
3539	pgmUnlock(pVM);
3540	return rc;
3541	}
3542
3543
3544	/**
3545	* Scans all shadow page tables for mappings of a physical page.
3546	*
3547	* This may be slow, but it's most likely more efficient than cleaning
3548	* out the entire page pool / cache.
3549	*
3550	* @returns VBox status code.
3551	* @retval VINF_SUCCESS if all references has been successfully cleared.
3552	* @retval VINF_PGM_GCPHYS_ALIASED if we're better off with a CR3 sync and
3553	* a page pool cleaning.
3554	*
3555	* @param pVM Pointer to the VM.
3556	* @param pPhysPage The guest page in question.
3557	*/
3558	int pgmPoolTrackFlushGCPhysPTsSlow(PVM pVM, PPGMPAGE pPhysPage)
3559	{
3560	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3561	STAM_PROFILE_START(&pPool->StatTrackFlushGCPhysPTsSlow, s);
3562	LogFlow(("pgmPoolTrackFlushGCPhysPTsSlow: cUsedPages=%d cPresent=%d pPhysPage=%R[pgmpage]\n",
3563	pPool->cUsedPages, pPool->cPresent, pPhysPage));
3564
3565	/*
3566	* There is a limit to what makes sense.
3567	*/
3568	if ( pPool->cPresent > 1024
3569	&& pVM->cCpus == 1)
3570	{
3571	LogFlow(("pgmPoolTrackFlushGCPhysPTsSlow: giving up... (cPresent=%d)\n", pPool->cPresent));
3572	STAM_PROFILE_STOP(&pPool->StatTrackFlushGCPhysPTsSlow, s);
3573	return VINF_PGM_GCPHYS_ALIASED;
3574	}
3575
3576	/*
3577	* Iterate all the pages until we've encountered all that in use.
3578	* This is simple but not quite optimal solution.
3579	*/
3580	const uint64_t u64 = PGM_PAGE_GET_HCPHYS(pPhysPage) \| X86_PTE_P; /** @todo drop X86_PTE_P here as we always test if present separately, anyway. */
3581	const uint32_t u32 = u64; /** @todo move into the 32BIT_PT_xx case */
3582	unsigned cLeft = pPool->cUsedPages;
3583	unsigned iPage = pPool->cCurPages;
3584	while (--iPage >= PGMPOOL_IDX_FIRST)
3585	{
3586	PPGMPOOLPAGE pPage = &pPool->aPages[iPage];
3587	if ( pPage->GCPhys != NIL_RTGCPHYS
3588	&& pPage->cPresent)
3589	{
3590	switch (pPage->enmKind)
3591	{
3592	/*
3593	* We only care about shadow page tables.
3594	*/
3595	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
3596	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
3597	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
3598	{
3599	unsigned cPresent = pPage->cPresent;
3600	PX86PT pPT = (PX86PT)PGMPOOL_PAGE_2_PTR(pVM, pPage);
3601	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pPT->a); i++)
3602	if (pPT->a[i].n.u1Present)
3603	{
3604	if ((pPT->a[i].u & (X86_PTE_PG_MASK \| X86_PTE_P)) == u32)
3605	{
3606	//Log4(("pgmPoolTrackFlushGCPhysPTsSlow: idx=%d i=%d pte=%RX32\n", iPage, i, pPT->a[i]));
3607	pPT->a[i].u = 0;
3608
3609	/* Update the counter as we're removing references. */
3610	Assert(pPage->cPresent);
3611	Assert(pPool->cPresent);
3612	pPage->cPresent--;
3613	pPool->cPresent--;
3614	}
3615	if (!--cPresent)
3616	break;
3617	}
3618	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPT);
3619	break;
3620	}
3621
3622	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
3623	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
3624	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
3625	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
3626	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
3627	{
3628	unsigned cPresent = pPage->cPresent;
3629	PPGMSHWPTPAE pPT = (PPGMSHWPTPAE)PGMPOOL_PAGE_2_PTR(pVM, pPage);
3630	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pPT->a); i++)
3631	if (PGMSHWPTEPAE_IS_P(pPT->a[i]))
3632	{
3633	if ((PGMSHWPTEPAE_GET_U(pPT->a[i]) & (X86_PTE_PAE_PG_MASK \| X86_PTE_P)) == u64)
3634	{
3635	//Log4(("pgmPoolTrackFlushGCPhysPTsSlow: idx=%d i=%d pte=%RX64\n", iPage, i, pPT->a[i]));
3636	PGMSHWPTEPAE_SET(pPT->a[i], 0); /// @todo why not atomic?
3637
3638	/* Update the counter as we're removing references. */
3639	Assert(pPage->cPresent);
3640	Assert(pPool->cPresent);
3641	pPage->cPresent--;
3642	pPool->cPresent--;
3643	}
3644	if (!--cPresent)
3645	break;
3646	}
3647	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPT);
3648	break;
3649	}
3650	#ifndef IN_RC
3651	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
3652	{
3653	unsigned cPresent = pPage->cPresent;
3654	PEPTPT pPT = (PEPTPT)PGMPOOL_PAGE_2_PTR(pVM, pPage);
3655	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pPT->a); i++)
3656	if (pPT->a[i].n.u1Present)
3657	{
3658	if ((pPT->a[i].u & (EPT_PTE_PG_MASK \| X86_PTE_P)) == u64)
3659	{
3660	//Log4(("pgmPoolTrackFlushGCPhysPTsSlow: idx=%d i=%d pte=%RX64\n", iPage, i, pPT->a[i]));
3661	pPT->a[i].u = 0;
3662
3663	/* Update the counter as we're removing references. */
3664	Assert(pPage->cPresent);
3665	Assert(pPool->cPresent);
3666	pPage->cPresent--;
3667	pPool->cPresent--;
3668	}
3669	if (!--cPresent)
3670	break;
3671	}
3672	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pPT);
3673	break;
3674	}
3675	#endif
3676	}
3677	if (!--cLeft)
3678	break;
3679	}
3680	}
3681
3682	PGM_PAGE_SET_TRACKING(pVM, pPhysPage, 0);
3683	STAM_PROFILE_STOP(&pPool->StatTrackFlushGCPhysPTsSlow, s);
3684
3685	/*
3686	* There is a limit to what makes sense. The above search is very expensive, so force a pgm pool flush.
3687	*/
3688	if (pPool->cPresent > 1024)
3689	{
3690	LogFlow(("pgmPoolTrackFlushGCPhysPTsSlow: giving up... (cPresent=%d)\n", pPool->cPresent));
3691	return VINF_PGM_GCPHYS_ALIASED;
3692	}
3693
3694	return VINF_SUCCESS;
3695	}
3696
3697
3698	/**
3699	* Clears the user entry in a user table.
3700	*
3701	* This is used to remove all references to a page when flushing it.
3702	*/
3703	static void pgmPoolTrackClearPageUser(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PCPGMPOOLUSER pUser)
3704	{
3705	Assert(pUser->iUser != NIL_PGMPOOL_IDX);
3706	Assert(pUser->iUser < pPool->cCurPages);
3707	uint32_t iUserTable = pUser->iUserTable;
3708
3709	/*
3710	* Map the user page.
3711	*/
3712	PPGMPOOLPAGE pUserPage = &pPool->aPages[pUser->iUser];
3713	union
3714	{
3715	uint64_t *pau64;
3716	uint32_t *pau32;
3717	} u;
3718	u.pau64 = (uint64_t *)PGMPOOL_PAGE_2_PTR(pPool->CTX_SUFF(pVM), pUserPage);
3719
3720	LogFlow(("pgmPoolTrackClearPageUser: clear %x in %s (%RGp) (flushing %s)\n", iUserTable, pgmPoolPoolKindToStr(pUserPage->enmKind), pUserPage->Core.Key, pgmPoolPoolKindToStr(pPage->enmKind)));
3721
3722	/* Safety precaution in case we change the paging for other modes too in the future. */
3723	Assert(!pgmPoolIsPageLocked(pPage));
3724
3725	#ifdef VBOX_STRICT
3726	/*
3727	* Some sanity checks.
3728	*/
3729	switch (pUserPage->enmKind)
3730	{
3731	case PGMPOOLKIND_32BIT_PD:
3732	case PGMPOOLKIND_32BIT_PD_PHYS:
3733	Assert(iUserTable < X86_PG_ENTRIES);
3734	break;
3735	case PGMPOOLKIND_PAE_PDPT:
3736	case PGMPOOLKIND_PAE_PDPT_FOR_32BIT:
3737	case PGMPOOLKIND_PAE_PDPT_PHYS:
3738	Assert(iUserTable < 4);
3739	Assert(!(u.pau64[iUserTable] & PGM_PLXFLAGS_PERMANENT));
3740	break;
3741	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
3742	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
3743	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
3744	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
3745	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
3746	case PGMPOOLKIND_PAE_PD_PHYS:
3747	Assert(iUserTable < X86_PG_PAE_ENTRIES);
3748	break;
3749	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
3750	Assert(iUserTable < X86_PG_PAE_ENTRIES);
3751	Assert(!(u.pau64[iUserTable] & PGM_PDFLAGS_MAPPING));
3752	break;
3753	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
3754	Assert(iUserTable < X86_PG_PAE_ENTRIES);
3755	Assert(!(u.pau64[iUserTable] & PGM_PLXFLAGS_PERMANENT));
3756	break;
3757	case PGMPOOLKIND_64BIT_PML4:
3758	Assert(!(u.pau64[iUserTable] & PGM_PLXFLAGS_PERMANENT));
3759	/* GCPhys >> PAGE_SHIFT is the index here */
3760	break;
3761	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
3762	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
3763	Assert(iUserTable < X86_PG_PAE_ENTRIES);
3764	break;
3765
3766	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
3767	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
3768	Assert(iUserTable < X86_PG_PAE_ENTRIES);
3769	break;
3770
3771	case PGMPOOLKIND_ROOT_NESTED:
3772	Assert(iUserTable < X86_PG_PAE_ENTRIES);
3773	break;
3774
3775	default:
3776	AssertMsgFailed(("enmKind=%d\n", pUserPage->enmKind));
3777	break;
3778	}
3779	#endif /* VBOX_STRICT */
3780
3781	/*
3782	* Clear the entry in the user page.
3783	*/
3784	switch (pUserPage->enmKind)
3785	{
3786	/* 32-bit entries */
3787	case PGMPOOLKIND_32BIT_PD:
3788	case PGMPOOLKIND_32BIT_PD_PHYS:
3789	ASMAtomicWriteU32(&u.pau32[iUserTable], 0);
3790	break;
3791
3792	/* 64-bit entries */
3793	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
3794	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
3795	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
3796	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
3797	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
3798	#ifdef IN_RC
3799	/*
3800	* In 32 bits PAE mode we must invalidate the TLB when changing a
3801	* PDPT entry; the CPU fetches them only during cr3 load, so any
3802	* non-present PDPT will continue to cause page faults.
3803	*/
3804	ASMReloadCR3();
3805	/* no break */
3806	#endif
3807	case PGMPOOLKIND_PAE_PD_PHYS:
3808	case PGMPOOLKIND_PAE_PDPT_PHYS:
3809	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
3810	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
3811	case PGMPOOLKIND_64BIT_PML4:
3812	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
3813	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
3814	case PGMPOOLKIND_PAE_PDPT:
3815	case PGMPOOLKIND_PAE_PDPT_FOR_32BIT:
3816	case PGMPOOLKIND_ROOT_NESTED:
3817	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
3818	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
3819	ASMAtomicWriteU64(&u.pau64[iUserTable], 0);
3820	break;
3821
3822	default:
3823	AssertFatalMsgFailed(("enmKind=%d iUser=%d iUserTable=%#x\n", pUserPage->enmKind, pUser->iUser, pUser->iUserTable));
3824	}
3825	PGM_DYNMAP_UNUSED_HINT_VM(pPool->CTX_SUFF(pVM), u.pau64);
3826	}
3827
3828
3829	/**
3830	* Clears all users of a page.
3831	*/
3832	static void pgmPoolTrackClearPageUsers(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
3833	{
3834	/*
3835	* Free all the user records.
3836	*/
3837	LogFlow(("pgmPoolTrackClearPageUsers %RGp\n", pPage->GCPhys));
3838
3839	PPGMPOOLUSER paUsers = pPool->CTX_SUFF(paUsers);
3840	uint16_t i = pPage->iUserHead;
3841	while (i != NIL_PGMPOOL_USER_INDEX)
3842	{
3843	/* Clear enter in user table. */
3844	pgmPoolTrackClearPageUser(pPool, pPage, &paUsers[i]);
3845
3846	/* Free it. */
3847	const uint16_t iNext = paUsers[i].iNext;
3848	paUsers[i].iUser = NIL_PGMPOOL_IDX;
3849	paUsers[i].iNext = pPool->iUserFreeHead;
3850	pPool->iUserFreeHead = i;
3851
3852	/* Next. */
3853	i = iNext;
3854	}
3855	pPage->iUserHead = NIL_PGMPOOL_USER_INDEX;
3856	}
3857
3858
3859	/**
3860	* Allocates a new physical cross reference extent.
3861	*
3862	* @returns Pointer to the allocated extent on success. NULL if we're out of them.
3863	* @param pVM Pointer to the VM.
3864	* @param piPhysExt Where to store the phys ext index.
3865	*/
3866	PPGMPOOLPHYSEXT pgmPoolTrackPhysExtAlloc(PVM pVM, uint16_t *piPhysExt)
3867	{
3868	PGM_LOCK_ASSERT_OWNER(pVM);
3869	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3870	uint16_t iPhysExt = pPool->iPhysExtFreeHead;
3871	if (iPhysExt == NIL_PGMPOOL_PHYSEXT_INDEX)
3872	{
3873	STAM_COUNTER_INC(&pPool->StamTrackPhysExtAllocFailures);
3874	return NULL;
3875	}
3876	PPGMPOOLPHYSEXT pPhysExt = &pPool->CTX_SUFF(paPhysExts)[iPhysExt];
3877	pPool->iPhysExtFreeHead = pPhysExt->iNext;
3878	pPhysExt->iNext = NIL_PGMPOOL_PHYSEXT_INDEX;
3879	*piPhysExt = iPhysExt;
3880	return pPhysExt;
3881	}
3882
3883
3884	/**
3885	* Frees a physical cross reference extent.
3886	*
3887	* @param pVM Pointer to the VM.
3888	* @param iPhysExt The extent to free.
3889	*/
3890	void pgmPoolTrackPhysExtFree(PVM pVM, uint16_t iPhysExt)
3891	{
3892	PGM_LOCK_ASSERT_OWNER(pVM);
3893	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3894	Assert(iPhysExt < pPool->cMaxPhysExts);
3895	PPGMPOOLPHYSEXT pPhysExt = &pPool->CTX_SUFF(paPhysExts)[iPhysExt];
3896	for (unsigned i = 0; i < RT_ELEMENTS(pPhysExt->aidx); i++)
3897	{
3898	pPhysExt->aidx[i] = NIL_PGMPOOL_IDX;
3899	pPhysExt->apte[i] = NIL_PGMPOOL_PHYSEXT_IDX_PTE;
3900	}
3901	pPhysExt->iNext = pPool->iPhysExtFreeHead;
3902	pPool->iPhysExtFreeHead = iPhysExt;
3903	}
3904
3905
3906	/**
3907	* Frees a physical cross reference extent.
3908	*
3909	* @param pVM Pointer to the VM.
3910	* @param iPhysExt The extent to free.
3911	*/
3912	void pgmPoolTrackPhysExtFreeList(PVM pVM, uint16_t iPhysExt)
3913	{
3914	PGM_LOCK_ASSERT_OWNER(pVM);
3915	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3916
3917	const uint16_t iPhysExtStart = iPhysExt;
3918	PPGMPOOLPHYSEXT pPhysExt;
3919	do
3920	{
3921	Assert(iPhysExt < pPool->cMaxPhysExts);
3922	pPhysExt = &pPool->CTX_SUFF(paPhysExts)[iPhysExt];
3923	for (unsigned i = 0; i < RT_ELEMENTS(pPhysExt->aidx); i++)
3924	{
3925	pPhysExt->aidx[i] = NIL_PGMPOOL_IDX;
3926	pPhysExt->apte[i] = NIL_PGMPOOL_PHYSEXT_IDX_PTE;
3927	}
3928
3929	/* next */
3930	iPhysExt = pPhysExt->iNext;
3931	} while (iPhysExt != NIL_PGMPOOL_PHYSEXT_INDEX);
3932
3933	pPhysExt->iNext = pPool->iPhysExtFreeHead;
3934	pPool->iPhysExtFreeHead = iPhysExtStart;
3935	}
3936
3937
3938	/**
3939	* Insert a reference into a list of physical cross reference extents.
3940	*
3941	* @returns The new tracking data for PGMPAGE.
3942	*
3943	* @param pVM Pointer to the VM.
3944	* @param iPhysExt The physical extent index of the list head.
3945	* @param iShwPT The shadow page table index.
3946	* @param iPte Page table entry
3947	*
3948	*/
3949	static uint16_t pgmPoolTrackPhysExtInsert(PVM pVM, uint16_t iPhysExt, uint16_t iShwPT, uint16_t iPte)
3950	{
3951	PGM_LOCK_ASSERT_OWNER(pVM);
3952	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
3953	PPGMPOOLPHYSEXT paPhysExts = pPool->CTX_SUFF(paPhysExts);
3954
3955	/*
3956	* Special common cases.
3957	*/
3958	if (paPhysExts[iPhysExt].aidx[1] == NIL_PGMPOOL_IDX)
3959	{
3960	paPhysExts[iPhysExt].aidx[1] = iShwPT;
3961	paPhysExts[iPhysExt].apte[1] = iPte;
3962	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatTrackAliasedMany);
3963	LogFlow(("pgmPoolTrackPhysExtInsert: %d:{,%d pte %d,}\n", iPhysExt, iShwPT, iPte));
3964	return PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, iPhysExt);
3965	}
3966	if (paPhysExts[iPhysExt].aidx[2] == NIL_PGMPOOL_IDX)
3967	{
3968	paPhysExts[iPhysExt].aidx[2] = iShwPT;
3969	paPhysExts[iPhysExt].apte[2] = iPte;
3970	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatTrackAliasedMany);
3971	LogFlow(("pgmPoolTrackPhysExtInsert: %d:{,,%d pte %d}\n", iPhysExt, iShwPT, iPte));
3972	return PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, iPhysExt);
3973	}
3974	AssertCompile(RT_ELEMENTS(paPhysExts[iPhysExt].aidx) == 3);
3975
3976	/*
3977	* General treatment.
3978	*/
3979	const uint16_t iPhysExtStart = iPhysExt;
3980	unsigned cMax = 15;
3981	for (;;)
3982	{
3983	Assert(iPhysExt < pPool->cMaxPhysExts);
3984	for (unsigned i = 0; i < RT_ELEMENTS(paPhysExts[iPhysExt].aidx); i++)
3985	if (paPhysExts[iPhysExt].aidx[i] == NIL_PGMPOOL_IDX)
3986	{
3987	paPhysExts[iPhysExt].aidx[i] = iShwPT;
3988	paPhysExts[iPhysExt].apte[i] = iPte;
3989	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatTrackAliasedMany);
3990	LogFlow(("pgmPoolTrackPhysExtInsert: %d:{%d pte %d} i=%d cMax=%d\n", iPhysExt, iShwPT, iPte, i, cMax));
3991	return PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, iPhysExtStart);
3992	}
3993	if (!--cMax)
3994	{
3995	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatTrackOverflows);
3996	pgmPoolTrackPhysExtFreeList(pVM, iPhysExtStart);
3997	LogFlow(("pgmPoolTrackPhysExtInsert: overflow (1) iShwPT=%d\n", iShwPT));
3998	return PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, PGMPOOL_TD_IDX_OVERFLOWED);
3999	}
4000
4001	/* advance */
4002	iPhysExt = paPhysExts[iPhysExt].iNext;
4003	if (iPhysExt == NIL_PGMPOOL_PHYSEXT_INDEX)
4004	break;
4005	}
4006
4007	/*
4008	* Add another extent to the list.
4009	*/
4010	PPGMPOOLPHYSEXT pNew = pgmPoolTrackPhysExtAlloc(pVM, &iPhysExt);
4011	if (!pNew)
4012	{
4013	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatTrackNoExtentsLeft);
4014	pgmPoolTrackPhysExtFreeList(pVM, iPhysExtStart);
4015	LogFlow(("pgmPoolTrackPhysExtInsert: pgmPoolTrackPhysExtAlloc failed iShwPT=%d\n", iShwPT));
4016	return PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, PGMPOOL_TD_IDX_OVERFLOWED);
4017	}
4018	pNew->iNext = iPhysExtStart;
4019	pNew->aidx[0] = iShwPT;
4020	pNew->apte[0] = iPte;
4021	LogFlow(("pgmPoolTrackPhysExtInsert: added new extent %d:{%d pte %d}->%d\n", iPhysExt, iShwPT, iPte, iPhysExtStart));
4022	return PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, iPhysExt);
4023	}
4024
4025
4026	/**
4027	* Add a reference to guest physical page where extents are in use.
4028	*
4029	* @returns The new tracking data for PGMPAGE.
4030	*
4031	* @param pVM Pointer to the VM.
4032	* @param pPhysPage Pointer to the aPages entry in the ram range.
4033	* @param u16 The ram range flags (top 16-bits).
4034	* @param iShwPT The shadow page table index.
4035	* @param iPte Page table entry
4036	*/
4037	uint16_t pgmPoolTrackPhysExtAddref(PVM pVM, PPGMPAGE pPhysPage, uint16_t u16, uint16_t iShwPT, uint16_t iPte)
4038	{
4039	pgmLock(pVM);
4040	if (PGMPOOL_TD_GET_CREFS(u16) != PGMPOOL_TD_CREFS_PHYSEXT)
4041	{
4042	/*
4043	* Convert to extent list.
4044	*/
4045	Assert(PGMPOOL_TD_GET_CREFS(u16) == 1);
4046	uint16_t iPhysExt;
4047	PPGMPOOLPHYSEXT pPhysExt = pgmPoolTrackPhysExtAlloc(pVM, &iPhysExt);
4048	if (pPhysExt)
4049	{
4050	LogFlow(("pgmPoolTrackPhysExtAddref: new extent: %d:{%d, %d}\n", iPhysExt, PGMPOOL_TD_GET_IDX(u16), iShwPT));
4051	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatTrackAliased);
4052	pPhysExt->aidx[0] = PGMPOOL_TD_GET_IDX(u16);
4053	pPhysExt->apte[0] = PGM_PAGE_GET_PTE_INDEX(pPhysPage);
4054	pPhysExt->aidx[1] = iShwPT;
4055	pPhysExt->apte[1] = iPte;
4056	u16 = PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, iPhysExt);
4057	}
4058	else
4059	u16 = PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, PGMPOOL_TD_IDX_OVERFLOWED);
4060	}
4061	else if (u16 != PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, PGMPOOL_TD_IDX_OVERFLOWED))
4062	{
4063	/*
4064	* Insert into the extent list.
4065	*/
4066	u16 = pgmPoolTrackPhysExtInsert(pVM, PGMPOOL_TD_GET_IDX(u16), iShwPT, iPte);
4067	}
4068	else
4069	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatTrackAliasedLots);
4070	pgmUnlock(pVM);
4071	return u16;
4072	}
4073
4074
4075	/**
4076	* Clear references to guest physical memory.
4077	*
4078	* @param pPool The pool.
4079	* @param pPage The page.
4080	* @param pPhysPage Pointer to the aPages entry in the ram range.
4081	* @param iPte Shadow PTE index
4082	*/
4083	void pgmPoolTrackPhysExtDerefGCPhys(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMPAGE pPhysPage, uint16_t iPte)
4084	{
4085	PVM pVM = pPool->CTX_SUFF(pVM);
4086	const unsigned cRefs = PGM_PAGE_GET_TD_CREFS(pPhysPage);
4087	AssertFatalMsg(cRefs == PGMPOOL_TD_CREFS_PHYSEXT, ("cRefs=%d pPhysPage=%R[pgmpage] pPage=%p:{.idx=%d}\n", cRefs, pPhysPage, pPage, pPage->idx));
4088
4089	uint16_t iPhysExt = PGM_PAGE_GET_TD_IDX(pPhysPage);
4090	if (iPhysExt != PGMPOOL_TD_IDX_OVERFLOWED)
4091	{
4092	pgmLock(pVM);
4093
4094	uint16_t iPhysExtPrev = NIL_PGMPOOL_PHYSEXT_INDEX;
4095	PPGMPOOLPHYSEXT paPhysExts = pPool->CTX_SUFF(paPhysExts);
4096	do
4097	{
4098	Assert(iPhysExt < pPool->cMaxPhysExts);
4099
4100	/*
4101	* Look for the shadow page and check if it's all freed.
4102	*/
4103	for (unsigned i = 0; i < RT_ELEMENTS(paPhysExts[iPhysExt].aidx); i++)
4104	{
4105	if ( paPhysExts[iPhysExt].aidx[i] == pPage->idx
4106	&& paPhysExts[iPhysExt].apte[i] == iPte)
4107	{
4108	paPhysExts[iPhysExt].aidx[i] = NIL_PGMPOOL_IDX;
4109	paPhysExts[iPhysExt].apte[i] = NIL_PGMPOOL_PHYSEXT_IDX_PTE;
4110
4111	for (i = 0; i < RT_ELEMENTS(paPhysExts[iPhysExt].aidx); i++)
4112	if (paPhysExts[iPhysExt].aidx[i] != NIL_PGMPOOL_IDX)
4113	{
4114	Log2(("pgmPoolTrackPhysExtDerefGCPhys: pPhysPage=%R[pgmpage] idx=%d\n", pPhysPage, pPage->idx));
4115	pgmUnlock(pVM);
4116	return;
4117	}
4118
4119	/* we can free the node. */
4120	const uint16_t iPhysExtNext = paPhysExts[iPhysExt].iNext;
4121	if ( iPhysExtPrev == NIL_PGMPOOL_PHYSEXT_INDEX
4122	&& iPhysExtNext == NIL_PGMPOOL_PHYSEXT_INDEX)
4123	{
4124	/* lonely node */
4125	pgmPoolTrackPhysExtFree(pVM, iPhysExt);
4126	Log2(("pgmPoolTrackPhysExtDerefGCPhys: pPhysPage=%R[pgmpage] idx=%d lonely\n", pPhysPage, pPage->idx));
4127	PGM_PAGE_SET_TRACKING(pVM, pPhysPage, 0);
4128	}
4129	else if (iPhysExtPrev == NIL_PGMPOOL_PHYSEXT_INDEX)
4130	{
4131	/* head */
4132	Log2(("pgmPoolTrackPhysExtDerefGCPhys: pPhysPage=%R[pgmpage] idx=%d head\n", pPhysPage, pPage->idx));
4133	PGM_PAGE_SET_TRACKING(pVM, pPhysPage, PGMPOOL_TD_MAKE(PGMPOOL_TD_CREFS_PHYSEXT, iPhysExtNext));
4134	pgmPoolTrackPhysExtFree(pVM, iPhysExt);
4135	}
4136	else
4137	{
4138	/* in list */
4139	Log2(("pgmPoolTrackPhysExtDerefGCPhys: pPhysPage=%R[pgmpage] idx=%d in list\n", pPhysPage, pPage->idx));
4140	paPhysExts[iPhysExtPrev].iNext = iPhysExtNext;
4141	pgmPoolTrackPhysExtFree(pVM, iPhysExt);
4142	}
4143	iPhysExt = iPhysExtNext;
4144	pgmUnlock(pVM);
4145	return;
4146	}
4147	}
4148
4149	/* next */
4150	iPhysExtPrev = iPhysExt;
4151	iPhysExt = paPhysExts[iPhysExt].iNext;
4152	} while (iPhysExt != NIL_PGMPOOL_PHYSEXT_INDEX);
4153
4154	pgmUnlock(pVM);
4155	AssertFatalMsgFailed(("not-found! cRefs=%d pPhysPage=%R[pgmpage] pPage=%p:{.idx=%d}\n", cRefs, pPhysPage, pPage, pPage->idx));
4156	}
4157	else /* nothing to do */
4158	Log2(("pgmPoolTrackPhysExtDerefGCPhys: pPhysPage=%R[pgmpage]\n", pPhysPage));
4159	}
4160
4161	/**
4162	* Clear references to guest physical memory.
4163	*
4164	* This is the same as pgmPoolTracDerefGCPhysHint except that the guest
4165	* physical address is assumed to be correct, so the linear search can be
4166	* skipped and we can assert at an earlier point.
4167	*
4168	* @param pPool The pool.
4169	* @param pPage The page.
4170	* @param HCPhys The host physical address corresponding to the guest page.
4171	* @param GCPhys The guest physical address corresponding to HCPhys.
4172	* @param iPte Shadow PTE index
4173	*/
4174	static void pgmPoolTracDerefGCPhys(PPGMPOOL pPool, PPGMPOOLPAGE pPage, RTHCPHYS HCPhys, RTGCPHYS GCPhys, uint16_t iPte)
4175	{
4176	/*
4177	* Lookup the page and check if it checks out before derefing it.
4178	*/
4179	PVM pVM = pPool->CTX_SUFF(pVM);
4180	PPGMPAGE pPhysPage = pgmPhysGetPage(pVM, GCPhys);
4181	if (pPhysPage)
4182	{
4183	Assert(PGM_PAGE_GET_HCPHYS(pPhysPage));
4184	#ifdef LOG_ENABLED
4185	RTHCPHYS HCPhysPage = PGM_PAGE_GET_HCPHYS(pPhysPage);
4186	Log2(("pgmPoolTracDerefGCPhys %RHp vs %RHp\n", HCPhysPage, HCPhys));
4187	#endif
4188	if (PGM_PAGE_GET_HCPHYS(pPhysPage) == HCPhys)
4189	{
4190	Assert(pPage->cPresent);
4191	Assert(pPool->cPresent);
4192	pPage->cPresent--;
4193	pPool->cPresent--;
4194	pgmTrackDerefGCPhys(pPool, pPage, pPhysPage, iPte);
4195	return;
4196	}
4197
4198	AssertFatalMsgFailed(("HCPhys=%RHp GCPhys=%RGp; found page has HCPhys=%RHp\n",
4199	HCPhys, GCPhys, PGM_PAGE_GET_HCPHYS(pPhysPage)));
4200	}
4201	AssertFatalMsgFailed(("HCPhys=%RHp GCPhys=%RGp\n", HCPhys, GCPhys));
4202	}
4203
4204
4205	/**
4206	* Clear references to guest physical memory.
4207	*
4208	* @param pPool The pool.
4209	* @param pPage The page.
4210	* @param HCPhys The host physical address corresponding to the guest page.
4211	* @param GCPhysHint The guest physical address which may corresponding to HCPhys.
4212	* @param iPte Shadow pte index
4213	*/
4214	void pgmPoolTracDerefGCPhysHint(PPGMPOOL pPool, PPGMPOOLPAGE pPage, RTHCPHYS HCPhys, RTGCPHYS GCPhysHint, uint16_t iPte)
4215	{
4216	Log4(("pgmPoolTracDerefGCPhysHint %RHp %RGp\n", HCPhys, GCPhysHint));
4217
4218	/*
4219	* Try the hint first.
4220	*/
4221	RTHCPHYS HCPhysHinted;
4222	PVM pVM = pPool->CTX_SUFF(pVM);
4223	PPGMPAGE pPhysPage = pgmPhysGetPage(pVM, GCPhysHint);
4224	if (pPhysPage)
4225	{
4226	HCPhysHinted = PGM_PAGE_GET_HCPHYS(pPhysPage);
4227	Assert(HCPhysHinted);
4228	if (HCPhysHinted == HCPhys)
4229	{
4230	Assert(pPage->cPresent);
4231	Assert(pPool->cPresent);
4232	pPage->cPresent--;
4233	pPool->cPresent--;
4234	pgmTrackDerefGCPhys(pPool, pPage, pPhysPage, iPte);
4235	return;
4236	}
4237	}
4238	else
4239	HCPhysHinted = UINT64_C(0xdeadbeefdeadbeef);
4240
4241	/*
4242	* Damn, the hint didn't work. We'll have to do an expensive linear search.
4243	*/
4244	STAM_COUNTER_INC(&pPool->StatTrackLinearRamSearches);
4245	PPGMRAMRANGE pRam = pPool->CTX_SUFF(pVM)->pgm.s.CTX_SUFF(pRamRangesX);
4246	while (pRam)
4247	{
4248	unsigned iPage = pRam->cb >> PAGE_SHIFT;
4249	while (iPage-- > 0)
4250	{
4251	if (PGM_PAGE_GET_HCPHYS(&pRam->aPages[iPage]) == HCPhys)
4252	{
4253	Log4(("pgmPoolTracDerefGCPhysHint: Linear HCPhys=%RHp GCPhysHint=%RGp GCPhysReal=%RGp\n",
4254	HCPhys, GCPhysHint, pRam->GCPhys + (iPage << PAGE_SHIFT)));
4255	Assert(pPage->cPresent);
4256	Assert(pPool->cPresent);
4257	pPage->cPresent--;
4258	pPool->cPresent--;
4259	pgmTrackDerefGCPhys(pPool, pPage, &pRam->aPages[iPage], iPte);
4260	return;
4261	}
4262	}
4263	pRam = pRam->CTX_SUFF(pNext);
4264	}
4265
4266	AssertFatalMsgFailed(("HCPhys=%RHp GCPhysHint=%RGp (Hinted page has HCPhys = %RHp)\n", HCPhys, GCPhysHint, HCPhysHinted));
4267	}
4268
4269
4270	/**
4271	* Clear references to guest physical memory in a 32-bit / 32-bit page table.
4272	*
4273	* @param pPool The pool.
4274	* @param pPage The page.
4275	* @param pShwPT The shadow page table (mapping of the page).
4276	* @param pGstPT The guest page table.
4277	*/
4278	DECLINLINE(void) pgmPoolTrackDerefPT32Bit32Bit(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PX86PT pShwPT, PCX86PT pGstPT)
4279	{
4280	RTGCPHYS32 const fPgMask = pPage->fA20Enabled ? X86_PTE_PG_MASK : X86_PTE_PG_MASK & ~RT_BIT_32(20);
4281	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++)
4282	{
4283	Assert(!(pShwPT->a[i].u & RT_BIT_32(10)));
4284	if (pShwPT->a[i].n.u1Present)
4285	{
4286	Log4(("pgmPoolTrackDerefPT32Bit32Bit: i=%d pte=%RX32 hint=%RX32\n",
4287	i, pShwPT->a[i].u & X86_PTE_PG_MASK, pGstPT->a[i].u & X86_PTE_PG_MASK));
4288	pgmPoolTracDerefGCPhysHint(pPool, pPage, pShwPT->a[i].u & X86_PTE_PG_MASK, pGstPT->a[i].u & fPgMask, i);
4289	if (!pPage->cPresent)
4290	break;
4291	}
4292	}
4293	}
4294
4295
4296	/**
4297	* Clear references to guest physical memory in a PAE / 32-bit page table.
4298	*
4299	* @param pPool The pool.
4300	* @param pPage The page.
4301	* @param pShwPT The shadow page table (mapping of the page).
4302	* @param pGstPT The guest page table (just a half one).
4303	*/
4304	DECLINLINE(void) pgmPoolTrackDerefPTPae32Bit(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMSHWPTPAE pShwPT, PCX86PT pGstPT)
4305	{
4306	RTGCPHYS32 const fPgMask = pPage->fA20Enabled ? X86_PTE_PG_MASK : X86_PTE_PG_MASK & ~RT_BIT_32(20);
4307	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++)
4308	{
4309	Assert( (PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & UINT64_C(0x7ff0000000000400)) == 0
4310	\|\| (PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & UINT64_C(0x7ff0000000000400)) == UINT64_C(0x7ff0000000000000));
4311	if (PGMSHWPTEPAE_IS_P(pShwPT->a[i]))
4312	{
4313	Log4(("pgmPoolTrackDerefPTPae32Bit: i=%d pte=%RX64 hint=%RX32\n",
4314	i, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), pGstPT->a[i].u & X86_PTE_PG_MASK));
4315	pgmPoolTracDerefGCPhysHint(pPool, pPage, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), pGstPT->a[i].u & fPgMask, i);
4316	if (!pPage->cPresent)
4317	break;
4318	}
4319	}
4320	}
4321
4322
4323	/**
4324	* Clear references to guest physical memory in a PAE / PAE page table.
4325	*
4326	* @param pPool The pool.
4327	* @param pPage The page.
4328	* @param pShwPT The shadow page table (mapping of the page).
4329	* @param pGstPT The guest page table.
4330	*/
4331	DECLINLINE(void) pgmPoolTrackDerefPTPaePae(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMSHWPTPAE pShwPT, PCX86PTPAE pGstPT)
4332	{
4333	RTGCPHYS const fPgMask = pPage->fA20Enabled ? X86_PTE_PAE_PG_MASK : X86_PTE_PAE_PG_MASK & ~RT_BIT_64(20);
4334	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++)
4335	{
4336	Assert( (PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & UINT64_C(0x7ff0000000000400)) == 0
4337	\|\| (PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & UINT64_C(0x7ff0000000000400)) == UINT64_C(0x7ff0000000000000));
4338	if (PGMSHWPTEPAE_IS_P(pShwPT->a[i]))
4339	{
4340	Log4(("pgmPoolTrackDerefPTPaePae: i=%d pte=%RX32 hint=%RX32\n",
4341	i, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), pGstPT->a[i].u & X86_PTE_PAE_PG_MASK));
4342	pgmPoolTracDerefGCPhysHint(pPool, pPage, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), pGstPT->a[i].u & fPgMask, i);
4343	if (!pPage->cPresent)
4344	break;
4345	}
4346	}
4347	}
4348
4349
4350	/**
4351	* Clear references to guest physical memory in a 32-bit / 4MB page table.
4352	*
4353	* @param pPool The pool.
4354	* @param pPage The page.
4355	* @param pShwPT The shadow page table (mapping of the page).
4356	*/
4357	DECLINLINE(void) pgmPoolTrackDerefPT32Bit4MB(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PX86PT pShwPT)
4358	{
4359	RTGCPHYS const GCPhysA20Mask = pPage->fA20Enabled ? UINT64_MAX : ~RT_BIT_64(20);
4360	RTGCPHYS GCPhys = pPage->GCPhys + PAGE_SIZE * pPage->iFirstPresent;
4361	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++, GCPhys += PAGE_SIZE)
4362	{
4363	Assert(!(pShwPT->a[i].u & RT_BIT_32(10)));
4364	if (pShwPT->a[i].n.u1Present)
4365	{
4366	Log4(("pgmPoolTrackDerefPT32Bit4MB: i=%d pte=%RX32 GCPhys=%RGp\n",
4367	i, pShwPT->a[i].u & X86_PTE_PG_MASK, GCPhys));
4368	pgmPoolTracDerefGCPhys(pPool, pPage, pShwPT->a[i].u & X86_PTE_PG_MASK, GCPhys & GCPhysA20Mask, i);
4369	if (!pPage->cPresent)
4370	break;
4371	}
4372	}
4373	}
4374
4375
4376	/**
4377	* Clear references to guest physical memory in a PAE / 2/4MB page table.
4378	*
4379	* @param pPool The pool.
4380	* @param pPage The page.
4381	* @param pShwPT The shadow page table (mapping of the page).
4382	*/
4383	DECLINLINE(void) pgmPoolTrackDerefPTPaeBig(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PPGMSHWPTPAE pShwPT)
4384	{
4385	RTGCPHYS const GCPhysA20Mask = pPage->fA20Enabled ? UINT64_MAX : ~RT_BIT_64(20);
4386	RTGCPHYS GCPhys = pPage->GCPhys + PAGE_SIZE * pPage->iFirstPresent;
4387	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++, GCPhys += PAGE_SIZE)
4388	{
4389	Assert( (PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & UINT64_C(0x7ff0000000000400)) == 0
4390	\|\| (PGMSHWPTEPAE_GET_U(pShwPT->a[i]) & UINT64_C(0x7ff0000000000400)) == UINT64_C(0x7ff0000000000000));
4391	if (PGMSHWPTEPAE_IS_P(pShwPT->a[i]))
4392	{
4393	Log4(("pgmPoolTrackDerefPTPaeBig: i=%d pte=%RX64 hint=%RGp\n",
4394	i, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), GCPhys));
4395	pgmPoolTracDerefGCPhys(pPool, pPage, PGMSHWPTEPAE_GET_HCPHYS(pShwPT->a[i]), GCPhys & GCPhysA20Mask, i);
4396	if (!pPage->cPresent)
4397	break;
4398	}
4399	}
4400	}
4401
4402
4403	/**
4404	* Clear references to shadowed pages in an EPT page table.
4405	*
4406	* @param pPool The pool.
4407	* @param pPage The page.
4408	* @param pShwPML4 The shadow page directory pointer table (mapping of the page).
4409	*/
4410	DECLINLINE(void) pgmPoolTrackDerefPTEPT(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PEPTPT pShwPT)
4411	{
4412	RTGCPHYS const GCPhysA20Mask = pPage->fA20Enabled ? UINT64_MAX : ~RT_BIT_64(20);
4413	RTGCPHYS GCPhys = pPage->GCPhys + PAGE_SIZE * pPage->iFirstPresent;
4414	for (unsigned i = pPage->iFirstPresent; i < RT_ELEMENTS(pShwPT->a); i++, GCPhys += PAGE_SIZE)
4415	{
4416	Assert((pShwPT->a[i].u & UINT64_C(0xfff0000000000f80)) == 0);
4417	if (pShwPT->a[i].n.u1Present)
4418	{
4419	Log4(("pgmPoolTrackDerefPTEPT: i=%d pte=%RX64 GCPhys=%RX64\n",
4420	i, pShwPT->a[i].u & EPT_PTE_PG_MASK, pPage->GCPhys));
4421	pgmPoolTracDerefGCPhys(pPool, pPage, pShwPT->a[i].u & EPT_PTE_PG_MASK, GCPhys & GCPhysA20Mask, i);
4422	if (!pPage->cPresent)
4423	break;
4424	}
4425	}
4426	}
4427
4428
4429	/**
4430	* Clear references to shadowed pages in a 32 bits page directory.
4431	*
4432	* @param pPool The pool.
4433	* @param pPage The page.
4434	* @param pShwPD The shadow page directory (mapping of the page).
4435	*/
4436	DECLINLINE(void) pgmPoolTrackDerefPD(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PX86PD pShwPD)
4437	{
4438	for (unsigned i = 0; i < RT_ELEMENTS(pShwPD->a); i++)
4439	{
4440	Assert(!(pShwPD->a[i].u & RT_BIT_32(9)));
4441	if ( pShwPD->a[i].n.u1Present
4442	&& !(pShwPD->a[i].u & PGM_PDFLAGS_MAPPING)
4443	)
4444	{
4445	PPGMPOOLPAGE pSubPage = (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, pShwPD->a[i].u & X86_PDE_PG_MASK);
4446	if (pSubPage)
4447	pgmPoolTrackFreeUser(pPool, pSubPage, pPage->idx, i);
4448	else
4449	AssertFatalMsgFailed(("%x\n", pShwPD->a[i].u & X86_PDE_PG_MASK));
4450	}
4451	}
4452	}
4453
4454
4455	/**
4456	* Clear references to shadowed pages in a PAE (legacy or 64 bits) page directory.
4457	*
4458	* @param pPool The pool.
4459	* @param pPage The page.
4460	* @param pShwPD The shadow page directory (mapping of the page).
4461	*/
4462	DECLINLINE(void) pgmPoolTrackDerefPDPae(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PX86PDPAE pShwPD)
4463	{
4464	for (unsigned i = 0; i < RT_ELEMENTS(pShwPD->a); i++)
4465	{
4466	if ( pShwPD->a[i].n.u1Present
4467	&& !(pShwPD->a[i].u & PGM_PDFLAGS_MAPPING))
4468	{
4469	#ifdef PGM_WITH_LARGE_PAGES
4470	if (pShwPD->a[i].b.u1Size)
4471	{
4472	Log4(("pgmPoolTrackDerefPDPae: i=%d pde=%RX64 GCPhys=%RX64\n",
4473	i, pShwPD->a[i].u & X86_PDE2M_PAE_PG_MASK, pPage->GCPhys));
4474	pgmPoolTracDerefGCPhys(pPool, pPage, pShwPD->a[i].u & X86_PDE2M_PAE_PG_MASK,
4475	pPage->GCPhys + i * 2 * _1M /* pPage->GCPhys = base address of the memory described by the PD */,
4476	i);
4477	}
4478	else
4479	#endif
4480	{
4481	Assert((pShwPD->a[i].u & (X86_PDE_PAE_MBZ_MASK_NX \| UINT64_C(0x7ff0000000000200))) == 0);
4482	PPGMPOOLPAGE pSubPage = (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, pShwPD->a[i].u & X86_PDE_PAE_PG_MASK);
4483	if (pSubPage)
4484	pgmPoolTrackFreeUser(pPool, pSubPage, pPage->idx, i);
4485	else
4486	AssertFatalMsgFailed(("%RX64\n", pShwPD->a[i].u & X86_PDE_PAE_PG_MASK));
4487	/** @todo 64-bit guests: have to ensure that we're not exhausting the dynamic mappings! */
4488	}
4489	}
4490	}
4491	}
4492
4493
4494	/**
4495	* Clear references to shadowed pages in a PAE page directory pointer table.
4496	*
4497	* @param pPool The pool.
4498	* @param pPage The page.
4499	* @param pShwPDPT The shadow page directory pointer table (mapping of the page).
4500	*/
4501	DECLINLINE(void) pgmPoolTrackDerefPDPTPae(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PX86PDPT pShwPDPT)
4502	{
4503	for (unsigned i = 0; i < X86_PG_PAE_PDPE_ENTRIES; i++)
4504	{
4505	Assert((pShwPDPT->a[i].u & (X86_PDPE_PAE_MBZ_MASK \| UINT64_C(0x7ff0000000000200))) == 0);
4506	if ( pShwPDPT->a[i].n.u1Present
4507	&& !(pShwPDPT->a[i].u & PGM_PLXFLAGS_MAPPING)
4508	)
4509	{
4510	PPGMPOOLPAGE pSubPage = (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, pShwPDPT->a[i].u & X86_PDPE_PG_MASK);
4511	if (pSubPage)
4512	pgmPoolTrackFreeUser(pPool, pSubPage, pPage->idx, i);
4513	else
4514	AssertFatalMsgFailed(("%RX64\n", pShwPDPT->a[i].u & X86_PDPE_PG_MASK));
4515	}
4516	}
4517	}
4518
4519
4520	/**
4521	* Clear references to shadowed pages in a 64-bit page directory pointer table.
4522	*
4523	* @param pPool The pool.
4524	* @param pPage The page.
4525	* @param pShwPDPT The shadow page directory pointer table (mapping of the page).
4526	*/
4527	DECLINLINE(void) pgmPoolTrackDerefPDPT64Bit(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PX86PDPT pShwPDPT)
4528	{
4529	for (unsigned i = 0; i < RT_ELEMENTS(pShwPDPT->a); i++)
4530	{
4531	Assert((pShwPDPT->a[i].u & (X86_PDPE_LM_MBZ_MASK_NX \| UINT64_C(0x7ff0000000000200))) == 0);
4532	if (pShwPDPT->a[i].n.u1Present)
4533	{
4534	PPGMPOOLPAGE pSubPage = (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, pShwPDPT->a[i].u & X86_PDPE_PG_MASK);
4535	if (pSubPage)
4536	pgmPoolTrackFreeUser(pPool, pSubPage, pPage->idx, i);
4537	else
4538	AssertFatalMsgFailed(("%RX64\n", pShwPDPT->a[i].u & X86_PDPE_PG_MASK));
4539	/** @todo 64-bit guests: have to ensure that we're not exhausting the dynamic mappings! */
4540	}
4541	}
4542	}
4543
4544
4545	/**
4546	* Clear references to shadowed pages in a 64-bit level 4 page table.
4547	*
4548	* @param pPool The pool.
4549	* @param pPage The page.
4550	* @param pShwPML4 The shadow page directory pointer table (mapping of the page).
4551	*/
4552	DECLINLINE(void) pgmPoolTrackDerefPML464Bit(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PX86PML4 pShwPML4)
4553	{
4554	for (unsigned i = 0; i < RT_ELEMENTS(pShwPML4->a); i++)
4555	{
4556	Assert((pShwPML4->a[i].u & (X86_PML4E_MBZ_MASK_NX \| UINT64_C(0x7ff0000000000200))) == 0);
4557	if (pShwPML4->a[i].n.u1Present)
4558	{
4559	PPGMPOOLPAGE pSubPage = (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, pShwPML4->a[i].u & X86_PDPE_PG_MASK);
4560	if (pSubPage)
4561	pgmPoolTrackFreeUser(pPool, pSubPage, pPage->idx, i);
4562	else
4563	AssertFatalMsgFailed(("%RX64\n", pShwPML4->a[i].u & X86_PML4E_PG_MASK));
4564	/** @todo 64-bit guests: have to ensure that we're not exhausting the dynamic mappings! */
4565	}
4566	}
4567	}
4568
4569
4570	/**
4571	* Clear references to shadowed pages in an EPT page directory.
4572	*
4573	* @param pPool The pool.
4574	* @param pPage The page.
4575	* @param pShwPD The shadow page directory (mapping of the page).
4576	*/
4577	DECLINLINE(void) pgmPoolTrackDerefPDEPT(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PEPTPD pShwPD)
4578	{
4579	for (unsigned i = 0; i < RT_ELEMENTS(pShwPD->a); i++)
4580	{
4581	Assert((pShwPD->a[i].u & UINT64_C(0xfff0000000000f80)) == 0);
4582	if (pShwPD->a[i].n.u1Present)
4583	{
4584	#ifdef PGM_WITH_LARGE_PAGES
4585	if (pShwPD->a[i].b.u1Size)
4586	{
4587	Log4(("pgmPoolTrackDerefPDEPT: i=%d pde=%RX64 GCPhys=%RX64\n",
4588	i, pShwPD->a[i].u & X86_PDE2M_PAE_PG_MASK, pPage->GCPhys));
4589	pgmPoolTracDerefGCPhys(pPool, pPage, pShwPD->a[i].u & X86_PDE2M_PAE_PG_MASK,
4590	pPage->GCPhys + i * 2 * _1M /* pPage->GCPhys = base address of the memory described by the PD */,
4591	i);
4592	}
4593	else
4594	#endif
4595	{
4596	PPGMPOOLPAGE pSubPage = (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, pShwPD->a[i].u & EPT_PDE_PG_MASK);
4597	if (pSubPage)
4598	pgmPoolTrackFreeUser(pPool, pSubPage, pPage->idx, i);
4599	else
4600	AssertFatalMsgFailed(("%RX64\n", pShwPD->a[i].u & EPT_PDE_PG_MASK));
4601	}
4602	/** @todo 64-bit guests: have to ensure that we're not exhausting the dynamic mappings! */
4603	}
4604	}
4605	}
4606
4607
4608	/**
4609	* Clear references to shadowed pages in an EPT page directory pointer table.
4610	*
4611	* @param pPool The pool.
4612	* @param pPage The page.
4613	* @param pShwPDPT The shadow page directory pointer table (mapping of the page).
4614	*/
4615	DECLINLINE(void) pgmPoolTrackDerefPDPTEPT(PPGMPOOL pPool, PPGMPOOLPAGE pPage, PEPTPDPT pShwPDPT)
4616	{
4617	for (unsigned i = 0; i < RT_ELEMENTS(pShwPDPT->a); i++)
4618	{
4619	Assert((pShwPDPT->a[i].u & UINT64_C(0xfff0000000000f80)) == 0);
4620	if (pShwPDPT->a[i].n.u1Present)
4621	{
4622	PPGMPOOLPAGE pSubPage = (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, pShwPDPT->a[i].u & EPT_PDPTE_PG_MASK);
4623	if (pSubPage)
4624	pgmPoolTrackFreeUser(pPool, pSubPage, pPage->idx, i);
4625	else
4626	AssertFatalMsgFailed(("%RX64\n", pShwPDPT->a[i].u & EPT_PDPTE_PG_MASK));
4627	/** @todo 64-bit guests: have to ensure that we're not exhausting the dynamic mappings! */
4628	}
4629	}
4630	}
4631
4632
4633	/**
4634	* Clears all references made by this page.
4635	*
4636	* This includes other shadow pages and GC physical addresses.
4637	*
4638	* @param pPool The pool.
4639	* @param pPage The page.
4640	*/
4641	static void pgmPoolTrackDeref(PPGMPOOL pPool, PPGMPOOLPAGE pPage)
4642	{
4643	/*
4644	* Map the shadow page and take action according to the page kind.
4645	*/
4646	PVM pVM = pPool->CTX_SUFF(pVM);
4647	void *pvShw = PGMPOOL_PAGE_2_PTR(pVM, pPage);
4648	switch (pPage->enmKind)
4649	{
4650	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
4651	{
4652	STAM_PROFILE_START(&pPool->StatTrackDerefGCPhys, g);
4653	void *pvGst;
4654	int rc = PGM_GCPHYS_2_PTR(pVM, pPage->GCPhys, &pvGst); AssertReleaseRC(rc);
4655	pgmPoolTrackDerefPT32Bit32Bit(pPool, pPage, (PX86PT)pvShw, (PCX86PT)pvGst);
4656	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvGst);
4657	STAM_PROFILE_STOP(&pPool->StatTrackDerefGCPhys, g);
4658	break;
4659	}
4660
4661	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
4662	{
4663	STAM_PROFILE_START(&pPool->StatTrackDerefGCPhys, g);
4664	void *pvGst;
4665	int rc = PGM_GCPHYS_2_PTR_EX(pVM, pPage->GCPhys, &pvGst); AssertReleaseRC(rc);
4666	pgmPoolTrackDerefPTPae32Bit(pPool, pPage, (PPGMSHWPTPAE)pvShw, (PCX86PT)pvGst);
4667	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvGst);
4668	STAM_PROFILE_STOP(&pPool->StatTrackDerefGCPhys, g);
4669	break;
4670	}
4671
4672	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
4673	{
4674	STAM_PROFILE_START(&pPool->StatTrackDerefGCPhys, g);
4675	void *pvGst;
4676	int rc = PGM_GCPHYS_2_PTR(pVM, pPage->GCPhys, &pvGst); AssertReleaseRC(rc);
4677	pgmPoolTrackDerefPTPaePae(pPool, pPage, (PPGMSHWPTPAE)pvShw, (PCX86PTPAE)pvGst);
4678	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvGst);
4679	STAM_PROFILE_STOP(&pPool->StatTrackDerefGCPhys, g);
4680	break;
4681	}
4682
4683	case PGMPOOLKIND_32BIT_PT_FOR_PHYS: /* treat it like a 4 MB page */
4684	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
4685	{
4686	STAM_PROFILE_START(&pPool->StatTrackDerefGCPhys, g);
4687	pgmPoolTrackDerefPT32Bit4MB(pPool, pPage, (PX86PT)pvShw);
4688	STAM_PROFILE_STOP(&pPool->StatTrackDerefGCPhys, g);
4689	break;
4690	}
4691
4692	case PGMPOOLKIND_PAE_PT_FOR_PHYS: /* treat it like a 2 MB page */
4693	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
4694	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
4695	{
4696	STAM_PROFILE_START(&pPool->StatTrackDerefGCPhys, g);
4697	pgmPoolTrackDerefPTPaeBig(pPool, pPage, (PPGMSHWPTPAE)pvShw);
4698	STAM_PROFILE_STOP(&pPool->StatTrackDerefGCPhys, g);
4699	break;
4700	}
4701
4702	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
4703	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
4704	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
4705	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
4706	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
4707	case PGMPOOLKIND_PAE_PD_PHYS:
4708	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
4709	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
4710	pgmPoolTrackDerefPDPae(pPool, pPage, (PX86PDPAE)pvShw);
4711	break;
4712
4713	case PGMPOOLKIND_32BIT_PD_PHYS:
4714	case PGMPOOLKIND_32BIT_PD:
4715	pgmPoolTrackDerefPD(pPool, pPage, (PX86PD)pvShw);
4716	break;
4717
4718	case PGMPOOLKIND_PAE_PDPT_FOR_32BIT:
4719	case PGMPOOLKIND_PAE_PDPT:
4720	case PGMPOOLKIND_PAE_PDPT_PHYS:
4721	pgmPoolTrackDerefPDPTPae(pPool, pPage, (PX86PDPT)pvShw);
4722	break;
4723
4724	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
4725	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
4726	pgmPoolTrackDerefPDPT64Bit(pPool, pPage, (PX86PDPT)pvShw);
4727	break;
4728
4729	case PGMPOOLKIND_64BIT_PML4:
4730	pgmPoolTrackDerefPML464Bit(pPool, pPage, (PX86PML4)pvShw);
4731	break;
4732
4733	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
4734	pgmPoolTrackDerefPTEPT(pPool, pPage, (PEPTPT)pvShw);
4735	break;
4736
4737	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
4738	pgmPoolTrackDerefPDEPT(pPool, pPage, (PEPTPD)pvShw);
4739	break;
4740
4741	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
4742	pgmPoolTrackDerefPDPTEPT(pPool, pPage, (PEPTPDPT)pvShw);
4743	break;
4744
4745	default:
4746	AssertFatalMsgFailed(("enmKind=%d\n", pPage->enmKind));
4747	}
4748
4749	/* paranoia, clear the shadow page. Remove this laser (i.e. let Alloc and ClearAll do it). */
4750	STAM_PROFILE_START(&pPool->StatZeroPage, z);
4751	ASMMemZeroPage(pvShw);
4752	STAM_PROFILE_STOP(&pPool->StatZeroPage, z);
4753	pPage->fZeroed = true;
4754	Assert(!pPage->cPresent);
4755	PGM_DYNMAP_UNUSED_HINT_VM(pVM, pvShw);
4756	}
4757
4758
4759	/**
4760	* Flushes a pool page.
4761	*
4762	* This moves the page to the free list after removing all user references to it.
4763	*
4764	* @returns VBox status code.
4765	* @retval VINF_SUCCESS on success.
4766	* @param pPool The pool.
4767	* @param HCPhys The HC physical address of the shadow page.
4768	* @param fFlush Flush the TLBS when required (should only be false in very specific use cases!!)
4769	*/
4770	int pgmPoolFlushPage(PPGMPOOL pPool, PPGMPOOLPAGE pPage, bool fFlush)
4771	{
4772	PVM pVM = pPool->CTX_SUFF(pVM);
4773	bool fFlushRequired = false;
4774
4775	int rc = VINF_SUCCESS;
4776	STAM_PROFILE_START(&pPool->StatFlushPage, f);
4777	LogFlow(("pgmPoolFlushPage: pPage=%p:{.Key=%RHp, .idx=%d, .enmKind=%s, .GCPhys=%RGp}\n",
4778	pPage, pPage->Core.Key, pPage->idx, pgmPoolPoolKindToStr(pPage->enmKind), pPage->GCPhys));
4779
4780	/*
4781	* Quietly reject any attempts at flushing any of the special root pages.
4782	*/
4783	if (pPage->idx < PGMPOOL_IDX_FIRST)
4784	{
4785	AssertFailed(); /* can no longer happen */
4786	Log(("pgmPoolFlushPage: special root page, rejected. enmKind=%s idx=%d\n", pgmPoolPoolKindToStr(pPage->enmKind), pPage->idx));
4787	return VINF_SUCCESS;
4788	}
4789
4790	pgmLock(pVM);
4791
4792	/*
4793	* Quietly reject any attempts at flushing the currently active shadow CR3 mapping
4794	*/
4795	if (pgmPoolIsPageLocked(pPage))
4796	{
4797	AssertMsg( pPage->enmKind == PGMPOOLKIND_64BIT_PML4
4798	\|\| pPage->enmKind == PGMPOOLKIND_PAE_PDPT
4799	\|\| pPage->enmKind == PGMPOOLKIND_PAE_PDPT_FOR_32BIT
4800	\|\| pPage->enmKind == PGMPOOLKIND_32BIT_PD
4801	\|\| pPage->enmKind == PGMPOOLKIND_PAE_PD_FOR_PAE_PD
4802	\|\| pPage->enmKind == PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD
4803	\|\| pPage->enmKind == PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD
4804	\|\| pPage->enmKind == PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD
4805	\|\| pPage->enmKind == PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD
4806	\|\| pPage->enmKind == PGMPOOLKIND_ROOT_NESTED,
4807	("Can't free the shadow CR3! (%RHp vs %RHp kind=%d\n", PGMGetHyperCR3(VMMGetCpu(pVM)), pPage->Core.Key, pPage->enmKind));
4808	Log(("pgmPoolFlushPage: current active shadow CR3, rejected. enmKind=%s idx=%d\n", pgmPoolPoolKindToStr(pPage->enmKind), pPage->idx));
4809	pgmUnlock(pVM);
4810	return VINF_SUCCESS;
4811	}
4812
4813	#if defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) \|\| defined(IN_RC)
4814	/* Start a subset so we won't run out of mapping space. */
4815	PVMCPU pVCpu = VMMGetCpu(pVM);
4816	uint32_t iPrevSubset = PGMRZDynMapPushAutoSubset(pVCpu);
4817	#endif
4818
4819	/*
4820	* Mark the page as being in need of an ASMMemZeroPage().
4821	*/
4822	pPage->fZeroed = false;
4823
4824	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
4825	if (pPage->fDirty)
4826	pgmPoolFlushDirtyPage(pVM, pPool, pPage->idxDirtyEntry, false /* do not remove */);
4827	#endif
4828
4829	/* If there are any users of this table, then we must issue a tlb flush on all VCPUs. */
4830	if (pPage->iUserHead != NIL_PGMPOOL_USER_INDEX)
4831	fFlushRequired = true;
4832
4833	/*
4834	* Clear the page.
4835	*/
4836	pgmPoolTrackClearPageUsers(pPool, pPage);
4837	STAM_PROFILE_START(&pPool->StatTrackDeref,a);
4838	pgmPoolTrackDeref(pPool, pPage);
4839	STAM_PROFILE_STOP(&pPool->StatTrackDeref,a);
4840
4841	/*
4842	* Flush it from the cache.
4843	*/
4844	pgmPoolCacheFlushPage(pPool, pPage);
4845
4846	#if defined(VBOX_WITH_2X_4GB_ADDR_SPACE_IN_R0) \|\| defined(IN_RC)
4847	/* Heavy stuff done. */
4848	PGMRZDynMapPopAutoSubset(pVCpu, iPrevSubset);
4849	#endif
4850
4851	/*
4852	* Deregistering the monitoring.
4853	*/
4854	if (pPage->fMonitored)
4855	rc = pgmPoolMonitorFlush(pPool, pPage);
4856
4857	/*
4858	* Free the page.
4859	*/
4860	Assert(pPage->iNext == NIL_PGMPOOL_IDX);
4861	pPage->iNext = pPool->iFreeHead;
4862	pPool->iFreeHead = pPage->idx;
4863	pPage->enmKind = PGMPOOLKIND_FREE;
4864	pPage->enmAccess = PGMPOOLACCESS_DONTCARE;
4865	pPage->GCPhys = NIL_RTGCPHYS;
4866	pPage->fReusedFlushPending = false;
4867
4868	pPool->cUsedPages--;
4869
4870	/* Flush the TLBs of all VCPUs if required. */
4871	if ( fFlushRequired
4872	&& fFlush)
4873	{
4874	PGM_INVL_ALL_VCPU_TLBS(pVM);
4875	}
4876
4877	pgmUnlock(pVM);
4878	STAM_PROFILE_STOP(&pPool->StatFlushPage, f);
4879	return rc;
4880	}
4881
4882
4883	/**
4884	* Frees a usage of a pool page.
4885	*
4886	* The caller is responsible to updating the user table so that it no longer
4887	* references the shadow page.
4888	*
4889	* @param pPool The pool.
4890	* @param HCPhys The HC physical address of the shadow page.
4891	* @param iUser The shadow page pool index of the user table.
4892	* @param iUserTable The index into the user table (shadowed).
4893	*/
4894	void pgmPoolFreeByPage(PPGMPOOL pPool, PPGMPOOLPAGE pPage, uint16_t iUser, uint32_t iUserTable)
4895	{
4896	PVM pVM = pPool->CTX_SUFF(pVM);
4897
4898	STAM_PROFILE_START(&pPool->StatFree, a);
4899	LogFlow(("pgmPoolFreeByPage: pPage=%p:{.Key=%RHp, .idx=%d, enmKind=%s} iUser=%d iUserTable=%#x\n",
4900	pPage, pPage->Core.Key, pPage->idx, pgmPoolPoolKindToStr(pPage->enmKind), iUser, iUserTable));
4901	Assert(pPage->idx >= PGMPOOL_IDX_FIRST);
4902	pgmLock(pVM);
4903	pgmPoolTrackFreeUser(pPool, pPage, iUser, iUserTable);
4904	if (!pPage->fCached)
4905	pgmPoolFlushPage(pPool, pPage);
4906	pgmUnlock(pVM);
4907	STAM_PROFILE_STOP(&pPool->StatFree, a);
4908	}
4909
4910
4911	/**
4912	* Makes one or more free page free.
4913	*
4914	* @returns VBox status code.
4915	* @retval VINF_SUCCESS on success.
4916	* @retval VERR_PGM_POOL_FLUSHED if the pool was flushed.
4917	*
4918	* @param pPool The pool.
4919	* @param enmKind Page table kind
4920	* @param iUser The user of the page.
4921	*/
4922	static int pgmPoolMakeMoreFreePages(PPGMPOOL pPool, PGMPOOLKIND enmKind, uint16_t iUser)
4923	{
4924	PVM pVM = pPool->CTX_SUFF(pVM);
4925	LogFlow(("pgmPoolMakeMoreFreePages: iUser=%d\n", iUser));
4926	NOREF(enmKind);
4927
4928	/*
4929	* If the pool isn't full grown yet, expand it.
4930	*/
4931	if ( pPool->cCurPages < pPool->cMaxPages
4932	#if defined(IN_RC)
4933	/* Hack alert: we can't deal with jumps to ring 3 when called from MapCR3 and allocating pages for PAE PDs. */
4934	&& enmKind != PGMPOOLKIND_PAE_PD_FOR_PAE_PD
4935	&& (enmKind < PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD \|\| enmKind > PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD)
4936	#endif
4937	)
4938	{
4939	STAM_PROFILE_ADV_SUSPEND(&pPool->StatAlloc, a);
4940	#ifdef IN_RING3
4941	int rc = PGMR3PoolGrow(pVM);
4942	#else
4943	int rc = VMMRZCallRing3NoCpu(pVM, VMMCALLRING3_PGM_POOL_GROW, 0);
4944	#endif
4945	if (RT_FAILURE(rc))
4946	return rc;
4947	STAM_PROFILE_ADV_RESUME(&pPool->StatAlloc, a);
4948	if (pPool->iFreeHead != NIL_PGMPOOL_IDX)
4949	return VINF_SUCCESS;
4950	}
4951
4952	/*
4953	* Free one cached page.
4954	*/
4955	return pgmPoolCacheFreeOne(pPool, iUser);
4956	}
4957
4958
4959	/**
4960	* Allocates a page from the pool.
4961	*
4962	* This page may actually be a cached page and not in need of any processing
4963	* on the callers part.
4964	*
4965	* @returns VBox status code.
4966	* @retval VINF_SUCCESS if a NEW page was allocated.
4967	* @retval VINF_PGM_CACHED_PAGE if a CACHED page was returned.
4968	* @retval VERR_PGM_POOL_FLUSHED if the pool was flushed.
4969	*
4970	* @param pVM Pointer to the VM.
4971	* @param GCPhys The GC physical address of the page we're gonna shadow.
4972	* For 4MB and 2MB PD entries, it's the first address the
4973	* shadow PT is covering.
4974	* @param enmKind The kind of mapping.
4975	* @param enmAccess Access type for the mapping (only relevant for big pages)
4976	* @param fA20Enabled Whether the A20 gate is enabled or not.
4977	* @param iUser The shadow page pool index of the user table.
4978	* @param iUserTable The index into the user table (shadowed).
4979	* @param fLockPage Lock the page
4980	* @param ppPage Where to store the pointer to the page. NULL is stored here on failure.
4981	*/
4982	int pgmPoolAlloc(PVM pVM, RTGCPHYS GCPhys, PGMPOOLKIND enmKind, PGMPOOLACCESS enmAccess, bool fA20Enabled,
4983	uint16_t iUser, uint32_t iUserTable, bool fLockPage, PPPGMPOOLPAGE ppPage)
4984	{
4985	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
4986	STAM_PROFILE_ADV_START(&pPool->StatAlloc, a);
4987	LogFlow(("pgmPoolAlloc: GCPhys=%RGp enmKind=%s iUser=%d iUserTable=%#x\n", GCPhys, pgmPoolPoolKindToStr(enmKind), iUser, iUserTable));
4988	*ppPage = NULL;
4989	/** @todo CSAM/PGMPrefetchPage messes up here during CSAMR3CheckGates
4990	* (TRPMR3SyncIDT) because of FF priority. Try fix that?
4991	* Assert(!(pVM->pgm.s.fGlobalSyncFlags & PGM_SYNC_CLEAR_PGM_POOL)); */
4992
4993	pgmLock(pVM);
4994
4995	if (pPool->fCacheEnabled)
4996	{
4997	int rc2 = pgmPoolCacheAlloc(pPool, GCPhys, enmKind, enmAccess, fA20Enabled, iUser, iUserTable, ppPage);
4998	if (RT_SUCCESS(rc2))
4999	{
5000	if (fLockPage)
5001	pgmPoolLockPage(pPool, *ppPage);
5002	pgmUnlock(pVM);
5003	STAM_PROFILE_ADV_STOP(&pPool->StatAlloc, a);
5004	LogFlow(("pgmPoolAlloc: cached returns %Rrc ppPage=%p:{.Key=%RHp, .idx=%d}\n", rc2, ppPage, (ppPage)->Core.Key, (ppPage)->idx));
5005	return rc2;
5006	}
5007	}
5008
5009	/*
5010	* Allocate a new one.
5011	*/
5012	int rc = VINF_SUCCESS;
5013	uint16_t iNew = pPool->iFreeHead;
5014	if (iNew == NIL_PGMPOOL_IDX)
5015	{
5016	rc = pgmPoolMakeMoreFreePages(pPool, enmKind, iUser);
5017	if (RT_FAILURE(rc))
5018	{
5019	pgmUnlock(pVM);
5020	Log(("pgmPoolAlloc: returns %Rrc (Free)\n", rc));
5021	STAM_PROFILE_ADV_STOP(&pPool->StatAlloc, a);
5022	return rc;
5023	}
5024	iNew = pPool->iFreeHead;
5025	AssertReleaseReturn(iNew != NIL_PGMPOOL_IDX, VERR_PGM_POOL_IPE);
5026	}
5027
5028	/* unlink the free head */
5029	PPGMPOOLPAGE pPage = &pPool->aPages[iNew];
5030	pPool->iFreeHead = pPage->iNext;
5031	pPage->iNext = NIL_PGMPOOL_IDX;
5032
5033	/*
5034	* Initialize it.
5035	*/
5036	pPool->cUsedPages++; /* physical handler registration / pgmPoolTrackFlushGCPhysPTsSlow requirement. */
5037	pPage->enmKind = enmKind;
5038	pPage->enmAccess = enmAccess;
5039	pPage->GCPhys = GCPhys;
5040	pPage->fA20Enabled = fA20Enabled;
5041	pPage->fSeenNonGlobal = false; /* Set this to 'true' to disable this feature. */
5042	pPage->fMonitored = false;
5043	pPage->fCached = false;
5044	pPage->fDirty = false;
5045	pPage->fReusedFlushPending = false;
5046	pPage->cModifications = 0;
5047	pPage->iModifiedNext = NIL_PGMPOOL_IDX;
5048	pPage->iModifiedPrev = NIL_PGMPOOL_IDX;
5049	pPage->cPresent = 0;
5050	pPage->iFirstPresent = NIL_PGMPOOL_PRESENT_INDEX;
5051	pPage->idxDirtyEntry = 0;
5052	pPage->GCPtrLastAccessHandlerFault = NIL_RTGCPTR;
5053	pPage->GCPtrLastAccessHandlerRip = NIL_RTGCPTR;
5054	pPage->cLastAccessHandler = 0;
5055	pPage->cLocked = 0;
5056	# ifdef VBOX_STRICT
5057	pPage->GCPtrDirtyFault = NIL_RTGCPTR;
5058	# endif
5059
5060	/*
5061	* Insert into the tracking and cache. If this fails, free the page.
5062	*/
5063	int rc3 = pgmPoolTrackInsert(pPool, pPage, GCPhys, iUser, iUserTable);
5064	if (RT_FAILURE(rc3))
5065	{
5066	pPool->cUsedPages--;
5067	pPage->enmKind = PGMPOOLKIND_FREE;
5068	pPage->enmAccess = PGMPOOLACCESS_DONTCARE;
5069	pPage->GCPhys = NIL_RTGCPHYS;
5070	pPage->iNext = pPool->iFreeHead;
5071	pPool->iFreeHead = pPage->idx;
5072	pgmUnlock(pVM);
5073	STAM_PROFILE_ADV_STOP(&pPool->StatAlloc, a);
5074	Log(("pgmPoolAlloc: returns %Rrc (Insert)\n", rc3));
5075	return rc3;
5076	}
5077
5078	/*
5079	* Commit the allocation, clear the page and return.
5080	*/
5081	#ifdef VBOX_WITH_STATISTICS
5082	if (pPool->cUsedPages > pPool->cUsedPagesHigh)
5083	pPool->cUsedPagesHigh = pPool->cUsedPages;
5084	#endif
5085
5086	if (!pPage->fZeroed)
5087	{
5088	STAM_PROFILE_START(&pPool->StatZeroPage, z);
5089	void *pv = PGMPOOL_PAGE_2_PTR(pVM, pPage);
5090	ASMMemZeroPage(pv);
5091	STAM_PROFILE_STOP(&pPool->StatZeroPage, z);
5092	}
5093
5094	*ppPage = pPage;
5095	if (fLockPage)
5096	pgmPoolLockPage(pPool, pPage);
5097	pgmUnlock(pVM);
5098	LogFlow(("pgmPoolAlloc: returns %Rrc *ppPage=%p:{.Key=%RHp, .idx=%d, .fCached=%RTbool, .fMonitored=%RTbool}\n",
5099	rc, pPage, pPage->Core.Key, pPage->idx, pPage->fCached, pPage->fMonitored));
5100	STAM_PROFILE_ADV_STOP(&pPool->StatAlloc, a);
5101	return rc;
5102	}
5103
5104
5105	/**
5106	* Frees a usage of a pool page.
5107	*
5108	* @param pVM Pointer to the VM.
5109	* @param HCPhys The HC physical address of the shadow page.
5110	* @param iUser The shadow page pool index of the user table.
5111	* @param iUserTable The index into the user table (shadowed).
5112	*/
5113	void pgmPoolFree(PVM pVM, RTHCPHYS HCPhys, uint16_t iUser, uint32_t iUserTable)
5114	{
5115	LogFlow(("pgmPoolFree: HCPhys=%RHp iUser=%d iUserTable=%#x\n", HCPhys, iUser, iUserTable));
5116	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
5117	pgmPoolFreeByPage(pPool, pgmPoolGetPage(pPool, HCPhys), iUser, iUserTable);
5118	}
5119
5120
5121	/**
5122	* Internal worker for finding a 'in-use' shadow page give by it's physical address.
5123	*
5124	* @returns Pointer to the shadow page structure.
5125	* @param pPool The pool.
5126	* @param HCPhys The HC physical address of the shadow page.
5127	*/
5128	PPGMPOOLPAGE pgmPoolGetPage(PPGMPOOL pPool, RTHCPHYS HCPhys)
5129	{
5130	PGM_LOCK_ASSERT_OWNER(pPool->CTX_SUFF(pVM));
5131
5132	/*
5133	* Look up the page.
5134	*/
5135	PPGMPOOLPAGE pPage = (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, HCPhys & X86_PTE_PAE_PG_MASK);
5136
5137	AssertFatalMsg(pPage && pPage->enmKind != PGMPOOLKIND_FREE, ("HCPhys=%RHp pPage=%p idx=%d\n", HCPhys, pPage, (pPage) ? pPage->idx : 0));
5138	return pPage;
5139	}
5140
5141
5142	/**
5143	* Internal worker for finding a page for debugging purposes, no assertions.
5144	*
5145	* @returns Pointer to the shadow page structure. NULL on if not found.
5146	* @param pPool The pool.
5147	* @param HCPhys The HC physical address of the shadow page.
5148	*/
5149	PPGMPOOLPAGE pgmPoolQueryPageForDbg(PPGMPOOL pPool, RTHCPHYS HCPhys)
5150	{
5151	PGM_LOCK_ASSERT_OWNER(pPool->CTX_SUFF(pVM));
5152	return (PPGMPOOLPAGE)RTAvloHCPhysGet(&pPool->HCPhysTree, HCPhys & X86_PTE_PAE_PG_MASK);
5153	}
5154
5155	#ifdef IN_RING3 /* currently only used in ring 3; save some space in the R0 & GC modules (left it here as we might need it elsewhere later on) */
5156
5157	/**
5158	* Flush the specified page if present
5159	*
5160	* @param pVM Pointer to the VM.
5161	* @param GCPhys Guest physical address of the page to flush
5162	*/
5163	void pgmPoolFlushPageByGCPhys(PVM pVM, RTGCPHYS GCPhys)
5164	{
5165	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
5166
5167	VM_ASSERT_EMT(pVM);
5168
5169	/*
5170	* Look up the GCPhys in the hash.
5171	*/
5172	GCPhys = GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK;
5173	unsigned i = pPool->aiHash[PGMPOOL_HASH(GCPhys)];
5174	if (i == NIL_PGMPOOL_IDX)
5175	return;
5176
5177	do
5178	{
5179	PPGMPOOLPAGE pPage = &pPool->aPages[i];
5180	if (pPage->GCPhys - GCPhys < PAGE_SIZE)
5181	{
5182	switch (pPage->enmKind)
5183	{
5184	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
5185	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
5186	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
5187	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
5188	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
5189	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
5190	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
5191	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
5192	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
5193	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
5194	case PGMPOOLKIND_64BIT_PML4:
5195	case PGMPOOLKIND_32BIT_PD:
5196	case PGMPOOLKIND_PAE_PDPT:
5197	{
5198	Log(("PGMPoolFlushPage: found pgm pool pages for %RGp\n", GCPhys));
5199	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
5200	if (pPage->fDirty)
5201	STAM_COUNTER_INC(&pPool->StatForceFlushDirtyPage);
5202	else
5203	#endif
5204	STAM_COUNTER_INC(&pPool->StatForceFlushPage);
5205	Assert(!pgmPoolIsPageLocked(pPage));
5206	pgmPoolMonitorChainFlush(pPool, pPage);
5207	return;
5208	}
5209
5210	/* ignore, no monitoring. */
5211	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
5212	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
5213	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
5214	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
5215	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
5216	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
5217	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
5218	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
5219	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
5220	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
5221	case PGMPOOLKIND_ROOT_NESTED:
5222	case PGMPOOLKIND_PAE_PD_PHYS:
5223	case PGMPOOLKIND_PAE_PDPT_PHYS:
5224	case PGMPOOLKIND_32BIT_PD_PHYS:
5225	case PGMPOOLKIND_PAE_PDPT_FOR_32BIT:
5226	break;
5227
5228	default:
5229	AssertFatalMsgFailed(("enmKind=%d idx=%d\n", pPage->enmKind, pPage->idx));
5230	}
5231	}
5232
5233	/* next */
5234	i = pPage->iNext;
5235	} while (i != NIL_PGMPOOL_IDX);
5236	return;
5237	}
5238
5239	#endif /* IN_RING3 */
5240	#ifdef IN_RING3
5241
5242	/**
5243	* Reset CPU on hot plugging.
5244	*
5245	* @param pVM Pointer to the VM.
5246	* @param pVCpu The virtual CPU.
5247	*/
5248	void pgmR3PoolResetUnpluggedCpu(PVM pVM, PVMCPU pVCpu)
5249	{
5250	pgmR3ExitShadowModeBeforePoolFlush(pVCpu);
5251
5252	pgmR3ReEnterShadowModeAfterPoolFlush(pVM, pVCpu);
5253	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
5254	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
5255	}
5256
5257
5258	/**
5259	* Flushes the entire cache.
5260	*
5261	* It will assert a global CR3 flush (FF) and assumes the caller is aware of
5262	* this and execute this CR3 flush.
5263	*
5264	* @param pPool The pool.
5265	*/
5266	void pgmR3PoolReset(PVM pVM)
5267	{
5268	PPGMPOOL pPool = pVM->pgm.s.CTX_SUFF(pPool);
5269
5270	PGM_LOCK_ASSERT_OWNER(pVM);
5271	STAM_PROFILE_START(&pPool->StatR3Reset, a);
5272	LogFlow(("pgmR3PoolReset:\n"));
5273
5274	/*
5275	* If there are no pages in the pool, there is nothing to do.
5276	*/
5277	if (pPool->cCurPages <= PGMPOOL_IDX_FIRST)
5278	{
5279	STAM_PROFILE_STOP(&pPool->StatR3Reset, a);
5280	return;
5281	}
5282
5283	/*
5284	* Exit the shadow mode since we're going to clear everything,
5285	* including the root page.
5286	*/
5287	for (VMCPUID i = 0; i < pVM->cCpus; i++)
5288	pgmR3ExitShadowModeBeforePoolFlush(&pVM->aCpus[i]);
5289
5290	/*
5291	* Nuke the free list and reinsert all pages into it.
5292	*/
5293	for (unsigned i = pPool->cCurPages - 1; i >= PGMPOOL_IDX_FIRST; i--)
5294	{
5295	PPGMPOOLPAGE pPage = &pPool->aPages[i];
5296
5297	Assert(pPage->Core.Key == MMPage2Phys(pVM, pPage->pvPageR3));
5298	if (pPage->fMonitored)
5299	pgmPoolMonitorFlush(pPool, pPage);
5300	pPage->iModifiedNext = NIL_PGMPOOL_IDX;
5301	pPage->iModifiedPrev = NIL_PGMPOOL_IDX;
5302	pPage->iMonitoredNext = NIL_PGMPOOL_IDX;
5303	pPage->iMonitoredPrev = NIL_PGMPOOL_IDX;
5304	pPage->cModifications = 0;
5305	pPage->GCPhys = NIL_RTGCPHYS;
5306	pPage->enmKind = PGMPOOLKIND_FREE;
5307	pPage->enmAccess = PGMPOOLACCESS_DONTCARE;
5308	Assert(pPage->idx == i);
5309	pPage->iNext = i + 1;
5310	pPage->fA20Enabled = true;
5311	pPage->fZeroed = false; /* This could probably be optimized, but better safe than sorry. */
5312	pPage->fSeenNonGlobal = false;
5313	pPage->fMonitored = false;
5314	pPage->fDirty = false;
5315	pPage->fCached = false;
5316	pPage->fReusedFlushPending = false;
5317	pPage->iUserHead = NIL_PGMPOOL_USER_INDEX;
5318	pPage->iAgeNext = NIL_PGMPOOL_IDX;
5319	pPage->iAgePrev = NIL_PGMPOOL_IDX;
5320	pPage->GCPtrLastAccessHandlerRip = NIL_RTGCPTR;
5321	pPage->GCPtrLastAccessHandlerFault = NIL_RTGCPTR;
5322	pPage->cLastAccessHandler = 0;
5323	pPage->cLocked = 0;
5324	#ifdef VBOX_STRICT
5325	pPage->GCPtrDirtyFault = NIL_RTGCPTR;
5326	#endif
5327	}
5328	pPool->aPages[pPool->cCurPages - 1].iNext = NIL_PGMPOOL_IDX;
5329	pPool->iFreeHead = PGMPOOL_IDX_FIRST;
5330	pPool->cUsedPages = 0;
5331
5332	/*
5333	* Zap and reinitialize the user records.
5334	*/
5335	pPool->cPresent = 0;
5336	pPool->iUserFreeHead = 0;
5337	PPGMPOOLUSER paUsers = pPool->CTX_SUFF(paUsers);
5338	const unsigned cMaxUsers = pPool->cMaxUsers;
5339	for (unsigned i = 0; i < cMaxUsers; i++)
5340	{
5341	paUsers[i].iNext = i + 1;
5342	paUsers[i].iUser = NIL_PGMPOOL_IDX;
5343	paUsers[i].iUserTable = 0xfffffffe;
5344	}
5345	paUsers[cMaxUsers - 1].iNext = NIL_PGMPOOL_USER_INDEX;
5346
5347	/*
5348	* Clear all the GCPhys links and rebuild the phys ext free list.
5349	*/
5350	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
5351	pRam;
5352	pRam = pRam->CTX_SUFF(pNext))
5353	{
5354	unsigned iPage = pRam->cb >> PAGE_SHIFT;
5355	while (iPage-- > 0)
5356	PGM_PAGE_SET_TRACKING(pVM, &pRam->aPages[iPage], 0);
5357	}
5358
5359	pPool->iPhysExtFreeHead = 0;
5360	PPGMPOOLPHYSEXT paPhysExts = pPool->CTX_SUFF(paPhysExts);
5361	const unsigned cMaxPhysExts = pPool->cMaxPhysExts;
5362	for (unsigned i = 0; i < cMaxPhysExts; i++)
5363	{
5364	paPhysExts[i].iNext = i + 1;
5365	paPhysExts[i].aidx[0] = NIL_PGMPOOL_IDX;
5366	paPhysExts[i].apte[0] = NIL_PGMPOOL_PHYSEXT_IDX_PTE;
5367	paPhysExts[i].aidx[1] = NIL_PGMPOOL_IDX;
5368	paPhysExts[i].apte[1] = NIL_PGMPOOL_PHYSEXT_IDX_PTE;
5369	paPhysExts[i].aidx[2] = NIL_PGMPOOL_IDX;
5370	paPhysExts[i].apte[2] = NIL_PGMPOOL_PHYSEXT_IDX_PTE;
5371	}
5372	paPhysExts[cMaxPhysExts - 1].iNext = NIL_PGMPOOL_PHYSEXT_INDEX;
5373
5374	/*
5375	* Just zap the modified list.
5376	*/
5377	pPool->cModifiedPages = 0;
5378	pPool->iModifiedHead = NIL_PGMPOOL_IDX;
5379
5380	/*
5381	* Clear the GCPhys hash and the age list.
5382	*/
5383	for (unsigned i = 0; i < RT_ELEMENTS(pPool->aiHash); i++)
5384	pPool->aiHash[i] = NIL_PGMPOOL_IDX;
5385	pPool->iAgeHead = NIL_PGMPOOL_IDX;
5386	pPool->iAgeTail = NIL_PGMPOOL_IDX;
5387
5388	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
5389	/* Clear all dirty pages. */
5390	pPool->idxFreeDirtyPage = 0;
5391	pPool->cDirtyPages = 0;
5392	for (unsigned i = 0; i < RT_ELEMENTS(pPool->aDirtyPages); i++)
5393	pPool->aDirtyPages[i].uIdx = NIL_PGMPOOL_IDX;
5394	#endif
5395
5396	/*
5397	* Reinsert active pages into the hash and ensure monitoring chains are correct.
5398	*/
5399	for (unsigned i = PGMPOOL_IDX_FIRST_SPECIAL; i < PGMPOOL_IDX_FIRST; i++)
5400	{
5401	PPGMPOOLPAGE pPage = &pPool->aPages[i];
5402	pPage->iNext = NIL_PGMPOOL_IDX;
5403	pPage->iModifiedNext = NIL_PGMPOOL_IDX;
5404	pPage->iModifiedPrev = NIL_PGMPOOL_IDX;
5405	pPage->cModifications = 0;
5406	/* ASSUMES that we're not sharing with any of the other special pages (safe for now). */
5407	pPage->iMonitoredNext = NIL_PGMPOOL_IDX;
5408	pPage->iMonitoredPrev = NIL_PGMPOOL_IDX;
5409	if (pPage->fMonitored)
5410	{
5411	int rc = PGMHandlerPhysicalChangeCallbacks(pVM, pPage->GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK,
5412	pPool->pfnAccessHandlerR3, MMHyperCCToR3(pVM, pPage),
5413	pPool->pfnAccessHandlerR0, MMHyperCCToR0(pVM, pPage),
5414	pPool->pfnAccessHandlerRC, MMHyperCCToRC(pVM, pPage),
5415	pPool->pszAccessHandler);
5416	AssertFatalRCSuccess(rc);
5417	pgmPoolHashInsert(pPool, pPage);
5418	}
5419	Assert(pPage->iUserHead == NIL_PGMPOOL_USER_INDEX); /* for now */
5420	Assert(pPage->iAgeNext == NIL_PGMPOOL_IDX);
5421	Assert(pPage->iAgePrev == NIL_PGMPOOL_IDX);
5422	}
5423
5424	for (VMCPUID i = 0; i < pVM->cCpus; i++)
5425	{
5426	/*
5427	* Re-enter the shadowing mode and assert Sync CR3 FF.
5428	*/
5429	PVMCPU pVCpu = &pVM->aCpus[i];
5430	pgmR3ReEnterShadowModeAfterPoolFlush(pVM, pVCpu);
5431	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
5432	VMCPU_FF_SET(pVCpu, VMCPU_FF_TLB_FLUSH);
5433	}
5434
5435	STAM_PROFILE_STOP(&pPool->StatR3Reset, a);
5436	}
5437
5438	#endif /* IN_RING3 */
5439
5440	#ifdef LOG_ENABLED
5441	/**
5442	* Stringifies a PGMPOOLKIND value.
5443	*/
5444	static const char *pgmPoolPoolKindToStr(uint8_t enmKind)
5445	{
5446	switch ((PGMPOOLKIND)enmKind)
5447	{
5448	case PGMPOOLKIND_INVALID:
5449	return "PGMPOOLKIND_INVALID";
5450	case PGMPOOLKIND_FREE:
5451	return "PGMPOOLKIND_FREE";
5452	case PGMPOOLKIND_32BIT_PT_FOR_PHYS:
5453	return "PGMPOOLKIND_32BIT_PT_FOR_PHYS";
5454	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT:
5455	return "PGMPOOLKIND_32BIT_PT_FOR_32BIT_PT";
5456	case PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB:
5457	return "PGMPOOLKIND_32BIT_PT_FOR_32BIT_4MB";
5458	case PGMPOOLKIND_PAE_PT_FOR_PHYS:
5459	return "PGMPOOLKIND_PAE_PT_FOR_PHYS";
5460	case PGMPOOLKIND_PAE_PT_FOR_32BIT_PT:
5461	return "PGMPOOLKIND_PAE_PT_FOR_32BIT_PT";
5462	case PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB:
5463	return "PGMPOOLKIND_PAE_PT_FOR_32BIT_4MB";
5464	case PGMPOOLKIND_PAE_PT_FOR_PAE_PT:
5465	return "PGMPOOLKIND_PAE_PT_FOR_PAE_PT";
5466	case PGMPOOLKIND_PAE_PT_FOR_PAE_2MB:
5467	return "PGMPOOLKIND_PAE_PT_FOR_PAE_2MB";
5468	case PGMPOOLKIND_32BIT_PD:
5469	return "PGMPOOLKIND_32BIT_PD";
5470	case PGMPOOLKIND_32BIT_PD_PHYS:
5471	return "PGMPOOLKIND_32BIT_PD_PHYS";
5472	case PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD:
5473	return "PGMPOOLKIND_PAE_PD0_FOR_32BIT_PD";
5474	case PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD:
5475	return "PGMPOOLKIND_PAE_PD1_FOR_32BIT_PD";
5476	case PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD:
5477	return "PGMPOOLKIND_PAE_PD2_FOR_32BIT_PD";
5478	case PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD:
5479	return "PGMPOOLKIND_PAE_PD3_FOR_32BIT_PD";
5480	case PGMPOOLKIND_PAE_PD_FOR_PAE_PD:
5481	return "PGMPOOLKIND_PAE_PD_FOR_PAE_PD";
5482	case PGMPOOLKIND_PAE_PD_PHYS:
5483	return "PGMPOOLKIND_PAE_PD_PHYS";
5484	case PGMPOOLKIND_PAE_PDPT_FOR_32BIT:
5485	return "PGMPOOLKIND_PAE_PDPT_FOR_32BIT";
5486	case PGMPOOLKIND_PAE_PDPT:
5487	return "PGMPOOLKIND_PAE_PDPT";
5488	case PGMPOOLKIND_PAE_PDPT_PHYS:
5489	return "PGMPOOLKIND_PAE_PDPT_PHYS";
5490	case PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT:
5491	return "PGMPOOLKIND_64BIT_PDPT_FOR_64BIT_PDPT";
5492	case PGMPOOLKIND_64BIT_PDPT_FOR_PHYS:
5493	return "PGMPOOLKIND_64BIT_PDPT_FOR_PHYS";
5494	case PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD:
5495	return "PGMPOOLKIND_64BIT_PD_FOR_64BIT_PD";
5496	case PGMPOOLKIND_64BIT_PD_FOR_PHYS:
5497	return "PGMPOOLKIND_64BIT_PD_FOR_PHYS";
5498	case PGMPOOLKIND_64BIT_PML4:
5499	return "PGMPOOLKIND_64BIT_PML4";
5500	case PGMPOOLKIND_EPT_PDPT_FOR_PHYS:
5501	return "PGMPOOLKIND_EPT_PDPT_FOR_PHYS";
5502	case PGMPOOLKIND_EPT_PD_FOR_PHYS:
5503	return "PGMPOOLKIND_EPT_PD_FOR_PHYS";
5504	case PGMPOOLKIND_EPT_PT_FOR_PHYS:
5505	return "PGMPOOLKIND_EPT_PT_FOR_PHYS";
5506	case PGMPOOLKIND_ROOT_NESTED:
5507	return "PGMPOOLKIND_ROOT_NESTED";
5508	}
5509	return "Unknown kind!";
5510	}
5511	#endif /* LOG_ENABLED*/
5512

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

source: vbox/trunk/src/VBox/VMM/VMMAll/PGMAllPool.cpp@ 41801

Download in other formats: