VirtualBox

About
Screenshots
Downloads
Documentation
End-user docs
Technical docs
Contribute
Community

Browse Source

Context Navigation

source: vbox/trunk/src/VBox/VMM/VMMR3/PGMPhys.cpp@ 84044

Last change on this file since 84044 was 82968, checked in by vboxsync, 4 years ago
Copyright year updates by scm.
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 199.4 KB

Line
1	/* $Id: PGMPhys.cpp 82968 2020-02-04 10:35:17Z vboxsync $ */
2	/** @file
3	* PGM - Page Manager and Monitor, Physical Memory Addressing.
4	*/
5
6	/*
7	* Copyright (C) 2006-2020 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_PHYS
23	#include <VBox/vmm/pgm.h>
24	#include <VBox/vmm/iem.h>
25	#include <VBox/vmm/iom.h>
26	#include <VBox/vmm/mm.h>
27	#include <VBox/vmm/nem.h>
28	#include <VBox/vmm/stam.h>
29	#include <VBox/vmm/pdmdev.h>
30	#include "PGMInternal.h"
31	#include <VBox/vmm/vmcc.h>
32
33	#include "PGMInline.h"
34
35	#include <VBox/sup.h>
36	#include <VBox/param.h>
37	#include <VBox/err.h>
38	#include <VBox/log.h>
39	#include <iprt/assert.h>
40	#include <iprt/alloc.h>
41	#include <iprt/asm.h>
42	#ifdef VBOX_STRICT
43	# include <iprt/crc.h>
44	#endif
45	#include <iprt/thread.h>
46	#include <iprt/string.h>
47	#include <iprt/system.h>
48
49
50	/*********************************************************************************************************************************
51	* Defined Constants And Macros *
52	*********************************************************************************************************************************/
53	/** The number of pages to free in one batch. */
54	#define PGMPHYS_FREE_PAGE_BATCH_SIZE 128
55
56
57	/*
58	* PGMR3PhysReadU8-64
59	* PGMR3PhysWriteU8-64
60	*/
61	#define PGMPHYSFN_READNAME PGMR3PhysReadU8
62	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU8
63	#define PGMPHYS_DATASIZE 1
64	#define PGMPHYS_DATATYPE uint8_t
65	#include "PGMPhysRWTmpl.h"
66
67	#define PGMPHYSFN_READNAME PGMR3PhysReadU16
68	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU16
69	#define PGMPHYS_DATASIZE 2
70	#define PGMPHYS_DATATYPE uint16_t
71	#include "PGMPhysRWTmpl.h"
72
73	#define PGMPHYSFN_READNAME PGMR3PhysReadU32
74	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU32
75	#define PGMPHYS_DATASIZE 4
76	#define PGMPHYS_DATATYPE uint32_t
77	#include "PGMPhysRWTmpl.h"
78
79	#define PGMPHYSFN_READNAME PGMR3PhysReadU64
80	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU64
81	#define PGMPHYS_DATASIZE 8
82	#define PGMPHYS_DATATYPE uint64_t
83	#include "PGMPhysRWTmpl.h"
84
85
86	/**
87	* EMT worker for PGMR3PhysReadExternal.
88	*/
89	static DECLCALLBACK(int) pgmR3PhysReadExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, void *pvBuf, size_t cbRead,
90	PGMACCESSORIGIN enmOrigin)
91	{
92	VBOXSTRICTRC rcStrict = PGMPhysRead(pVM, *pGCPhys, pvBuf, cbRead, enmOrigin);
93	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
94	return VINF_SUCCESS;
95	}
96
97
98	/**
99	* Read from physical memory, external users.
100	*
101	* @returns VBox status code.
102	* @retval VINF_SUCCESS.
103	*
104	* @param pVM The cross context VM structure.
105	* @param GCPhys Physical address to read from.
106	* @param pvBuf Where to read into.
107	* @param cbRead How many bytes to read.
108	* @param enmOrigin Who is calling.
109	*
110	* @thread Any but EMTs.
111	*/
112	VMMR3DECL(int) PGMR3PhysReadExternal(PVM pVM, RTGCPHYS GCPhys, void *pvBuf, size_t cbRead, PGMACCESSORIGIN enmOrigin)
113	{
114	VM_ASSERT_OTHER_THREAD(pVM);
115
116	AssertMsgReturn(cbRead > 0, ("don't even think about reading zero bytes!\n"), VINF_SUCCESS);
117	LogFlow(("PGMR3PhysReadExternal: %RGp %d\n", GCPhys, cbRead));
118
119	pgmLock(pVM);
120
121	/*
122	* Copy loop on ram ranges.
123	*/
124	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
125	for (;;)
126	{
127	/* Inside range or not? */
128	if (pRam && GCPhys >= pRam->GCPhys)
129	{
130	/*
131	* Must work our way thru this page by page.
132	*/
133	RTGCPHYS off = GCPhys - pRam->GCPhys;
134	while (off < pRam->cb)
135	{
136	unsigned iPage = off >> PAGE_SHIFT;
137	PPGMPAGE pPage = &pRam->aPages[iPage];
138
139	/*
140	* If the page has an ALL access handler, we'll have to
141	* delegate the job to EMT.
142	*/
143	if ( PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
144	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
145	{
146	pgmUnlock(pVM);
147
148	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysReadExternalEMT, 5,
149	pVM, &GCPhys, pvBuf, cbRead, enmOrigin);
150	}
151	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
152
153	/*
154	* Simple stuff, go ahead.
155	*/
156	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
157	if (cb > cbRead)
158	cb = cbRead;
159	PGMPAGEMAPLOCK PgMpLck;
160	const void *pvSrc;
161	int rc = pgmPhysGCPhys2CCPtrInternalReadOnly(pVM, pPage, pRam->GCPhys + off, &pvSrc, &PgMpLck);
162	if (RT_SUCCESS(rc))
163	{
164	memcpy(pvBuf, pvSrc, cb);
165	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
166	}
167	else
168	{
169	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternalReadOnly failed on %RGp / %R[pgmpage] -> %Rrc\n",
170	pRam->GCPhys + off, pPage, rc));
171	memset(pvBuf, 0xff, cb);
172	}
173
174	/* next page */
175	if (cb >= cbRead)
176	{
177	pgmUnlock(pVM);
178	return VINF_SUCCESS;
179	}
180	cbRead -= cb;
181	off += cb;
182	GCPhys += cb;
183	pvBuf = (char *)pvBuf + cb;
184	} /* walk pages in ram range. */
185	}
186	else
187	{
188	LogFlow(("PGMPhysRead: Unassigned %RGp size=%u\n", GCPhys, cbRead));
189
190	/*
191	* Unassigned address space.
192	*/
193	size_t cb = pRam ? pRam->GCPhys - GCPhys : ~(size_t)0;
194	if (cb >= cbRead)
195	{
196	memset(pvBuf, 0xff, cbRead);
197	break;
198	}
199	memset(pvBuf, 0xff, cb);
200
201	cbRead -= cb;
202	pvBuf = (char *)pvBuf + cb;
203	GCPhys += cb;
204	}
205
206	/* Advance range if necessary. */
207	while (pRam && GCPhys > pRam->GCPhysLast)
208	pRam = pRam->CTX_SUFF(pNext);
209	} /* Ram range walk */
210
211	pgmUnlock(pVM);
212
213	return VINF_SUCCESS;
214	}
215
216
217	/**
218	* EMT worker for PGMR3PhysWriteExternal.
219	*/
220	static DECLCALLBACK(int) pgmR3PhysWriteExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, const void *pvBuf, size_t cbWrite,
221	PGMACCESSORIGIN enmOrigin)
222	{
223	/** @todo VERR_EM_NO_MEMORY */
224	VBOXSTRICTRC rcStrict = PGMPhysWrite(pVM, *pGCPhys, pvBuf, cbWrite, enmOrigin);
225	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
226	return VINF_SUCCESS;
227	}
228
229
230	/**
231	* Write to physical memory, external users.
232	*
233	* @returns VBox status code.
234	* @retval VINF_SUCCESS.
235	* @retval VERR_EM_NO_MEMORY.
236	*
237	* @param pVM The cross context VM structure.
238	* @param GCPhys Physical address to write to.
239	* @param pvBuf What to write.
240	* @param cbWrite How many bytes to write.
241	* @param enmOrigin Who is calling.
242	*
243	* @thread Any but EMTs.
244	*/
245	VMMDECL(int) PGMR3PhysWriteExternal(PVM pVM, RTGCPHYS GCPhys, const void *pvBuf, size_t cbWrite, PGMACCESSORIGIN enmOrigin)
246	{
247	VM_ASSERT_OTHER_THREAD(pVM);
248
249	AssertMsg(!pVM->pgm.s.fNoMorePhysWrites,
250	("Calling PGMR3PhysWriteExternal after pgmR3Save()! GCPhys=%RGp cbWrite=%#x enmOrigin=%d\n",
251	GCPhys, cbWrite, enmOrigin));
252	AssertMsgReturn(cbWrite > 0, ("don't even think about writing zero bytes!\n"), VINF_SUCCESS);
253	LogFlow(("PGMR3PhysWriteExternal: %RGp %d\n", GCPhys, cbWrite));
254
255	pgmLock(pVM);
256
257	/*
258	* Copy loop on ram ranges, stop when we hit something difficult.
259	*/
260	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
261	for (;;)
262	{
263	/* Inside range or not? */
264	if (pRam && GCPhys >= pRam->GCPhys)
265	{
266	/*
267	* Must work our way thru this page by page.
268	*/
269	RTGCPTR off = GCPhys - pRam->GCPhys;
270	while (off < pRam->cb)
271	{
272	RTGCPTR iPage = off >> PAGE_SHIFT;
273	PPGMPAGE pPage = &pRam->aPages[iPage];
274
275	/*
276	* Is the page problematic, we have to do the work on the EMT.
277	*
278	* Allocating writable pages and access handlers are
279	* problematic, write monitored pages are simple and can be
280	* dealt with here.
281	*/
282	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
283	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
284	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
285	{
286	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
287	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
288	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, GCPhys);
289	else
290	{
291	pgmUnlock(pVM);
292
293	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysWriteExternalEMT, 5,
294	pVM, &GCPhys, pvBuf, cbWrite, enmOrigin);
295	}
296	}
297	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
298
299	/*
300	* Simple stuff, go ahead.
301	*/
302	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
303	if (cb > cbWrite)
304	cb = cbWrite;
305	PGMPAGEMAPLOCK PgMpLck;
306	void *pvDst;
307	int rc = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, pRam->GCPhys + off, &pvDst, &PgMpLck);
308	if (RT_SUCCESS(rc))
309	{
310	memcpy(pvDst, pvBuf, cb);
311	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
312	}
313	else
314	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternal failed on %RGp / %R[pgmpage] -> %Rrc\n",
315	pRam->GCPhys + off, pPage, rc));
316
317	/* next page */
318	if (cb >= cbWrite)
319	{
320	pgmUnlock(pVM);
321	return VINF_SUCCESS;
322	}
323
324	cbWrite -= cb;
325	off += cb;
326	GCPhys += cb;
327	pvBuf = (const char *)pvBuf + cb;
328	} /* walk pages in ram range */
329	}
330	else
331	{
332	/*
333	* Unassigned address space, skip it.
334	*/
335	if (!pRam)
336	break;
337	size_t cb = pRam->GCPhys - GCPhys;
338	if (cb >= cbWrite)
339	break;
340	cbWrite -= cb;
341	pvBuf = (const char *)pvBuf + cb;
342	GCPhys += cb;
343	}
344
345	/* Advance range if necessary. */
346	while (pRam && GCPhys > pRam->GCPhysLast)
347	pRam = pRam->CTX_SUFF(pNext);
348	} /* Ram range walk */
349
350	pgmUnlock(pVM);
351	return VINF_SUCCESS;
352	}
353
354
355	/**
356	* VMR3ReqCall worker for PGMR3PhysGCPhys2CCPtrExternal to make pages writable.
357	*
358	* @returns see PGMR3PhysGCPhys2CCPtrExternal
359	* @param pVM The cross context VM structure.
360	* @param pGCPhys Pointer to the guest physical address.
361	* @param ppv Where to store the mapping address.
362	* @param pLock Where to store the lock.
363	*/
364	static DECLCALLBACK(int) pgmR3PhysGCPhys2CCPtrDelegated(PVM pVM, PRTGCPHYS pGCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
365	{
366	/*
367	* Just hand it to PGMPhysGCPhys2CCPtr and check that it's not a page with
368	* an access handler after it succeeds.
369	*/
370	int rc = pgmLock(pVM);
371	AssertRCReturn(rc, rc);
372
373	rc = PGMPhysGCPhys2CCPtr(pVM, *pGCPhys, ppv, pLock);
374	if (RT_SUCCESS(rc))
375	{
376	PPGMPAGEMAPTLBE pTlbe;
377	int rc2 = pgmPhysPageQueryTlbe(pVM, *pGCPhys, &pTlbe);
378	AssertFatalRC(rc2);
379	PPGMPAGE pPage = pTlbe->pPage;
380	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
381	{
382	PGMPhysReleasePageMappingLock(pVM, pLock);
383	rc = VERR_PGM_PHYS_PAGE_RESERVED;
384	}
385	else if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
386	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
387	\|\| pgmPoolIsDirtyPage(pVM, *pGCPhys)
388	#endif
389	)
390	{
391	/* We must flush any corresponding pgm pool page here, otherwise we'll
392	* not be informed about writes and keep bogus gst->shw mappings around.
393	*/
394	pgmPoolFlushPageByGCPhys(pVM, *pGCPhys);
395	Assert(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage));
396	/** @todo r=bird: return VERR_PGM_PHYS_PAGE_RESERVED here if it still has
397	* active handlers, see the PGMR3PhysGCPhys2CCPtrExternal docs. */
398	}
399	}
400
401	pgmUnlock(pVM);
402	return rc;
403	}
404
405
406	/**
407	* Requests the mapping of a guest page into ring-3, external threads.
408	*
409	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
410	* release it.
411	*
412	* This API will assume your intention is to write to the page, and will
413	* therefore replace shared and zero pages. If you do not intend to modify the
414	* page, use the PGMR3PhysGCPhys2CCPtrReadOnlyExternal() API.
415	*
416	* @returns VBox status code.
417	* @retval VINF_SUCCESS on success.
418	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
419	* backing or if the page has any active access handlers. The caller
420	* must fall back on using PGMR3PhysWriteExternal.
421	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
422	*
423	* @param pVM The cross context VM structure.
424	* @param GCPhys The guest physical address of the page that should be mapped.
425	* @param ppv Where to store the address corresponding to GCPhys.
426	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
427	*
428	* @remark Avoid calling this API from within critical sections (other than the
429	* PGM one) because of the deadlock risk when we have to delegating the
430	* task to an EMT.
431	* @thread Any.
432	*/
433	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrExternal(PVM pVM, RTGCPHYS GCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
434	{
435	AssertPtr(ppv);
436	AssertPtr(pLock);
437
438	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
439
440	int rc = pgmLock(pVM);
441	AssertRCReturn(rc, rc);
442
443	/*
444	* Query the Physical TLB entry for the page (may fail).
445	*/
446	PPGMPAGEMAPTLBE pTlbe;
447	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
448	if (RT_SUCCESS(rc))
449	{
450	PPGMPAGE pPage = pTlbe->pPage;
451	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
452	rc = VERR_PGM_PHYS_PAGE_RESERVED;
453	else
454	{
455	/*
456	* If the page is shared, the zero page, or being write monitored
457	* it must be converted to an page that's writable if possible.
458	* We can only deal with write monitored pages here, the rest have
459	* to be on an EMT.
460	*/
461	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
462	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
463	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
464	\|\| pgmPoolIsDirtyPage(pVM, GCPhys)
465	#endif
466	)
467	{
468	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
469	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
470	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
471	&& !pgmPoolIsDirtyPage(pVM, GCPhys) /** @todo we're very likely doing this twice. */
472	#endif
473	)
474	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, GCPhys);
475	else
476	{
477	pgmUnlock(pVM);
478
479	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
480	pVM, &GCPhys, ppv, pLock);
481	}
482	}
483
484	/*
485	* Now, just perform the locking and calculate the return address.
486	*/
487	PPGMPAGEMAP pMap = pTlbe->pMap;
488	if (pMap)
489	pMap->cRefs++;
490
491	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
492	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
493	{
494	if (cLocks == 0)
495	pVM->pgm.s.cWriteLockedPages++;
496	PGM_PAGE_INC_WRITE_LOCKS(pPage);
497	}
498	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
499	{
500	PGM_PAGE_INC_WRITE_LOCKS(pPage);
501	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", GCPhys, pPage));
502	if (pMap)
503	pMap->cRefs++; /* Extra ref to prevent it from going away. */
504	}
505
506	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
507	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
508	pLock->pvMap = pMap;
509	}
510	}
511
512	pgmUnlock(pVM);
513	return rc;
514	}
515
516
517	/**
518	* Requests the mapping of a guest page into ring-3, external threads.
519	*
520	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
521	* release it.
522	*
523	* @returns VBox status code.
524	* @retval VINF_SUCCESS on success.
525	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
526	* backing or if the page as an active ALL access handler. The caller
527	* must fall back on using PGMPhysRead.
528	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
529	*
530	* @param pVM The cross context VM structure.
531	* @param GCPhys The guest physical address of the page that should be mapped.
532	* @param ppv Where to store the address corresponding to GCPhys.
533	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
534	*
535	* @remark Avoid calling this API from within critical sections (other than
536	* the PGM one) because of the deadlock risk.
537	* @thread Any.
538	*/
539	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrReadOnlyExternal(PVM pVM, RTGCPHYS GCPhys, void const **ppv, PPGMPAGEMAPLOCK pLock)
540	{
541	int rc = pgmLock(pVM);
542	AssertRCReturn(rc, rc);
543
544	/*
545	* Query the Physical TLB entry for the page (may fail).
546	*/
547	PPGMPAGEMAPTLBE pTlbe;
548	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
549	if (RT_SUCCESS(rc))
550	{
551	PPGMPAGE pPage = pTlbe->pPage;
552	#if 1
553	/* MMIO pages doesn't have any readable backing. */
554	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
555	rc = VERR_PGM_PHYS_PAGE_RESERVED;
556	#else
557	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
558	rc = VERR_PGM_PHYS_PAGE_RESERVED;
559	#endif
560	else
561	{
562	/*
563	* Now, just perform the locking and calculate the return address.
564	*/
565	PPGMPAGEMAP pMap = pTlbe->pMap;
566	if (pMap)
567	pMap->cRefs++;
568
569	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
570	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
571	{
572	if (cLocks == 0)
573	pVM->pgm.s.cReadLockedPages++;
574	PGM_PAGE_INC_READ_LOCKS(pPage);
575	}
576	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
577	{
578	PGM_PAGE_INC_READ_LOCKS(pPage);
579	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", GCPhys, pPage));
580	if (pMap)
581	pMap->cRefs++; /* Extra ref to prevent it from going away. */
582	}
583
584	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
585	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
586	pLock->pvMap = pMap;
587	}
588	}
589
590	pgmUnlock(pVM);
591	return rc;
592	}
593
594
595	/**
596	* Requests the mapping of multiple guest page into ring-3, external threads.
597	*
598	* When you're done with the pages, call PGMPhysBulkReleasePageMappingLock()
599	* ASAP to release them.
600	*
601	* This API will assume your intention is to write to the pages, and will
602	* therefore replace shared and zero pages. If you do not intend to modify the
603	* pages, use the PGMR3PhysBulkGCPhys2CCPtrReadOnlyExternal() API.
604	*
605	* @returns VBox status code.
606	* @retval VINF_SUCCESS on success.
607	* @retval VERR_PGM_PHYS_PAGE_RESERVED if any of the pages has no physical
608	* backing or if any of the pages the page has any active access
609	* handlers. The caller must fall back on using PGMR3PhysWriteExternal.
610	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if @a paGCPhysPages contains
611	* an invalid physical address.
612	*
613	* @param pVM The cross context VM structure.
614	* @param cPages Number of pages to lock.
615	* @param paGCPhysPages The guest physical address of the pages that
616	* should be mapped (@a cPages entries).
617	* @param papvPages Where to store the ring-3 mapping addresses
618	* corresponding to @a paGCPhysPages.
619	* @param paLocks Where to store the locking information that
620	* pfnPhysBulkReleasePageMappingLock needs (@a cPages
621	* in length).
622	*
623	* @remark Avoid calling this API from within critical sections (other than the
624	* PGM one) because of the deadlock risk when we have to delegating the
625	* task to an EMT.
626	* @thread Any.
627	*/
628	VMMR3DECL(int) PGMR3PhysBulkGCPhys2CCPtrExternal(PVM pVM, uint32_t cPages, PCRTGCPHYS paGCPhysPages,
629	void **papvPages, PPGMPAGEMAPLOCK paLocks)
630	{
631	Assert(cPages > 0);
632	AssertPtr(papvPages);
633	AssertPtr(paLocks);
634
635	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
636
637	int rc = pgmLock(pVM);
638	AssertRCReturn(rc, rc);
639
640	/*
641	* Lock the pages one by one.
642	* The loop body is similar to PGMR3PhysGCPhys2CCPtrExternal.
643	*/
644	int32_t cNextYield = 128;
645	uint32_t iPage;
646	for (iPage = 0; iPage < cPages; iPage++)
647	{
648	if (--cNextYield > 0)
649	{ /* likely */ }
650	else
651	{
652	pgmUnlock(pVM);
653	ASMNopPause();
654	pgmLock(pVM);
655	cNextYield = 128;
656	}
657
658	/*
659	* Query the Physical TLB entry for the page (may fail).
660	*/
661	PPGMPAGEMAPTLBE pTlbe;
662	rc = pgmPhysPageQueryTlbe(pVM, paGCPhysPages[iPage], &pTlbe);
663	if (RT_SUCCESS(rc))
664	{ }
665	else
666	break;
667	PPGMPAGE pPage = pTlbe->pPage;
668
669	/*
670	* No MMIO or active access handlers.
671	*/
672	if ( !PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage)
673	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
674	{ }
675	else
676	{
677	rc = VERR_PGM_PHYS_PAGE_RESERVED;
678	break;
679	}
680
681	/*
682	* The page must be in the allocated state and not be a dirty pool page.
683	* We can handle converting a write monitored page to an allocated one, but
684	* anything more complicated must be delegated to an EMT.
685	*/
686	bool fDelegateToEmt = false;
687	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_ALLOCATED)
688	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
689	fDelegateToEmt = pgmPoolIsDirtyPage(pVM, paGCPhysPages[iPage]);
690	#else
691	fDelegateToEmt = false;
692	#endif
693	else if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
694	{
695	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
696	if (!pgmPoolIsDirtyPage(pVM, paGCPhysPages[iPage]))
697	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, paGCPhysPages[iPage]);
698	else
699	fDelegateToEmt = true;
700	#endif
701	}
702	else
703	fDelegateToEmt = true;
704	if (!fDelegateToEmt)
705	{ }
706	else
707	{
708	/* We could do this delegation in bulk, but considered too much work vs gain. */
709	pgmUnlock(pVM);
710	rc = VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
711	pVM, &paGCPhysPages[iPage], &papvPages[iPage], &paLocks[iPage]);
712	pgmLock(pVM);
713	if (RT_FAILURE(rc))
714	break;
715	cNextYield = 128;
716	}
717
718	/*
719	* Now, just perform the locking and address calculation.
720	*/
721	PPGMPAGEMAP pMap = pTlbe->pMap;
722	if (pMap)
723	pMap->cRefs++;
724
725	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
726	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
727	{
728	if (cLocks == 0)
729	pVM->pgm.s.cWriteLockedPages++;
730	PGM_PAGE_INC_WRITE_LOCKS(pPage);
731	}
732	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
733	{
734	PGM_PAGE_INC_WRITE_LOCKS(pPage);
735	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", paGCPhysPages[iPage], pPage));
736	if (pMap)
737	pMap->cRefs++; /* Extra ref to prevent it from going away. */
738	}
739
740	papvPages[iPage] = (void *)((uintptr_t)pTlbe->pv \| (uintptr_t)(paGCPhysPages[iPage] & PAGE_OFFSET_MASK));
741	paLocks[iPage].uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
742	paLocks[iPage].pvMap = pMap;
743	}
744
745	pgmUnlock(pVM);
746
747	/*
748	* On failure we must unlock any pages we managed to get already.
749	*/
750	if (RT_FAILURE(rc) && iPage > 0)
751	PGMPhysBulkReleasePageMappingLocks(pVM, iPage, paLocks);
752
753	return rc;
754	}
755
756
757	/**
758	* Requests the mapping of multiple guest page into ring-3, for reading only,
759	* external threads.
760	*
761	* When you're done with the pages, call PGMPhysReleasePageMappingLock() ASAP
762	* to release them.
763	*
764	* @returns VBox status code.
765	* @retval VINF_SUCCESS on success.
766	* @retval VERR_PGM_PHYS_PAGE_RESERVED if any of the pages has no physical
767	* backing or if any of the pages the page has an active ALL access
768	* handler. The caller must fall back on using PGMR3PhysWriteExternal.
769	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if @a paGCPhysPages contains
770	* an invalid physical address.
771	*
772	* @param pVM The cross context VM structure.
773	* @param cPages Number of pages to lock.
774	* @param paGCPhysPages The guest physical address of the pages that
775	* should be mapped (@a cPages entries).
776	* @param papvPages Where to store the ring-3 mapping addresses
777	* corresponding to @a paGCPhysPages.
778	* @param paLocks Where to store the lock information that
779	* pfnPhysReleasePageMappingLock needs (@a cPages
780	* in length).
781	*
782	* @remark Avoid calling this API from within critical sections (other than
783	* the PGM one) because of the deadlock risk.
784	* @thread Any.
785	*/
786	VMMR3DECL(int) PGMR3PhysBulkGCPhys2CCPtrReadOnlyExternal(PVM pVM, uint32_t cPages, PCRTGCPHYS paGCPhysPages,
787	void const **papvPages, PPGMPAGEMAPLOCK paLocks)
788	{
789	Assert(cPages > 0);
790	AssertPtr(papvPages);
791	AssertPtr(paLocks);
792
793	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
794
795	int rc = pgmLock(pVM);
796	AssertRCReturn(rc, rc);
797
798	/*
799	* Lock the pages one by one.
800	* The loop body is similar to PGMR3PhysGCPhys2CCPtrReadOnlyExternal.
801	*/
802	int32_t cNextYield = 256;
803	uint32_t iPage;
804	for (iPage = 0; iPage < cPages; iPage++)
805	{
806	if (--cNextYield > 0)
807	{ /* likely */ }
808	else
809	{
810	pgmUnlock(pVM);
811	ASMNopPause();
812	pgmLock(pVM);
813	cNextYield = 256;
814	}
815
816	/*
817	* Query the Physical TLB entry for the page (may fail).
818	*/
819	PPGMPAGEMAPTLBE pTlbe;
820	rc = pgmPhysPageQueryTlbe(pVM, paGCPhysPages[iPage], &pTlbe);
821	if (RT_SUCCESS(rc))
822	{ }
823	else
824	break;
825	PPGMPAGE pPage = pTlbe->pPage;
826
827	/*
828	* No MMIO or active all access handlers, everything else can be accessed.
829	*/
830	if ( !PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage)
831	&& !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
832	{ }
833	else
834	{
835	rc = VERR_PGM_PHYS_PAGE_RESERVED;
836	break;
837	}
838
839	/*
840	* Now, just perform the locking and address calculation.
841	*/
842	PPGMPAGEMAP pMap = pTlbe->pMap;
843	if (pMap)
844	pMap->cRefs++;
845
846	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
847	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
848	{
849	if (cLocks == 0)
850	pVM->pgm.s.cReadLockedPages++;
851	PGM_PAGE_INC_READ_LOCKS(pPage);
852	}
853	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
854	{
855	PGM_PAGE_INC_READ_LOCKS(pPage);
856	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", paGCPhysPages[iPage], pPage));
857	if (pMap)
858	pMap->cRefs++; /* Extra ref to prevent it from going away. */
859	}
860
861	papvPages[iPage] = (void *)((uintptr_t)pTlbe->pv \| (uintptr_t)(paGCPhysPages[iPage] & PAGE_OFFSET_MASK));
862	paLocks[iPage].uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
863	paLocks[iPage].pvMap = pMap;
864	}
865
866	pgmUnlock(pVM);
867
868	/*
869	* On failure we must unlock any pages we managed to get already.
870	*/
871	if (RT_FAILURE(rc) && iPage > 0)
872	PGMPhysBulkReleasePageMappingLocks(pVM, iPage, paLocks);
873
874	return rc;
875	}
876
877
878	#define MAKE_LEAF(a_pNode) \
879	do { \
880	(a_pNode)->pLeftR3 = NIL_RTR3PTR; \
881	(a_pNode)->pRightR3 = NIL_RTR3PTR; \
882	(a_pNode)->pLeftR0 = NIL_RTR0PTR; \
883	(a_pNode)->pRightR0 = NIL_RTR0PTR; \
884	} while (0)
885
886	#define INSERT_LEFT(a_pParent, a_pNode) \
887	do { \
888	(a_pParent)->pLeftR3 = (a_pNode); \
889	(a_pParent)->pLeftR0 = (a_pNode)->pSelfR0; \
890	} while (0)
891	#define INSERT_RIGHT(a_pParent, a_pNode) \
892	do { \
893	(a_pParent)->pRightR3 = (a_pNode); \
894	(a_pParent)->pRightR0 = (a_pNode)->pSelfR0; \
895	} while (0)
896
897
898	/**
899	* Recursive tree builder.
900	*
901	* @param ppRam Pointer to the iterator variable.
902	* @param iDepth The current depth. Inserts a leaf node if 0.
903	*/
904	static PPGMRAMRANGE pgmR3PhysRebuildRamRangeSearchTreesRecursively(PPGMRAMRANGE *ppRam, int iDepth)
905	{
906	PPGMRAMRANGE pRam;
907	if (iDepth <= 0)
908	{
909	/*
910	* Leaf node.
911	*/
912	pRam = *ppRam;
913	if (pRam)
914	{
915	*ppRam = pRam->pNextR3;
916	MAKE_LEAF(pRam);
917	}
918	}
919	else
920	{
921
922	/*
923	* Intermediate node.
924	*/
925	PPGMRAMRANGE pLeft = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
926
927	pRam = *ppRam;
928	if (!pRam)
929	return pLeft;
930	*ppRam = pRam->pNextR3;
931	MAKE_LEAF(pRam);
932	INSERT_LEFT(pRam, pLeft);
933
934	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
935	if (pRight)
936	INSERT_RIGHT(pRam, pRight);
937	}
938	return pRam;
939	}
940
941
942	/**
943	* Rebuilds the RAM range search trees.
944	*
945	* @param pVM The cross context VM structure.
946	*/
947	static void pgmR3PhysRebuildRamRangeSearchTrees(PVM pVM)
948	{
949
950	/*
951	* Create the reasonably balanced tree in a sequential fashion.
952	* For simplicity (laziness) we use standard recursion here.
953	*/
954	int iDepth = 0;
955	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
956	PPGMRAMRANGE pRoot = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, 0);
957	while (pRam)
958	{
959	PPGMRAMRANGE pLeft = pRoot;
960
961	pRoot = pRam;
962	pRam = pRam->pNextR3;
963	MAKE_LEAF(pRoot);
964	INSERT_LEFT(pRoot, pLeft);
965
966	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, iDepth);
967	if (pRight)
968	INSERT_RIGHT(pRoot, pRight);
969	/** @todo else: rotate the tree. */
970
971	iDepth++;
972	}
973
974	pVM->pgm.s.pRamRangeTreeR3 = pRoot;
975	pVM->pgm.s.pRamRangeTreeR0 = pRoot ? pRoot->pSelfR0 : NIL_RTR0PTR;
976
977	#ifdef VBOX_STRICT
978	/*
979	* Verify that the above code works.
980	*/
981	unsigned cRanges = 0;
982	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
983	cRanges++;
984	Assert(cRanges > 0);
985
986	unsigned cMaxDepth = ASMBitLastSetU32(cRanges);
987	if ((1U << cMaxDepth) < cRanges)
988	cMaxDepth++;
989
990	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
991	{
992	unsigned cDepth = 0;
993	PPGMRAMRANGE pRam2 = pVM->pgm.s.pRamRangeTreeR3;
994	for (;;)
995	{
996	if (pRam == pRam2)
997	break;
998	Assert(pRam2);
999	if (pRam->GCPhys < pRam2->GCPhys)
1000	pRam2 = pRam2->pLeftR3;
1001	else
1002	pRam2 = pRam2->pRightR3;
1003	}
1004	AssertMsg(cDepth <= cMaxDepth, ("cDepth=%d cMaxDepth=%d\n", cDepth, cMaxDepth));
1005	}
1006	#endif /* VBOX_STRICT */
1007	}
1008
1009	#undef MAKE_LEAF
1010	#undef INSERT_LEFT
1011	#undef INSERT_RIGHT
1012
1013	/**
1014	* Relinks the RAM ranges using the pSelfRC and pSelfR0 pointers.
1015	*
1016	* Called when anything was relocated.
1017	*
1018	* @param pVM The cross context VM structure.
1019	*/
1020	void pgmR3PhysRelinkRamRanges(PVM pVM)
1021	{
1022	PPGMRAMRANGE pCur;
1023
1024	#ifdef VBOX_STRICT
1025	for (pCur = pVM->pgm.s.pRamRangesXR3; pCur; pCur = pCur->pNextR3)
1026	{
1027	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfR0 == MMHyperCCToR0(pVM, pCur));
1028	Assert((pCur->GCPhys & PAGE_OFFSET_MASK) == 0);
1029	Assert((pCur->GCPhysLast & PAGE_OFFSET_MASK) == PAGE_OFFSET_MASK);
1030	Assert((pCur->cb & PAGE_OFFSET_MASK) == 0);
1031	Assert(pCur->cb == pCur->GCPhysLast - pCur->GCPhys + 1);
1032	for (PPGMRAMRANGE pCur2 = pVM->pgm.s.pRamRangesXR3; pCur2; pCur2 = pCur2->pNextR3)
1033	Assert( pCur2 == pCur
1034	\|\| strcmp(pCur2->pszDesc, pCur->pszDesc)); /** @todo fix MMIO ranges!! */
1035	}
1036	#endif
1037
1038	pCur = pVM->pgm.s.pRamRangesXR3;
1039	if (pCur)
1040	{
1041	pVM->pgm.s.pRamRangesXR0 = pCur->pSelfR0;
1042
1043	for (; pCur->pNextR3; pCur = pCur->pNextR3)
1044	pCur->pNextR0 = pCur->pNextR3->pSelfR0;
1045
1046	Assert(pCur->pNextR0 == NIL_RTR0PTR);
1047	}
1048	else
1049	{
1050	Assert(pVM->pgm.s.pRamRangesXR0 == NIL_RTR0PTR);
1051	}
1052	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1053
1054	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1055	}
1056
1057
1058	/**
1059	* Links a new RAM range into the list.
1060	*
1061	* @param pVM The cross context VM structure.
1062	* @param pNew Pointer to the new list entry.
1063	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
1064	*/
1065	static void pgmR3PhysLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, PPGMRAMRANGE pPrev)
1066	{
1067	AssertMsg(pNew->pszDesc, ("%RGp-%RGp\n", pNew->GCPhys, pNew->GCPhysLast));
1068	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfR0 == MMHyperCCToR0(pVM, pNew));
1069
1070	pgmLock(pVM);
1071
1072	PPGMRAMRANGE pRam = pPrev ? pPrev->pNextR3 : pVM->pgm.s.pRamRangesXR3;
1073	pNew->pNextR3 = pRam;
1074	pNew->pNextR0 = pRam ? pRam->pSelfR0 : NIL_RTR0PTR;
1075
1076	if (pPrev)
1077	{
1078	pPrev->pNextR3 = pNew;
1079	pPrev->pNextR0 = pNew->pSelfR0;
1080	}
1081	else
1082	{
1083	pVM->pgm.s.pRamRangesXR3 = pNew;
1084	pVM->pgm.s.pRamRangesXR0 = pNew->pSelfR0;
1085	}
1086	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1087
1088	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1089	pgmUnlock(pVM);
1090	}
1091
1092
1093	/**
1094	* Unlink an existing RAM range from the list.
1095	*
1096	* @param pVM The cross context VM structure.
1097	* @param pRam Pointer to the new list entry.
1098	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
1099	*/
1100	static void pgmR3PhysUnlinkRamRange2(PVM pVM, PPGMRAMRANGE pRam, PPGMRAMRANGE pPrev)
1101	{
1102	Assert(pPrev ? pPrev->pNextR3 == pRam : pVM->pgm.s.pRamRangesXR3 == pRam);
1103	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfR0 == MMHyperCCToR0(pVM, pRam));
1104
1105	pgmLock(pVM);
1106
1107	PPGMRAMRANGE pNext = pRam->pNextR3;
1108	if (pPrev)
1109	{
1110	pPrev->pNextR3 = pNext;
1111	pPrev->pNextR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
1112	}
1113	else
1114	{
1115	Assert(pVM->pgm.s.pRamRangesXR3 == pRam);
1116	pVM->pgm.s.pRamRangesXR3 = pNext;
1117	pVM->pgm.s.pRamRangesXR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
1118	}
1119	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1120
1121	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1122	pgmUnlock(pVM);
1123	}
1124
1125
1126	/**
1127	* Unlink an existing RAM range from the list.
1128	*
1129	* @param pVM The cross context VM structure.
1130	* @param pRam Pointer to the new list entry.
1131	*/
1132	static void pgmR3PhysUnlinkRamRange(PVM pVM, PPGMRAMRANGE pRam)
1133	{
1134	pgmLock(pVM);
1135
1136	/* find prev. */
1137	PPGMRAMRANGE pPrev = NULL;
1138	PPGMRAMRANGE pCur = pVM->pgm.s.pRamRangesXR3;
1139	while (pCur != pRam)
1140	{
1141	pPrev = pCur;
1142	pCur = pCur->pNextR3;
1143	}
1144	AssertFatal(pCur);
1145
1146	pgmR3PhysUnlinkRamRange2(pVM, pRam, pPrev);
1147	pgmUnlock(pVM);
1148	}
1149
1150
1151	/**
1152	* Frees a range of pages, replacing them with ZERO pages of the specified type.
1153	*
1154	* @returns VBox status code.
1155	* @param pVM The cross context VM structure.
1156	* @param pRam The RAM range in which the pages resides.
1157	* @param GCPhys The address of the first page.
1158	* @param GCPhysLast The address of the last page.
1159	* @param enmType The page type to replace then with.
1160	*/
1161	static int pgmR3PhysFreePageRange(PVM pVM, PPGMRAMRANGE pRam, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast, PGMPAGETYPE enmType)
1162	{
1163	PGM_LOCK_ASSERT_OWNER(pVM);
1164	uint32_t cPendingPages = 0;
1165	PGMMFREEPAGESREQ pReq;
1166	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1167	AssertLogRelRCReturn(rc, rc);
1168
1169	/* Iterate the pages. */
1170	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
1171	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
1172	while (cPagesLeft-- > 0)
1173	{
1174	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys, enmType);
1175	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
1176
1177	PGM_PAGE_SET_TYPE(pVM, pPageDst, enmType);
1178
1179	GCPhys += PAGE_SIZE;
1180	pPageDst++;
1181	}
1182
1183	if (cPendingPages)
1184	{
1185	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1186	AssertLogRelRCReturn(rc, rc);
1187	}
1188	GMMR3FreePagesCleanup(pReq);
1189
1190	return rc;
1191	}
1192
1193	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
1194
1195	/**
1196	* Rendezvous callback used by PGMR3ChangeMemBalloon that changes the memory balloon size
1197	*
1198	* This is only called on one of the EMTs while the other ones are waiting for
1199	* it to complete this function.
1200	*
1201	* @returns VINF_SUCCESS (VBox strict status code).
1202	* @param pVM The cross context VM structure.
1203	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
1204	* @param pvUser User parameter
1205	*/
1206	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysChangeMemBalloonRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
1207	{
1208	uintptr_t paUser = (uintptr_t )pvUser;
1209	bool fInflate = !!paUser[0];
1210	unsigned cPages = paUser[1];
1211	RTGCPHYS paPhysPage = (RTGCPHYS )paUser[2];
1212	uint32_t cPendingPages = 0;
1213	PGMMFREEPAGESREQ pReq;
1214	int rc;
1215
1216	Log(("pgmR3PhysChangeMemBalloonRendezvous: %s %x pages\n", (fInflate) ? "inflate" : "deflate", cPages));
1217	pgmLock(pVM);
1218
1219	if (fInflate)
1220	{
1221	/* Flush the PGM pool cache as we might have stale references to pages that we just freed. */
1222	pgmR3PoolClearAllRendezvous(pVM, pVCpu, NULL);
1223
1224	/* Replace pages with ZERO pages. */
1225	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1226	if (RT_FAILURE(rc))
1227	{
1228	pgmUnlock(pVM);
1229	AssertLogRelRC(rc);
1230	return rc;
1231	}
1232
1233	/* Iterate the pages. */
1234	for (unsigned i = 0; i < cPages; i++)
1235	{
1236	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
1237	if ( pPage == NULL
1238	\|\| PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM)
1239	{
1240	Log(("pgmR3PhysChangeMemBalloonRendezvous: invalid physical page %RGp pPage->u3Type=%d\n", paPhysPage[i], pPage ? PGM_PAGE_GET_TYPE(pPage) : 0));
1241	break;
1242	}
1243
1244	LogFlow(("balloon page: %RGp\n", paPhysPage[i]));
1245
1246	/* Flush the shadow PT if this page was previously used as a guest page table. */
1247	pgmPoolFlushPageByGCPhys(pVM, paPhysPage[i]);
1248
1249	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, paPhysPage[i], (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage));
1250	if (RT_FAILURE(rc))
1251	{
1252	pgmUnlock(pVM);
1253	AssertLogRelRC(rc);
1254	return rc;
1255	}
1256	Assert(PGM_PAGE_IS_ZERO(pPage));
1257	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_BALLOONED);
1258	}
1259
1260	if (cPendingPages)
1261	{
1262	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1263	if (RT_FAILURE(rc))
1264	{
1265	pgmUnlock(pVM);
1266	AssertLogRelRC(rc);
1267	return rc;
1268	}
1269	}
1270	GMMR3FreePagesCleanup(pReq);
1271	}
1272	else
1273	{
1274	/* Iterate the pages. */
1275	for (unsigned i = 0; i < cPages; i++)
1276	{
1277	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
1278	AssertBreak(pPage && PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM);
1279
1280	LogFlow(("Free ballooned page: %RGp\n", paPhysPage[i]));
1281
1282	Assert(PGM_PAGE_IS_BALLOONED(pPage));
1283
1284	/* Change back to zero page. (NEM does not need to be informed.) */
1285	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
1286	}
1287
1288	/* Note that we currently do not map any ballooned pages in our shadow page tables, so no need to flush the pgm pool. */
1289	}
1290
1291	/* Notify GMM about the balloon change. */
1292	rc = GMMR3BalloonedPages(pVM, (fInflate) ? GMMBALLOONACTION_INFLATE : GMMBALLOONACTION_DEFLATE, cPages);
1293	if (RT_SUCCESS(rc))
1294	{
1295	if (!fInflate)
1296	{
1297	Assert(pVM->pgm.s.cBalloonedPages >= cPages);
1298	pVM->pgm.s.cBalloonedPages -= cPages;
1299	}
1300	else
1301	pVM->pgm.s.cBalloonedPages += cPages;
1302	}
1303
1304	pgmUnlock(pVM);
1305
1306	/* Flush the recompiler's TLB as well. */
1307	for (VMCPUID i = 0; i < pVM->cCpus; i++)
1308	CPUMSetChangedFlags(pVM->apCpusR3[i], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1309
1310	AssertLogRelRC(rc);
1311	return rc;
1312	}
1313
1314
1315	/**
1316	* Frees a range of ram pages, replacing them with ZERO pages; helper for PGMR3PhysFreeRamPages
1317	*
1318	* @returns VBox status code.
1319	* @param pVM The cross context VM structure.
1320	* @param fInflate Inflate or deflate memory balloon
1321	* @param cPages Number of pages to free
1322	* @param paPhysPage Array of guest physical addresses
1323	*/
1324	static DECLCALLBACK(void) pgmR3PhysChangeMemBalloonHelper(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1325	{
1326	uintptr_t paUser[3];
1327
1328	paUser[0] = fInflate;
1329	paUser[1] = cPages;
1330	paUser[2] = (uintptr_t)paPhysPage;
1331	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1332	AssertRC(rc);
1333
1334	/* Made a copy in PGMR3PhysFreeRamPages; free it here. */
1335	RTMemFree(paPhysPage);
1336	}
1337
1338	#endif /* 64-bit host && (Windows \|\| Solaris \|\| Linux \|\| FreeBSD) */
1339
1340	/**
1341	* Inflate or deflate a memory balloon
1342	*
1343	* @returns VBox status code.
1344	* @param pVM The cross context VM structure.
1345	* @param fInflate Inflate or deflate memory balloon
1346	* @param cPages Number of pages to free
1347	* @param paPhysPage Array of guest physical addresses
1348	*/
1349	VMMR3DECL(int) PGMR3PhysChangeMemBalloon(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1350	{
1351	/* This must match GMMR0Init; currently we only support memory ballooning on all 64-bit hosts except Mac OS X */
1352	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
1353	int rc;
1354
1355	/* Older additions (ancient non-functioning balloon code) pass wrong physical addresses. */
1356	AssertReturn(!(paPhysPage[0] & 0xfff), VERR_INVALID_PARAMETER);
1357
1358	/* We own the IOM lock here and could cause a deadlock by waiting for another VCPU that is blocking on the IOM lock.
1359	* In the SMP case we post a request packet to postpone the job.
1360	*/
1361	if (pVM->cCpus > 1)
1362	{
1363	unsigned cbPhysPage = cPages * sizeof(paPhysPage[0]);
1364	RTGCPHYS paPhysPageCopy = (RTGCPHYS )RTMemAlloc(cbPhysPage);
1365	AssertReturn(paPhysPageCopy, VERR_NO_MEMORY);
1366
1367	memcpy(paPhysPageCopy, paPhysPage, cbPhysPage);
1368
1369	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysChangeMemBalloonHelper, 4, pVM, fInflate, cPages, paPhysPageCopy);
1370	AssertRC(rc);
1371	}
1372	else
1373	{
1374	uintptr_t paUser[3];
1375
1376	paUser[0] = fInflate;
1377	paUser[1] = cPages;
1378	paUser[2] = (uintptr_t)paPhysPage;
1379	rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1380	AssertRC(rc);
1381	}
1382	return rc;
1383
1384	#else
1385	NOREF(pVM); NOREF(fInflate); NOREF(cPages); NOREF(paPhysPage);
1386	return VERR_NOT_IMPLEMENTED;
1387	#endif
1388	}
1389
1390
1391	/**
1392	* Rendezvous callback used by PGMR3WriteProtectRAM that write protects all
1393	* physical RAM.
1394	*
1395	* This is only called on one of the EMTs while the other ones are waiting for
1396	* it to complete this function.
1397	*
1398	* @returns VINF_SUCCESS (VBox strict status code).
1399	* @param pVM The cross context VM structure.
1400	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
1401	* @param pvUser User parameter, unused.
1402	*/
1403	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysWriteProtectRAMRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
1404	{
1405	int rc = VINF_SUCCESS;
1406	NOREF(pvUser); NOREF(pVCpu);
1407
1408	pgmLock(pVM);
1409	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
1410	pgmPoolResetDirtyPages(pVM);
1411	#endif
1412
1413	/** @todo pointless to write protect the physical page pointed to by RSP. */
1414
1415	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
1416	pRam;
1417	pRam = pRam->CTX_SUFF(pNext))
1418	{
1419	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
1420	for (uint32_t iPage = 0; iPage < cPages; iPage++)
1421	{
1422	PPGMPAGE pPage = &pRam->aPages[iPage];
1423	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
1424
1425	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
1426	\|\| enmPageType == PGMPAGETYPE_MMIO2)
1427	{
1428	/*
1429	* A RAM page.
1430	*/
1431	switch (PGM_PAGE_GET_STATE(pPage))
1432	{
1433	case PGM_PAGE_STATE_ALLOCATED:
1434	/** @todo Optimize this: Don't always re-enable write
1435	* monitoring if the page is known to be very busy. */
1436	if (PGM_PAGE_IS_WRITTEN_TO(pPage))
1437	PGM_PAGE_CLEAR_WRITTEN_TO(pVM, pPage);
1438
1439	pgmPhysPageWriteMonitor(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
1440	break;
1441
1442	case PGM_PAGE_STATE_SHARED:
1443	AssertFailed();
1444	break;
1445
1446	case PGM_PAGE_STATE_WRITE_MONITORED: /* nothing to change. */
1447	default:
1448	break;
1449	}
1450	}
1451	}
1452	}
1453	pgmR3PoolWriteProtectPages(pVM);
1454	PGM_INVL_ALL_VCPU_TLBS(pVM);
1455	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
1456	CPUMSetChangedFlags(pVM->apCpusR3[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1457
1458	pgmUnlock(pVM);
1459	return rc;
1460	}
1461
1462	/**
1463	* Protect all physical RAM to monitor writes
1464	*
1465	* @returns VBox status code.
1466	* @param pVM The cross context VM structure.
1467	*/
1468	VMMR3DECL(int) PGMR3PhysWriteProtectRAM(PVM pVM)
1469	{
1470	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1471
1472	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysWriteProtectRAMRendezvous, NULL);
1473	AssertRC(rc);
1474	return rc;
1475	}
1476
1477
1478	/**
1479	* Gets the number of ram ranges.
1480	*
1481	* @returns Number of ram ranges. Returns UINT32_MAX if @a pVM is invalid.
1482	* @param pVM The cross context VM structure.
1483	*/
1484	VMMR3DECL(uint32_t) PGMR3PhysGetRamRangeCount(PVM pVM)
1485	{
1486	VM_ASSERT_VALID_EXT_RETURN(pVM, UINT32_MAX);
1487
1488	pgmLock(pVM);
1489	uint32_t cRamRanges = 0;
1490	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext))
1491	cRamRanges++;
1492	pgmUnlock(pVM);
1493	return cRamRanges;
1494	}
1495
1496
1497	/**
1498	* Get information about a range.
1499	*
1500	* @returns VINF_SUCCESS or VERR_OUT_OF_RANGE.
1501	* @param pVM The cross context VM structure.
1502	* @param iRange The ordinal of the range.
1503	* @param pGCPhysStart Where to return the start of the range. Optional.
1504	* @param pGCPhysLast Where to return the address of the last byte in the
1505	* range. Optional.
1506	* @param ppszDesc Where to return the range description. Optional.
1507	* @param pfIsMmio Where to indicate that this is a pure MMIO range.
1508	* Optional.
1509	*/
1510	VMMR3DECL(int) PGMR3PhysGetRange(PVM pVM, uint32_t iRange, PRTGCPHYS pGCPhysStart, PRTGCPHYS pGCPhysLast,
1511	const char *ppszDesc, bool pfIsMmio)
1512	{
1513	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1514
1515	pgmLock(pVM);
1516	uint32_t iCurRange = 0;
1517	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext), iCurRange++)
1518	if (iCurRange == iRange)
1519	{
1520	if (pGCPhysStart)
1521	*pGCPhysStart = pCur->GCPhys;
1522	if (pGCPhysLast)
1523	*pGCPhysLast = pCur->GCPhysLast;
1524	if (ppszDesc)
1525	*ppszDesc = pCur->pszDesc;
1526	if (pfIsMmio)
1527	*pfIsMmio = !!(pCur->fFlags & PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO);
1528
1529	pgmUnlock(pVM);
1530	return VINF_SUCCESS;
1531	}
1532	pgmUnlock(pVM);
1533	return VERR_OUT_OF_RANGE;
1534	}
1535
1536
1537	/**
1538	* Query the amount of free memory inside VMMR0
1539	*
1540	* @returns VBox status code.
1541	* @param pUVM The user mode VM handle.
1542	* @param pcbAllocMem Where to return the amount of memory allocated
1543	* by VMs.
1544	* @param pcbFreeMem Where to return the amount of memory that is
1545	* allocated from the host but not currently used
1546	* by any VMs.
1547	* @param pcbBallonedMem Where to return the sum of memory that is
1548	* currently ballooned by the VMs.
1549	* @param pcbSharedMem Where to return the amount of memory that is
1550	* currently shared.
1551	*/
1552	VMMR3DECL(int) PGMR3QueryGlobalMemoryStats(PUVM pUVM, uint64_t pcbAllocMem, uint64_t pcbFreeMem,
1553	uint64_t pcbBallonedMem, uint64_t pcbSharedMem)
1554	{
1555	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1556	VM_ASSERT_VALID_EXT_RETURN(pUVM->pVM, VERR_INVALID_VM_HANDLE);
1557
1558	uint64_t cAllocPages = 0;
1559	uint64_t cFreePages = 0;
1560	uint64_t cBalloonPages = 0;
1561	uint64_t cSharedPages = 0;
1562	int rc = GMMR3QueryHypervisorMemoryStats(pUVM->pVM, &cAllocPages, &cFreePages, &cBalloonPages, &cSharedPages);
1563	AssertRCReturn(rc, rc);
1564
1565	if (pcbAllocMem)
1566	pcbAllocMem = cAllocPages _4K;
1567
1568	if (pcbFreeMem)
1569	pcbFreeMem = cFreePages _4K;
1570
1571	if (pcbBallonedMem)
1572	pcbBallonedMem = cBalloonPages _4K;
1573
1574	if (pcbSharedMem)
1575	pcbSharedMem = cSharedPages _4K;
1576
1577	Log(("PGMR3QueryVMMMemoryStats: all=%llx free=%llx ballooned=%llx shared=%llx\n",
1578	cAllocPages, cFreePages, cBalloonPages, cSharedPages));
1579	return VINF_SUCCESS;
1580	}
1581
1582
1583	/**
1584	* Query memory stats for the VM.
1585	*
1586	* @returns VBox status code.
1587	* @param pUVM The user mode VM handle.
1588	* @param pcbTotalMem Where to return total amount memory the VM may
1589	* possibly use.
1590	* @param pcbPrivateMem Where to return the amount of private memory
1591	* currently allocated.
1592	* @param pcbSharedMem Where to return the amount of actually shared
1593	* memory currently used by the VM.
1594	* @param pcbZeroMem Where to return the amount of memory backed by
1595	* zero pages.
1596	*
1597	* @remarks The total mem is normally larger than the sum of the three
1598	* components. There are two reasons for this, first the amount of
1599	* shared memory is what we're sure is shared instead of what could
1600	* possibly be shared with someone. Secondly, because the total may
1601	* include some pure MMIO pages that doesn't go into any of the three
1602	* sub-counts.
1603	*
1604	* @todo Why do we return reused shared pages instead of anything that could
1605	* potentially be shared? Doesn't this mean the first VM gets a much
1606	* lower number of shared pages?
1607	*/
1608	VMMR3DECL(int) PGMR3QueryMemoryStats(PUVM pUVM, uint64_t pcbTotalMem, uint64_t pcbPrivateMem,
1609	uint64_t pcbSharedMem, uint64_t pcbZeroMem)
1610	{
1611	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1612	PVM pVM = pUVM->pVM;
1613	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1614
1615	if (pcbTotalMem)
1616	pcbTotalMem = (uint64_t)pVM->pgm.s.cAllPages PAGE_SIZE;
1617
1618	if (pcbPrivateMem)
1619	pcbPrivateMem = (uint64_t)pVM->pgm.s.cPrivatePages PAGE_SIZE;
1620
1621	if (pcbSharedMem)
1622	pcbSharedMem = (uint64_t)pVM->pgm.s.cReusedSharedPages PAGE_SIZE;
1623
1624	if (pcbZeroMem)
1625	pcbZeroMem = (uint64_t)pVM->pgm.s.cZeroPages PAGE_SIZE;
1626
1627	Log(("PGMR3QueryMemoryStats: all=%x private=%x reused=%x zero=%x\n", pVM->pgm.s.cAllPages, pVM->pgm.s.cPrivatePages, pVM->pgm.s.cReusedSharedPages, pVM->pgm.s.cZeroPages));
1628	return VINF_SUCCESS;
1629	}
1630
1631
1632	/**
1633	* PGMR3PhysRegisterRam worker that initializes and links a RAM range.
1634	*
1635	* @param pVM The cross context VM structure.
1636	* @param pNew The new RAM range.
1637	* @param GCPhys The address of the RAM range.
1638	* @param GCPhysLast The last address of the RAM range.
1639	* @param RCPtrNew The RC address if the range is floating. NIL_RTRCPTR
1640	* if in HMA.
1641	* @param R0PtrNew Ditto for R0.
1642	* @param pszDesc The description.
1643	* @param pPrev The previous RAM range (for linking).
1644	*/
1645	static void pgmR3PhysInitAndLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast,
1646	RTRCPTR RCPtrNew, RTR0PTR R0PtrNew, const char *pszDesc, PPGMRAMRANGE pPrev)
1647	{
1648	/*
1649	* Initialize the range.
1650	*/
1651	pNew->pSelfR0 = R0PtrNew != NIL_RTR0PTR ? R0PtrNew : MMHyperCCToR0(pVM, pNew);
1652	pNew->GCPhys = GCPhys;
1653	pNew->GCPhysLast = GCPhysLast;
1654	pNew->cb = GCPhysLast - GCPhys + 1;
1655	pNew->pszDesc = pszDesc;
1656	pNew->fFlags = RCPtrNew != NIL_RTRCPTR ? PGM_RAM_RANGE_FLAGS_FLOATING : 0;
1657	pNew->pvR3 = NULL;
1658	pNew->paLSPages = NULL;
1659
1660	uint32_t const cPages = pNew->cb >> PAGE_SHIFT;
1661	RTGCPHYS iPage = cPages;
1662	while (iPage-- > 0)
1663	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_RAM);
1664
1665	/* Update the page count stats. */
1666	pVM->pgm.s.cZeroPages += cPages;
1667	pVM->pgm.s.cAllPages += cPages;
1668
1669	/*
1670	* Link it.
1671	*/
1672	pgmR3PhysLinkRamRange(pVM, pNew, pPrev);
1673	}
1674
1675
1676	#ifndef PGM_WITHOUT_MAPPINGS
1677	/**
1678	* @callback_method_impl{FNPGMRELOCATE, Relocate a floating RAM range.}
1679	* @sa pgmR3PhysMMIO2ExRangeRelocate
1680	*/
1681	static DECLCALLBACK(bool) pgmR3PhysRamRangeRelocate(PVM pVM, RTGCPTR GCPtrOld, RTGCPTR GCPtrNew,
1682	PGMRELOCATECALL enmMode, void *pvUser)
1683	{
1684	PPGMRAMRANGE pRam = (PPGMRAMRANGE)pvUser;
1685	Assert(pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING);
1686	Assert(pRam->pSelfRC == GCPtrOld + PAGE_SIZE); RT_NOREF_PV(GCPtrOld);
1687
1688	switch (enmMode)
1689	{
1690	case PGMRELOCATECALL_SUGGEST:
1691	return true;
1692
1693	case PGMRELOCATECALL_RELOCATE:
1694	{
1695	/*
1696	* Update myself, then relink all the ranges and flush the RC TLB.
1697	*/
1698	pgmLock(pVM);
1699
1700	pRam->pSelfRC = (RTRCPTR)(GCPtrNew + PAGE_SIZE);
1701
1702	pgmR3PhysRelinkRamRanges(pVM);
1703	for (unsigned i = 0; i < PGM_RAMRANGE_TLB_ENTRIES; i++)
1704	pVM->pgm.s.apRamRangesTlbRC[i] = NIL_RTRCPTR;
1705
1706	pgmUnlock(pVM);
1707	return true;
1708	}
1709
1710	default:
1711	AssertFailedReturn(false);
1712	}
1713	}
1714	#endif /* !PGM_WITHOUT_MAPPINGS */
1715
1716
1717	/**
1718	* PGMR3PhysRegisterRam worker that registers a high chunk.
1719	*
1720	* @returns VBox status code.
1721	* @param pVM The cross context VM structure.
1722	* @param GCPhys The address of the RAM.
1723	* @param cRamPages The number of RAM pages to register.
1724	* @param cbChunk The size of the PGMRAMRANGE guest mapping.
1725	* @param iChunk The chunk number.
1726	* @param pszDesc The RAM range description.
1727	* @param ppPrev Previous RAM range pointer. In/Out.
1728	*/
1729	static int pgmR3PhysRegisterHighRamChunk(PVM pVM, RTGCPHYS GCPhys, uint32_t cRamPages,
1730	uint32_t cbChunk, uint32_t iChunk, const char *pszDesc,
1731	PPGMRAMRANGE *ppPrev)
1732	{
1733	const char *pszDescChunk = iChunk == 0
1734	? pszDesc
1735	: MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s (#%u)", pszDesc, iChunk + 1);
1736	AssertReturn(pszDescChunk, VERR_NO_MEMORY);
1737
1738	/*
1739	* Allocate memory for the new chunk.
1740	*/
1741	size_t const cChunkPages = RT_ALIGN_Z(RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cRamPages]), PAGE_SIZE) >> PAGE_SHIFT;
1742	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
1743	AssertReturn(paChunkPages, VERR_NO_TMP_MEMORY);
1744	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
1745	void *pvChunk = NULL;
1746	int rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk, &R0PtrChunk, paChunkPages);
1747	if (RT_SUCCESS(rc))
1748	{
1749	Assert(R0PtrChunk != NIL_RTR0PTR);
1750	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
1751
1752	PPGMRAMRANGE pNew = (PPGMRAMRANGE)pvChunk;
1753
1754	/*
1755	* Create a mapping and map the pages into it.
1756	* We push these in below the HMA.
1757	*/
1758	RTGCPTR GCPtrChunkMap = pVM->pgm.s.GCPtrPrevRamRangeMapping - cbChunk;
1759	#ifndef PGM_WITHOUT_MAPPINGS
1760	rc = PGMR3MapPT(pVM, GCPtrChunkMap, cbChunk, 0 /fFlags/, pgmR3PhysRamRangeRelocate, pNew, pszDescChunk);
1761	if (RT_SUCCESS(rc))
1762	#endif /* !PGM_WITHOUT_MAPPINGS */
1763	{
1764	pVM->pgm.s.GCPtrPrevRamRangeMapping = GCPtrChunkMap;
1765
1766	RTGCPTR const GCPtrChunk = GCPtrChunkMap + PAGE_SIZE;
1767	#ifndef PGM_WITHOUT_MAPPINGS
1768	RTGCPTR GCPtrPage = GCPtrChunk;
1769	for (uint32_t iPage = 0; iPage < cChunkPages && RT_SUCCESS(rc); iPage++, GCPtrPage += PAGE_SIZE)
1770	rc = PGMMap(pVM, GCPtrPage, paChunkPages[iPage].Phys, PAGE_SIZE, 0);
1771	if (RT_SUCCESS(rc))
1772	#endif /* !PGM_WITHOUT_MAPPINGS */
1773	{
1774	/*
1775	* Ok, init and link the range.
1776	*/
1777	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhys + ((RTGCPHYS)cRamPages << PAGE_SHIFT) - 1,
1778	(RTRCPTR)GCPtrChunk, R0PtrChunk, pszDescChunk, *ppPrev);
1779	*ppPrev = pNew;
1780	}
1781	}
1782
1783	if (RT_FAILURE(rc))
1784	SUPR3PageFreeEx(pvChunk, cChunkPages);
1785	}
1786
1787	RTMemTmpFree(paChunkPages);
1788	return rc;
1789	}
1790
1791
1792	/**
1793	* Sets up a range RAM.
1794	*
1795	* This will check for conflicting registrations, make a resource
1796	* reservation for the memory (with GMM), and setup the per-page
1797	* tracking structures (PGMPAGE).
1798	*
1799	* @returns VBox status code.
1800	* @param pVM The cross context VM structure.
1801	* @param GCPhys The physical address of the RAM.
1802	* @param cb The size of the RAM.
1803	* @param pszDesc The description - not copied, so, don't free or change it.
1804	*/
1805	VMMR3DECL(int) PGMR3PhysRegisterRam(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, const char *pszDesc)
1806	{
1807	/*
1808	* Validate input.
1809	*/
1810	Log(("PGMR3PhysRegisterRam: GCPhys=%RGp cb=%RGp pszDesc=%s\n", GCPhys, cb, pszDesc));
1811	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
1812	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
1813	AssertReturn(cb > 0, VERR_INVALID_PARAMETER);
1814	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
1815	AssertMsgReturn(GCPhysLast > GCPhys, ("The range wraps! GCPhys=%RGp cb=%RGp\n", GCPhys, cb), VERR_INVALID_PARAMETER);
1816	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
1817	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1818
1819	pgmLock(pVM);
1820
1821	/*
1822	* Find range location and check for conflicts.
1823	* (We don't lock here because the locking by EMT is only required on update.)
1824	*/
1825	PPGMRAMRANGE pPrev = NULL;
1826	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
1827	while (pRam && GCPhysLast >= pRam->GCPhys)
1828	{
1829	if ( GCPhysLast >= pRam->GCPhys
1830	&& GCPhys <= pRam->GCPhysLast)
1831	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
1832	GCPhys, GCPhysLast, pszDesc,
1833	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
1834	VERR_PGM_RAM_CONFLICT);
1835
1836	/* next */
1837	pPrev = pRam;
1838	pRam = pRam->pNextR3;
1839	}
1840
1841	/*
1842	* Register it with GMM (the API bitches).
1843	*/
1844	const RTGCPHYS cPages = cb >> PAGE_SHIFT;
1845	int rc = MMR3IncreaseBaseReservation(pVM, cPages);
1846	if (RT_FAILURE(rc))
1847	{
1848	pgmUnlock(pVM);
1849	return rc;
1850	}
1851
1852	if ( GCPhys >= _4G
1853	&& cPages > 256)
1854	{
1855	/*
1856	* The PGMRAMRANGE structures for the high memory can get very big.
1857	* In order to avoid SUPR3PageAllocEx allocation failures due to the
1858	* allocation size limit there and also to avoid being unable to find
1859	* guest mapping space for them, we split this memory up into 4MB in
1860	* (potential) raw-mode configs and 16MB chunks in forced AMD-V/VT-x
1861	* mode.
1862	*
1863	* The first and last page of each mapping are guard pages and marked
1864	* not-present. So, we've got 4186112 and 16769024 bytes available for
1865	* the PGMRAMRANGE structure.
1866	*
1867	* Note! The sizes used here will influence the saved state.
1868	*/
1869	uint32_t cbChunk = 16U*_1M;
1870	uint32_t cPagesPerChunk = 1048048; /* max ~1048059 */
1871	AssertCompile(sizeof(PGMRAMRANGE) + sizeof(PGMPAGE) * 1048048 < 16U_1M - PAGE_SIZE 2);
1872	AssertRelease(RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPagesPerChunk]) + PAGE_SIZE * 2 <= cbChunk);
1873
1874	RTGCPHYS cPagesLeft = cPages;
1875	RTGCPHYS GCPhysChunk = GCPhys;
1876	uint32_t iChunk = 0;
1877	while (cPagesLeft > 0)
1878	{
1879	uint32_t cPagesInChunk = cPagesLeft;
1880	if (cPagesInChunk > cPagesPerChunk)
1881	cPagesInChunk = cPagesPerChunk;
1882
1883	rc = pgmR3PhysRegisterHighRamChunk(pVM, GCPhysChunk, cPagesInChunk, cbChunk, iChunk, pszDesc, &pPrev);
1884	AssertRCReturn(rc, rc);
1885
1886	/* advance */
1887	GCPhysChunk += (RTGCPHYS)cPagesInChunk << PAGE_SHIFT;
1888	cPagesLeft -= cPagesInChunk;
1889	iChunk++;
1890	}
1891	}
1892	else
1893	{
1894	/*
1895	* Allocate, initialize and link the new RAM range.
1896	*/
1897	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
1898	PPGMRAMRANGE pNew;
1899	rc = MMR3HyperAllocOnceNoRel(pVM, cbRamRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
1900	AssertLogRelMsgRCReturn(rc, ("cbRamRange=%zu\n", cbRamRange), rc);
1901
1902	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhysLast, NIL_RTRCPTR, NIL_RTR0PTR, pszDesc, pPrev);
1903	}
1904	pgmPhysInvalidatePageMapTLB(pVM);
1905
1906	/*
1907	* Notify NEM while holding the lock (experimental) and REM without (like always).
1908	*/
1909	rc = NEMR3NotifyPhysRamRegister(pVM, GCPhys, cb);
1910	pgmUnlock(pVM);
1911	return rc;
1912	}
1913
1914
1915	/**
1916	* Worker called by PGMR3InitFinalize if we're configured to pre-allocate RAM.
1917	*
1918	* We do this late in the init process so that all the ROM and MMIO ranges have
1919	* been registered already and we don't go wasting memory on them.
1920	*
1921	* @returns VBox status code.
1922	*
1923	* @param pVM The cross context VM structure.
1924	*/
1925	int pgmR3PhysRamPreAllocate(PVM pVM)
1926	{
1927	Assert(pVM->pgm.s.fRamPreAlloc);
1928	Log(("pgmR3PhysRamPreAllocate: enter\n"));
1929
1930	/*
1931	* Walk the RAM ranges and allocate all RAM pages, halt at
1932	* the first allocation error.
1933	*/
1934	uint64_t cPages = 0;
1935	uint64_t NanoTS = RTTimeNanoTS();
1936	pgmLock(pVM);
1937	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1938	{
1939	PPGMPAGE pPage = &pRam->aPages[0];
1940	RTGCPHYS GCPhys = pRam->GCPhys;
1941	uint32_t cLeft = pRam->cb >> PAGE_SHIFT;
1942	while (cLeft-- > 0)
1943	{
1944	if (PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1945	{
1946	switch (PGM_PAGE_GET_STATE(pPage))
1947	{
1948	case PGM_PAGE_STATE_ZERO:
1949	{
1950	int rc = pgmPhysAllocPage(pVM, pPage, GCPhys);
1951	if (RT_FAILURE(rc))
1952	{
1953	LogRel(("PGM: RAM Pre-allocation failed at %RGp (in %s) with rc=%Rrc\n", GCPhys, pRam->pszDesc, rc));
1954	pgmUnlock(pVM);
1955	return rc;
1956	}
1957	cPages++;
1958	break;
1959	}
1960
1961	case PGM_PAGE_STATE_BALLOONED:
1962	case PGM_PAGE_STATE_ALLOCATED:
1963	case PGM_PAGE_STATE_WRITE_MONITORED:
1964	case PGM_PAGE_STATE_SHARED:
1965	/* nothing to do here. */
1966	break;
1967	}
1968	}
1969
1970	/* next */
1971	pPage++;
1972	GCPhys += PAGE_SIZE;
1973	}
1974	}
1975	pgmUnlock(pVM);
1976	NanoTS = RTTimeNanoTS() - NanoTS;
1977
1978	LogRel(("PGM: Pre-allocated %llu pages in %llu ms\n", cPages, NanoTS / 1000000));
1979	Log(("pgmR3PhysRamPreAllocate: returns VINF_SUCCESS\n"));
1980	return VINF_SUCCESS;
1981	}
1982
1983
1984	/**
1985	* Checks shared page checksums.
1986	*
1987	* @param pVM The cross context VM structure.
1988	*/
1989	void pgmR3PhysAssertSharedPageChecksums(PVM pVM)
1990	{
1991	#ifdef VBOX_STRICT
1992	pgmLock(pVM);
1993
1994	if (pVM->pgm.s.cSharedPages > 0)
1995	{
1996	/*
1997	* Walk the ram ranges.
1998	*/
1999	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2000	{
2001	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2002	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2003
2004	while (iPage-- > 0)
2005	{
2006	PPGMPAGE pPage = &pRam->aPages[iPage];
2007	if (PGM_PAGE_IS_SHARED(pPage))
2008	{
2009	uint32_t u32Checksum = pPage->s.u2Unused0/* \| ((uint32_t)pPage->s.u2Unused1 << 8)*/;
2010	if (!u32Checksum)
2011	{
2012	RTGCPHYS GCPhysPage = pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT);
2013	void const *pvPage;
2014	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhysPage, &pvPage);
2015	if (RT_SUCCESS(rc))
2016	{
2017	uint32_t u32Checksum2 = RTCrc32(pvPage, PAGE_SIZE);
2018	# if 0
2019	AssertMsg((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum, ("GCPhysPage=%RGp\n", GCPhysPage));
2020	# else
2021	if ((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum)
2022	LogFlow(("shpg %#x @ %RGp %#x [OK]\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
2023	else
2024	AssertMsgFailed(("shpg %#x @ %RGp %#x\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
2025	# endif
2026	}
2027	else
2028	AssertRC(rc);
2029	}
2030	}
2031
2032	} /* for each page */
2033
2034	} /* for each ram range */
2035	}
2036
2037	pgmUnlock(pVM);
2038	#endif /* VBOX_STRICT */
2039	NOREF(pVM);
2040	}
2041
2042
2043	/**
2044	* Resets the physical memory state.
2045	*
2046	* ASSUMES that the caller owns the PGM lock.
2047	*
2048	* @returns VBox status code.
2049	* @param pVM The cross context VM structure.
2050	*/
2051	int pgmR3PhysRamReset(PVM pVM)
2052	{
2053	PGM_LOCK_ASSERT_OWNER(pVM);
2054
2055	/* Reset the memory balloon. */
2056	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2057	AssertRC(rc);
2058
2059	#ifdef VBOX_WITH_PAGE_SHARING
2060	/* Clear all registered shared modules. */
2061	pgmR3PhysAssertSharedPageChecksums(pVM);
2062	rc = GMMR3ResetSharedModules(pVM);
2063	AssertRC(rc);
2064	#endif
2065	/* Reset counters. */
2066	pVM->pgm.s.cReusedSharedPages = 0;
2067	pVM->pgm.s.cBalloonedPages = 0;
2068
2069	return VINF_SUCCESS;
2070	}
2071
2072
2073	/**
2074	* Resets (zeros) the RAM after all devices and components have been reset.
2075	*
2076	* ASSUMES that the caller owns the PGM lock.
2077	*
2078	* @returns VBox status code.
2079	* @param pVM The cross context VM structure.
2080	*/
2081	int pgmR3PhysRamZeroAll(PVM pVM)
2082	{
2083	PGM_LOCK_ASSERT_OWNER(pVM);
2084
2085	/*
2086	* We batch up pages that should be freed instead of calling GMM for
2087	* each and every one of them.
2088	*/
2089	uint32_t cPendingPages = 0;
2090	PGMMFREEPAGESREQ pReq;
2091	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2092	AssertLogRelRCReturn(rc, rc);
2093
2094	/*
2095	* Walk the ram ranges.
2096	*/
2097	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2098	{
2099	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2100	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2101
2102	if ( !pVM->pgm.s.fRamPreAlloc
2103	&& pVM->pgm.s.fZeroRamPagesOnReset)
2104	{
2105	/* Replace all RAM pages by ZERO pages. */
2106	while (iPage-- > 0)
2107	{
2108	PPGMPAGE pPage = &pRam->aPages[iPage];
2109	switch (PGM_PAGE_GET_TYPE(pPage))
2110	{
2111	case PGMPAGETYPE_RAM:
2112	/* Do not replace pages part of a 2 MB continuous range
2113	with zero pages, but zero them instead. */
2114	if ( PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE
2115	\|\| PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE_DISABLED)
2116	{
2117	void *pvPage;
2118	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2119	AssertLogRelRCReturn(rc, rc);
2120	ASMMemZeroPage(pvPage);
2121	}
2122	else if (PGM_PAGE_IS_BALLOONED(pPage))
2123	{
2124	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2125	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2126	}
2127	else if (!PGM_PAGE_IS_ZERO(pPage))
2128	{
2129	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2130	PGMPAGETYPE_RAM);
2131	AssertLogRelRCReturn(rc, rc);
2132	}
2133	break;
2134
2135	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2136	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2137	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2138	true /fDoAccounting/);
2139	break;
2140
2141	case PGMPAGETYPE_MMIO2:
2142	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2143	case PGMPAGETYPE_ROM:
2144	case PGMPAGETYPE_MMIO:
2145	break;
2146	default:
2147	AssertFailed();
2148	}
2149	} /* for each page */
2150	}
2151	else
2152	{
2153	/* Zero the memory. */
2154	while (iPage-- > 0)
2155	{
2156	PPGMPAGE pPage = &pRam->aPages[iPage];
2157	switch (PGM_PAGE_GET_TYPE(pPage))
2158	{
2159	case PGMPAGETYPE_RAM:
2160	switch (PGM_PAGE_GET_STATE(pPage))
2161	{
2162	case PGM_PAGE_STATE_ZERO:
2163	break;
2164
2165	case PGM_PAGE_STATE_BALLOONED:
2166	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2167	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2168	break;
2169
2170	case PGM_PAGE_STATE_SHARED:
2171	case PGM_PAGE_STATE_WRITE_MONITORED:
2172	rc = pgmPhysPageMakeWritable(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2173	AssertLogRelRCReturn(rc, rc);
2174	RT_FALL_THRU();
2175
2176	case PGM_PAGE_STATE_ALLOCATED:
2177	if (pVM->pgm.s.fZeroRamPagesOnReset)
2178	{
2179	void *pvPage;
2180	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2181	AssertLogRelRCReturn(rc, rc);
2182	ASMMemZeroPage(pvPage);
2183	}
2184	break;
2185	}
2186	break;
2187
2188	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2189	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2190	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2191	true /fDoAccounting/);
2192	break;
2193
2194	case PGMPAGETYPE_MMIO2:
2195	case PGMPAGETYPE_ROM_SHADOW:
2196	case PGMPAGETYPE_ROM:
2197	case PGMPAGETYPE_MMIO:
2198	break;
2199	default:
2200	AssertFailed();
2201
2202	}
2203	} /* for each page */
2204	}
2205
2206	}
2207
2208	/*
2209	* Finish off any pages pending freeing.
2210	*/
2211	if (cPendingPages)
2212	{
2213	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2214	AssertLogRelRCReturn(rc, rc);
2215	}
2216	GMMR3FreePagesCleanup(pReq);
2217	return VINF_SUCCESS;
2218	}
2219
2220
2221	/**
2222	* Frees all RAM during VM termination
2223	*
2224	* ASSUMES that the caller owns the PGM lock.
2225	*
2226	* @returns VBox status code.
2227	* @param pVM The cross context VM structure.
2228	*/
2229	int pgmR3PhysRamTerm(PVM pVM)
2230	{
2231	PGM_LOCK_ASSERT_OWNER(pVM);
2232
2233	/* Reset the memory balloon. */
2234	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2235	AssertRC(rc);
2236
2237	#ifdef VBOX_WITH_PAGE_SHARING
2238	/*
2239	* Clear all registered shared modules.
2240	*/
2241	pgmR3PhysAssertSharedPageChecksums(pVM);
2242	rc = GMMR3ResetSharedModules(pVM);
2243	AssertRC(rc);
2244
2245	/*
2246	* Flush the handy pages updates to make sure no shared pages are hiding
2247	* in there. (No unlikely if the VM shuts down, apparently.)
2248	*/
2249	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_FLUSH_HANDY_PAGES, 0, NULL);
2250	#endif
2251
2252	/*
2253	* We batch up pages that should be freed instead of calling GMM for
2254	* each and every one of them.
2255	*/
2256	uint32_t cPendingPages = 0;
2257	PGMMFREEPAGESREQ pReq;
2258	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2259	AssertLogRelRCReturn(rc, rc);
2260
2261	/*
2262	* Walk the ram ranges.
2263	*/
2264	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2265	{
2266	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2267	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2268
2269	while (iPage-- > 0)
2270	{
2271	PPGMPAGE pPage = &pRam->aPages[iPage];
2272	switch (PGM_PAGE_GET_TYPE(pPage))
2273	{
2274	case PGMPAGETYPE_RAM:
2275	/* Free all shared pages. Private pages are automatically freed during GMM VM cleanup. */
2276	/** @todo change this to explicitly free private pages here. */
2277	if (PGM_PAGE_IS_SHARED(pPage))
2278	{
2279	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2280	PGMPAGETYPE_RAM);
2281	AssertLogRelRCReturn(rc, rc);
2282	}
2283	break;
2284
2285	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2286	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO:
2287	case PGMPAGETYPE_MMIO2:
2288	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2289	case PGMPAGETYPE_ROM:
2290	case PGMPAGETYPE_MMIO:
2291	break;
2292	default:
2293	AssertFailed();
2294	}
2295	} /* for each page */
2296	}
2297
2298	/*
2299	* Finish off any pages pending freeing.
2300	*/
2301	if (cPendingPages)
2302	{
2303	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2304	AssertLogRelRCReturn(rc, rc);
2305	}
2306	GMMR3FreePagesCleanup(pReq);
2307	return VINF_SUCCESS;
2308	}
2309
2310
2311	/**
2312	* This is the interface IOM is using to register an MMIO region.
2313	*
2314	* It will check for conflicts and ensure that a RAM range structure
2315	* is present before calling the PGMR3HandlerPhysicalRegister API to
2316	* register the callbacks.
2317	*
2318	* @returns VBox status code.
2319	*
2320	* @param pVM The cross context VM structure.
2321	* @param GCPhys The start of the MMIO region.
2322	* @param cb The size of the MMIO region.
2323	* @param hType The physical access handler type registration.
2324	* @param pvUserR3 The user argument for R3.
2325	* @param pvUserR0 The user argument for R0.
2326	* @param pvUserRC The user argument for RC.
2327	* @param pszDesc The description of the MMIO region.
2328	*/
2329	VMMR3DECL(int) PGMR3PhysMMIORegister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMPHYSHANDLERTYPE hType,
2330	RTR3PTR pvUserR3, RTR0PTR pvUserR0, RTRCPTR pvUserRC, const char *pszDesc)
2331	{
2332	/*
2333	* Assert on some assumption.
2334	*/
2335	VM_ASSERT_EMT(pVM);
2336	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2337	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2338	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2339	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2340	Assert(((PPGMPHYSHANDLERTYPEINT)MMHyperHeapOffsetToPtr(pVM, hType))->enmKind == PGMPHYSHANDLERKIND_MMIO);
2341
2342	int rc = pgmLock(pVM);
2343	AssertRCReturn(rc, rc);
2344
2345	/*
2346	* Make sure there's a RAM range structure for the region.
2347	*/
2348	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2349	bool fRamExists = false;
2350	PPGMRAMRANGE pRamPrev = NULL;
2351	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2352	while (pRam && GCPhysLast >= pRam->GCPhys)
2353	{
2354	if ( GCPhysLast >= pRam->GCPhys
2355	&& GCPhys <= pRam->GCPhysLast)
2356	{
2357	/* Simplification: all within the same range. */
2358	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
2359	&& GCPhysLast <= pRam->GCPhysLast,
2360	("%RGp-%RGp (MMIO/%s) falls partly outside %RGp-%RGp (%s)\n",
2361	GCPhys, GCPhysLast, pszDesc,
2362	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2363	pgmUnlock(pVM),
2364	VERR_PGM_RAM_CONFLICT);
2365
2366	/* Check that it's all RAM or MMIO pages. */
2367	PCPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2368	uint32_t cLeft = cb >> PAGE_SHIFT;
2369	while (cLeft-- > 0)
2370	{
2371	AssertLogRelMsgReturnStmt( PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM
2372	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO,
2373	("%RGp-%RGp (MMIO/%s): %RGp is not a RAM or MMIO page - type=%d desc=%s\n",
2374	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, PGM_PAGE_GET_TYPE(pPage), pRam->pszDesc),
2375	pgmUnlock(pVM),
2376	VERR_PGM_RAM_CONFLICT);
2377	pPage++;
2378	}
2379
2380	/* Looks good. */
2381	fRamExists = true;
2382	break;
2383	}
2384
2385	/* next */
2386	pRamPrev = pRam;
2387	pRam = pRam->pNextR3;
2388	}
2389	PPGMRAMRANGE pNew;
2390	if (fRamExists)
2391	{
2392	pNew = NULL;
2393
2394	/*
2395	* Make all the pages in the range MMIO/ZERO pages, freeing any
2396	* RAM pages currently mapped here. This might not be 100% correct
2397	* for PCI memory, but we're doing the same thing for MMIO2 pages.
2398	*/
2399	rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, PGMPAGETYPE_MMIO);
2400	AssertRCReturnStmt(rc, pgmUnlock(pVM), rc);
2401
2402	/* Force a PGM pool flush as guest ram references have been changed. */
2403	/** @todo not entirely SMP safe; assuming for now the guest takes
2404	* care of this internally (not touch mapped mmio while changing the
2405	* mapping). */
2406	PVMCPU pVCpu = VMMGetCpu(pVM);
2407	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2408	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2409	}
2410	else
2411	{
2412
2413	/*
2414	* No RAM range, insert an ad hoc one.
2415	*
2416	* Note that we don't have to tell REM about this range because
2417	* PGMHandlerPhysicalRegisterEx will do that for us.
2418	*/
2419	Log(("PGMR3PhysMMIORegister: Adding ad hoc MMIO range for %RGp-%RGp %s\n", GCPhys, GCPhysLast, pszDesc));
2420
2421	const uint32_t cPages = cb >> PAGE_SHIFT;
2422	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
2423	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), 16, MM_TAG_PGM_PHYS, (void **)&pNew);
2424	AssertLogRelMsgRCReturnStmt(rc, ("cbRamRange=%zu\n", cbRamRange), pgmUnlock(pVM), rc);
2425
2426	/* Initialize the range. */
2427	pNew->pSelfR0 = MMHyperCCToR0(pVM, pNew);
2428	pNew->GCPhys = GCPhys;
2429	pNew->GCPhysLast = GCPhysLast;
2430	pNew->cb = cb;
2431	pNew->pszDesc = pszDesc;
2432	pNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO;
2433	pNew->pvR3 = NULL;
2434	pNew->paLSPages = NULL;
2435
2436	uint32_t iPage = cPages;
2437	while (iPage-- > 0)
2438	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_MMIO);
2439	Assert(PGM_PAGE_GET_TYPE(&pNew->aPages[0]) == PGMPAGETYPE_MMIO);
2440
2441	/* update the page count stats. */
2442	pVM->pgm.s.cPureMmioPages += cPages;
2443	pVM->pgm.s.cAllPages += cPages;
2444
2445	/* link it */
2446	pgmR3PhysLinkRamRange(pVM, pNew, pRamPrev);
2447	}
2448
2449	/*
2450	* Register the access handler.
2451	*/
2452	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, hType, pvUserR3, pvUserR0, pvUserRC, pszDesc);
2453	if ( RT_FAILURE(rc)
2454	&& !fRamExists)
2455	{
2456	pVM->pgm.s.cPureMmioPages -= cb >> PAGE_SHIFT;
2457	pVM->pgm.s.cAllPages -= cb >> PAGE_SHIFT;
2458
2459	/* remove the ad hoc range. */
2460	pgmR3PhysUnlinkRamRange2(pVM, pNew, pRamPrev);
2461	pNew->cb = pNew->GCPhys = pNew->GCPhysLast = NIL_RTGCPHYS;
2462	MMHyperFree(pVM, pRam);
2463	}
2464	pgmPhysInvalidatePageMapTLB(pVM);
2465
2466	pgmUnlock(pVM);
2467	return rc;
2468	}
2469
2470
2471	/**
2472	* This is the interface IOM is using to register an MMIO region.
2473	*
2474	* It will take care of calling PGMHandlerPhysicalDeregister and clean up
2475	* any ad hoc PGMRAMRANGE left behind.
2476	*
2477	* @returns VBox status code.
2478	* @param pVM The cross context VM structure.
2479	* @param GCPhys The start of the MMIO region.
2480	* @param cb The size of the MMIO region.
2481	*/
2482	VMMR3DECL(int) PGMR3PhysMMIODeregister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb)
2483	{
2484	VM_ASSERT_EMT(pVM);
2485
2486	int rc = pgmLock(pVM);
2487	AssertRCReturn(rc, rc);
2488
2489	/*
2490	* First deregister the handler, then check if we should remove the ram range.
2491	*/
2492	rc = PGMHandlerPhysicalDeregister(pVM, GCPhys);
2493	if (RT_SUCCESS(rc))
2494	{
2495	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2496	PPGMRAMRANGE pRamPrev = NULL;
2497	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2498	while (pRam && GCPhysLast >= pRam->GCPhys)
2499	{
2500	/** @todo We're being a bit too careful here. rewrite. */
2501	if ( GCPhysLast == pRam->GCPhysLast
2502	&& GCPhys == pRam->GCPhys)
2503	{
2504	Assert(pRam->cb == cb);
2505
2506	/*
2507	* See if all the pages are dead MMIO pages.
2508	*/
2509	uint32_t const cPages = cb >> PAGE_SHIFT;
2510	bool fAllMMIO = true;
2511	uint32_t iPage = 0;
2512	uint32_t cLeft = cPages;
2513	while (cLeft-- > 0)
2514	{
2515	PPGMPAGE pPage = &pRam->aPages[iPage];
2516	if ( !PGM_PAGE_IS_MMIO_OR_ALIAS(pPage)
2517	/\|\| not-out-of-action later /)
2518	{
2519	fAllMMIO = false;
2520	AssertMsgFailed(("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2521	break;
2522	}
2523	Assert( PGM_PAGE_IS_ZERO(pPage)
2524	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2525	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO);
2526	pPage++;
2527	}
2528	if (fAllMMIO)
2529	{
2530	/*
2531	* Ad-hoc range, unlink and free it.
2532	*/
2533	Log(("PGMR3PhysMMIODeregister: Freeing ad hoc MMIO range for %RGp-%RGp %s\n",
2534	GCPhys, GCPhysLast, pRam->pszDesc));
2535
2536	pVM->pgm.s.cAllPages -= cPages;
2537	pVM->pgm.s.cPureMmioPages -= cPages;
2538
2539	pgmR3PhysUnlinkRamRange2(pVM, pRam, pRamPrev);
2540	pRam->cb = pRam->GCPhys = pRam->GCPhysLast = NIL_RTGCPHYS;
2541	MMHyperFree(pVM, pRam);
2542	break;
2543	}
2544	}
2545
2546	/*
2547	* Range match? It will all be within one range (see PGMAllHandler.cpp).
2548	*/
2549	if ( GCPhysLast >= pRam->GCPhys
2550	&& GCPhys <= pRam->GCPhysLast)
2551	{
2552	Assert(GCPhys >= pRam->GCPhys);
2553	Assert(GCPhysLast <= pRam->GCPhysLast);
2554
2555	/*
2556	* Turn the pages back into RAM pages.
2557	*/
2558	uint32_t iPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2559	uint32_t cLeft = cb >> PAGE_SHIFT;
2560	while (cLeft--)
2561	{
2562	PPGMPAGE pPage = &pRam->aPages[iPage];
2563	AssertMsg( (PGM_PAGE_IS_MMIO(pPage) && PGM_PAGE_IS_ZERO(pPage))
2564	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2565	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO,
2566	("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2567	if (PGM_PAGE_IS_MMIO_OR_ALIAS(pPage))
2568	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_RAM);
2569	}
2570	break;
2571	}
2572
2573	/* next */
2574	pRamPrev = pRam;
2575	pRam = pRam->pNextR3;
2576	}
2577	}
2578
2579	/* Force a PGM pool flush as guest ram references have been changed. */
2580	/** @todo Not entirely SMP safe; assuming for now the guest takes care of
2581	* this internally (not touch mapped mmio while changing the mapping). */
2582	PVMCPU pVCpu = VMMGetCpu(pVM);
2583	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2584	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2585
2586	pgmPhysInvalidatePageMapTLB(pVM);
2587	pgmPhysInvalidRamRangeTlbs(pVM);
2588	pgmUnlock(pVM);
2589	return rc;
2590	}
2591
2592
2593	/**
2594	* Locate a MMIO2 range.
2595	*
2596	* @returns Pointer to the MMIO2 range.
2597	* @param pVM The cross context VM structure.
2598	* @param pDevIns The device instance owning the region.
2599	* @param iSubDev The sub-device number.
2600	* @param iRegion The region.
2601	* @param hMmio2 Handle to look up. If NIL, use the @a iSubDev and
2602	* @a iRegion.
2603	*/
2604	DECLINLINE(PPGMREGMMIO2RANGE) pgmR3PhysMmio2Find(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev,
2605	uint32_t iRegion, PGMMMIO2HANDLE hMmio2)
2606	{
2607	if (hMmio2 != NIL_PGMMMIO2HANDLE)
2608	{
2609	if (hMmio2 <= RT_ELEMENTS(pVM->pgm.s.apMmio2RangesR3) && hMmio2 != 0)
2610	{
2611	PPGMREGMMIO2RANGE pCur = pVM->pgm.s.apMmio2RangesR3[hMmio2 - 1];
2612	if (pCur && pCur->pDevInsR3 == pDevIns)
2613	{
2614	Assert(pCur->idMmio2 == hMmio2);
2615	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_MMIO2, NULL);
2616	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, NULL);
2617	return pCur;
2618	}
2619	Assert(!pCur);
2620	}
2621	for (PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2622	if (pCur->idMmio2 == hMmio2)
2623	{
2624	AssertBreak(pCur->pDevInsR3 == pDevIns);
2625	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_MMIO2, NULL);
2626	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, NULL);
2627	return pCur;
2628	}
2629	}
2630	else
2631	{
2632	/*
2633	* Search the list. There shouldn't be many entries.
2634	*/
2635	/** @todo Optimize this lookup! There may now be many entries and it'll
2636	* become really slow when doing MMR3HyperMapMMIO2 and similar. */
2637	for (PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2638	if ( pCur->pDevInsR3 == pDevIns
2639	&& pCur->iRegion == iRegion
2640	&& pCur->iSubDev == iSubDev)
2641	return pCur;
2642	}
2643	return NULL;
2644	}
2645
2646
2647	#ifndef PGM_WITHOUT_MAPPINGS
2648	/**
2649	* @callback_method_impl{FNPGMRELOCATE, Relocate a floating MMIO/MMIO2 range.}
2650	* @sa pgmR3PhysRamRangeRelocate
2651	*/
2652	static DECLCALLBACK(bool) pgmR3PhysMmio2RangeRelocate(PVM pVM, RTGCPTR GCPtrOld, RTGCPTR GCPtrNew,
2653	PGMRELOCATECALL enmMode, void *pvUser)
2654	{
2655	PPGMREGMMIO2RANGE pMmio = (PPGMREGMMIO2RANGE)pvUser;
2656	Assert(pMmio->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING);
2657	Assert(pMmio->RamRange.pSelfRC == GCPtrOld + PAGE_SIZE + RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange)); RT_NOREF_PV(GCPtrOld);
2658
2659	switch (enmMode)
2660	{
2661	case PGMRELOCATECALL_SUGGEST:
2662	return true;
2663
2664	case PGMRELOCATECALL_RELOCATE:
2665	{
2666	/*
2667	* Update myself, then relink all the ranges and flush the RC TLB.
2668	*/
2669	pgmLock(pVM);
2670
2671	pMmio->RamRange.pSelfRC = (RTRCPTR)(GCPtrNew + PAGE_SIZE + RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange));
2672
2673	pgmR3PhysRelinkRamRanges(pVM);
2674	for (unsigned i = 0; i < PGM_RAMRANGE_TLB_ENTRIES; i++)
2675	pVM->pgm.s.apRamRangesTlbRC[i] = NIL_RTRCPTR;
2676
2677	pgmUnlock(pVM);
2678	return true;
2679	}
2680
2681	default:
2682	AssertFailedReturn(false);
2683	}
2684	}
2685	#endif /* !PGM_WITHOUT_MAPPINGS */
2686
2687
2688	/**
2689	* Calculates the number of chunks
2690	*
2691	* @returns Number of registration chunk needed.
2692	* @param pVM The cross context VM structure.
2693	* @param cb The size of the MMIO/MMIO2 range.
2694	* @param pcPagesPerChunk Where to return the number of pages tracked by each
2695	* chunk. Optional.
2696	* @param pcbChunk Where to return the guest mapping size for a chunk.
2697	*/
2698	static uint16_t pgmR3PhysMmio2CalcChunkCount(PVM pVM, RTGCPHYS cb, uint32_t pcPagesPerChunk, uint32_t pcbChunk)
2699	{
2700	RT_NOREF_PV(pVM); /* without raw mode */
2701
2702	/*
2703	* This is the same calculation as PGMR3PhysRegisterRam does, except we'll be
2704	* needing a few bytes extra the PGMREGMMIO2RANGE structure.
2705	*
2706	* Note! In additions, we've got a 24 bit sub-page range for MMIO2 ranges, leaving
2707	* us with an absolute maximum of 16777215 pages per chunk (close to 64 GB).
2708	*/
2709	uint32_t cbChunk = 16U*_1M;
2710	uint32_t cPagesPerChunk = 1048048; /* max ~1048059 */
2711	AssertCompile(sizeof(PGMREGMMIO2RANGE) + sizeof(PGMPAGE) * 1048048 < 16U_1M - PAGE_SIZE 2);
2712	AssertRelease(cPagesPerChunk <= PGM_MMIO2_MAX_PAGE_COUNT); /* See above note. */
2713	AssertRelease(RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPagesPerChunk]) + PAGE_SIZE * 2 <= cbChunk);
2714	if (pcbChunk)
2715	*pcbChunk = cbChunk;
2716	if (pcPagesPerChunk)
2717	*pcPagesPerChunk = cPagesPerChunk;
2718
2719	/* Calc the number of chunks we need. */
2720	RTGCPHYS const cPages = cb >> X86_PAGE_SHIFT;
2721	uint16_t cChunks = (uint16_t)((cPages + cPagesPerChunk - 1) / cPagesPerChunk);
2722	AssertRelease((RTGCPHYS)cChunks * cPagesPerChunk >= cPages);
2723	return cChunks;
2724	}
2725
2726
2727	/**
2728	* Worker for PGMR3PhysMMIO2Register that allocates and the PGMREGMMIO2RANGE
2729	* structures and does basic initialization.
2730	*
2731	* Caller must set type specfic members and initialize the PGMPAGE structures.
2732	*
2733	* This was previously also used by PGMR3PhysMmio2PreRegister, a function for
2734	* pre-registering MMIO that was later (6.1) replaced by a new handle based IOM
2735	* interface. The reference to caller and type above is purely historical.
2736	*
2737	* @returns VBox status code.
2738	* @param pVM The cross context VM structure.
2739	* @param pDevIns The device instance owning the region.
2740	* @param iSubDev The sub-device number (internal PCI config number).
2741	* @param iRegion The region number. If the MMIO2 memory is a PCI
2742	* I/O region this number has to be the number of that
2743	* region. Otherwise it can be any number safe
2744	* UINT8_MAX.
2745	* @param cb The size of the region. Must be page aligned.
2746	* @param pszDesc The description.
2747	* @param ppHeadRet Where to return the pointer to the first
2748	* registration chunk.
2749	*
2750	* @thread EMT
2751	*/
2752	static int pgmR3PhysMmio2Create(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb,
2753	const char pszDesc, PPGMREGMMIO2RANGE ppHeadRet)
2754	{
2755	/*
2756	* Figure out how many chunks we need and of which size.
2757	*/
2758	uint32_t cPagesPerChunk;
2759	uint16_t cChunks = pgmR3PhysMmio2CalcChunkCount(pVM, cb, &cPagesPerChunk, NULL);
2760	AssertReturn(cChunks, VERR_PGM_PHYS_MMIO_EX_IPE);
2761
2762	/*
2763	* Allocate the chunks.
2764	*/
2765	PPGMREGMMIO2RANGE *ppNext = ppHeadRet;
2766	*ppNext = NULL;
2767
2768	int rc = VINF_SUCCESS;
2769	uint32_t cPagesLeft = cb >> X86_PAGE_SHIFT;
2770	for (uint16_t iChunk = 0; iChunk < cChunks && RT_SUCCESS(rc); iChunk++)
2771	{
2772	/*
2773	* We currently do a single RAM range for the whole thing. This will
2774	* probably have to change once someone needs really large MMIO regions,
2775	* as we will be running into SUPR3PageAllocEx limitations and such.
2776	*/
2777	const uint32_t cPagesTrackedByChunk = RT_MIN(cPagesLeft, cPagesPerChunk);
2778	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPagesTrackedByChunk]);
2779	PPGMREGMMIO2RANGE pNew = NULL;
2780	if ( iChunk + 1 < cChunks
2781	\|\| cbRange >= _1M)
2782	{
2783	/*
2784	* Allocate memory for the registration structure.
2785	*/
2786	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2787	size_t const cbChunk = (1 + cChunkPages + 1) << PAGE_SHIFT;
2788	AssertLogRelBreakStmt(cbChunk == (uint32_t)cbChunk, rc = VERR_OUT_OF_RANGE);
2789	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
2790	AssertBreakStmt(paChunkPages, rc = VERR_NO_TMP_MEMORY);
2791	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
2792	void *pvChunk = NULL;
2793	rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk, &R0PtrChunk, paChunkPages);
2794	AssertLogRelMsgRCBreakStmt(rc, ("rc=%Rrc, cChunkPages=%#zx\n", rc, cChunkPages), RTMemTmpFree(paChunkPages));
2795
2796	Assert(R0PtrChunk != NIL_RTR0PTR);
2797	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
2798
2799	pNew = (PPGMREGMMIO2RANGE)pvChunk;
2800	pNew->RamRange.fFlags = PGM_RAM_RANGE_FLAGS_FLOATING;
2801	pNew->RamRange.pSelfR0 = R0PtrChunk + RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange);
2802
2803	RTMemTmpFree(paChunkPages);
2804	}
2805	/*
2806	* Not so big, do a one time hyper allocation.
2807	*/
2808	else
2809	{
2810	rc = MMR3HyperAllocOnceNoRel(pVM, cbRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
2811	AssertLogRelMsgRCBreak(rc, ("cbRange=%zu\n", cbRange));
2812
2813	/*
2814	* Initialize allocation specific items.
2815	*/
2816	//pNew->RamRange.fFlags = 0;
2817	pNew->RamRange.pSelfR0 = MMHyperCCToR0(pVM, &pNew->RamRange);
2818	}
2819
2820	/*
2821	* Initialize the registration structure (caller does specific bits).
2822	*/
2823	pNew->pDevInsR3 = pDevIns;
2824	//pNew->pvR3 = NULL;
2825	//pNew->pNext = NULL;
2826	//pNew->fFlags = 0;
2827	if (iChunk == 0)
2828	pNew->fFlags \|= PGMREGMMIO2RANGE_F_FIRST_CHUNK;
2829	if (iChunk + 1 == cChunks)
2830	pNew->fFlags \|= PGMREGMMIO2RANGE_F_LAST_CHUNK;
2831	pNew->iSubDev = iSubDev;
2832	pNew->iRegion = iRegion;
2833	pNew->idSavedState = UINT8_MAX;
2834	pNew->idMmio2 = UINT8_MAX;
2835	//pNew->pPhysHandlerR3 = NULL;
2836	//pNew->paLSPages = NULL;
2837	pNew->RamRange.GCPhys = NIL_RTGCPHYS;
2838	pNew->RamRange.GCPhysLast = NIL_RTGCPHYS;
2839	pNew->RamRange.pszDesc = pszDesc;
2840	pNew->RamRange.cb = pNew->cbReal = (RTGCPHYS)cPagesTrackedByChunk << X86_PAGE_SHIFT;
2841	pNew->RamRange.fFlags \|= PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO_EX;
2842	//pNew->RamRange.pvR3 = NULL;
2843	//pNew->RamRange.paLSPages = NULL;
2844
2845	*ppNext = pNew;
2846	ASMCompilerBarrier();
2847	cPagesLeft -= cPagesTrackedByChunk;
2848	ppNext = &pNew->pNextR3;
2849	}
2850	Assert(cPagesLeft == 0);
2851
2852	if (RT_SUCCESS(rc))
2853	{
2854	Assert((*ppHeadRet)->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
2855	return VINF_SUCCESS;
2856	}
2857
2858	/*
2859	* Free floating ranges.
2860	*/
2861	while (*ppHeadRet)
2862	{
2863	PPGMREGMMIO2RANGE pFree = *ppHeadRet;
2864	*ppHeadRet = pFree->pNextR3;
2865
2866	if (pFree->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
2867	{
2868	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[pFree->RamRange.cb >> X86_PAGE_SHIFT]);
2869	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2870	SUPR3PageFreeEx(pFree, cChunkPages);
2871	}
2872	}
2873
2874	return rc;
2875	}
2876
2877
2878	/**
2879	* Common worker PGMR3PhysMmio2PreRegister & PGMR3PhysMMIO2Register that links a
2880	* complete registration entry into the lists and lookup tables.
2881	*
2882	* @param pVM The cross context VM structure.
2883	* @param pNew The new MMIO / MMIO2 registration to link.
2884	*/
2885	static void pgmR3PhysMmio2Link(PVM pVM, PPGMREGMMIO2RANGE pNew)
2886	{
2887	/*
2888	* Link it into the list (order doesn't matter, so insert it at the head).
2889	*
2890	* Note! The range we're linking may consist of multiple chunks, so we
2891	* have to find the last one.
2892	*/
2893	PPGMREGMMIO2RANGE pLast = pNew;
2894	for (pLast = pNew; ; pLast = pLast->pNextR3)
2895	{
2896	if (pLast->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2897	break;
2898	Assert(pLast->pNextR3);
2899	Assert(pLast->pNextR3->pDevInsR3 == pNew->pDevInsR3);
2900	Assert(pLast->pNextR3->iSubDev == pNew->iSubDev);
2901	Assert(pLast->pNextR3->iRegion == pNew->iRegion);
2902	Assert((pLast->pNextR3->fFlags & PGMREGMMIO2RANGE_F_MMIO2) == (pNew->fFlags & PGMREGMMIO2RANGE_F_MMIO2));
2903	Assert(pLast->pNextR3->idMmio2 == (pLast->fFlags & PGMREGMMIO2RANGE_F_MMIO2 ? pLast->idMmio2 + 1 : UINT8_MAX));
2904	}
2905
2906	pgmLock(pVM);
2907
2908	/* Link in the chain of ranges at the head of the list. */
2909	pLast->pNextR3 = pVM->pgm.s.pRegMmioRangesR3;
2910	pVM->pgm.s.pRegMmioRangesR3 = pNew;
2911
2912	/* If MMIO, insert the MMIO2 range/page IDs. */
2913	uint8_t idMmio2 = pNew->idMmio2;
2914	if (idMmio2 != UINT8_MAX)
2915	{
2916	for (;;)
2917	{
2918	Assert(pNew->fFlags & PGMREGMMIO2RANGE_F_MMIO2);
2919	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == NULL);
2920	Assert(pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] == NIL_RTR0PTR);
2921	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = pNew;
2922	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = pNew->RamRange.pSelfR0 - RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange);
2923	if (pNew->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2924	break;
2925	pNew = pNew->pNextR3;
2926	idMmio2++;
2927	}
2928	}
2929	else
2930	Assert(!(pNew->fFlags & PGMREGMMIO2RANGE_F_MMIO2));
2931
2932	pgmPhysInvalidatePageMapTLB(pVM);
2933	pgmUnlock(pVM);
2934	}
2935
2936
2937	/**
2938	* Allocate and register an MMIO2 region.
2939	*
2940	* As mentioned elsewhere, MMIO2 is just RAM spelled differently. It's RAM
2941	* associated with a device. It is also non-shared memory with a permanent
2942	* ring-3 mapping and page backing (presently).
2943	*
2944	* A MMIO2 range may overlap with base memory if a lot of RAM is configured for
2945	* the VM, in which case we'll drop the base memory pages. Presently we will
2946	* make no attempt to preserve anything that happens to be present in the base
2947	* memory that is replaced, this is of course incorrect but it's too much
2948	* effort.
2949	*
2950	* @returns VBox status code.
2951	* @retval VINF_SUCCESS on success, *ppv pointing to the R3 mapping of the
2952	* memory.
2953	* @retval VERR_ALREADY_EXISTS if the region already exists.
2954	*
2955	* @param pVM The cross context VM structure.
2956	* @param pDevIns The device instance owning the region.
2957	* @param iSubDev The sub-device number.
2958	* @param iRegion The region number. If the MMIO2 memory is a PCI
2959	* I/O region this number has to be the number of that
2960	* region. Otherwise it can be any number save
2961	* UINT8_MAX.
2962	* @param cb The size of the region. Must be page aligned.
2963	* @param fFlags Reserved for future use, must be zero.
2964	* @param pszDesc The description.
2965	* @param ppv Where to store the pointer to the ring-3 mapping of
2966	* the memory.
2967	* @param phRegion Where to return the MMIO2 region handle. Optional.
2968	* @thread EMT
2969	*/
2970	VMMR3_INT_DECL(int) PGMR3PhysMmio2Register(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb,
2971	uint32_t fFlags, const char pszDesc, void ppv, PGMMMIO2HANDLE phRegion)
2972	{
2973	/*
2974	* Validate input.
2975	*/
2976	AssertPtrReturn(ppv, VERR_INVALID_POINTER);
2977	*ppv = NULL;
2978	if (phRegion)
2979	{
2980	AssertPtrReturn(phRegion, VERR_INVALID_POINTER);
2981	*phRegion = NIL_PGMMMIO2HANDLE;
2982	}
2983	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2984	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2985	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
2986	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
2987	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2988	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2989	AssertReturn(pgmR3PhysMmio2Find(pVM, pDevIns, iSubDev, iRegion, NIL_PGMMMIO2HANDLE) == NULL, VERR_ALREADY_EXISTS);
2990	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2991	AssertReturn(cb, VERR_INVALID_PARAMETER);
2992	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2993
2994	const uint32_t cPages = cb >> PAGE_SHIFT;
2995	AssertLogRelReturn(((RTGCPHYS)cPages << PAGE_SHIFT) == cb, VERR_INVALID_PARAMETER);
2996	AssertLogRelReturn(cPages <= (MM_MMIO_64_MAX >> X86_PAGE_SHIFT), VERR_OUT_OF_RANGE);
2997	AssertLogRelReturn(cPages <= PGM_MMIO2_MAX_PAGE_COUNT, VERR_OUT_OF_RANGE);
2998
2999	/*
3000	* For the 2nd+ instance, mangle the description string so it's unique.
3001	*/
3002	if (pDevIns->iInstance > 0) /** @todo Move to PDMDevHlp.cpp and use a real string cache. */
3003	{
3004	pszDesc = MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s [%u]", pszDesc, pDevIns->iInstance);
3005	if (!pszDesc)
3006	return VERR_NO_MEMORY;
3007	}
3008
3009	/*
3010	* Allocate an MMIO2 range ID (not freed on failure).
3011	*
3012	* The zero ID is not used as it could be confused with NIL_GMM_PAGEID, so
3013	* the IDs goes from 1 thru PGM_MMIO2_MAX_RANGES.
3014	*/
3015	unsigned cChunks = pgmR3PhysMmio2CalcChunkCount(pVM, cb, NULL, NULL);
3016	pgmLock(pVM);
3017	uint8_t idMmio2 = pVM->pgm.s.cMmio2Regions + 1;
3018	unsigned cNewMmio2Regions = pVM->pgm.s.cMmio2Regions + cChunks;
3019	if (cNewMmio2Regions > PGM_MMIO2_MAX_RANGES)
3020	{
3021	pgmUnlock(pVM);
3022	AssertLogRelFailedReturn(VERR_PGM_TOO_MANY_MMIO2_RANGES);
3023	}
3024	pVM->pgm.s.cMmio2Regions = cNewMmio2Regions;
3025	pgmUnlock(pVM);
3026
3027	/*
3028	* Try reserve and allocate the backing memory first as this is what is
3029	* most likely to fail.
3030	*/
3031	int rc = MMR3AdjustFixedReservation(pVM, cPages, pszDesc);
3032	if (RT_SUCCESS(rc))
3033	{
3034	PSUPPAGE paPages = (PSUPPAGE)RTMemTmpAlloc(cPages * sizeof(SUPPAGE));
3035	if (RT_SUCCESS(rc))
3036	{
3037	void *pvPages;
3038	#if defined(VBOX_WITH_RAM_IN_KERNEL) && !defined(VBOX_WITH_LINEAR_HOST_PHYS_MEM)
3039	RTR0PTR pvPagesR0;
3040	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, &pvPagesR0, paPages);
3041	#else
3042	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, NULL /pR0Ptr/, paPages);
3043	#endif
3044	if (RT_SUCCESS(rc))
3045	{
3046	memset(pvPages, 0, cPages * PAGE_SIZE);
3047
3048	/*
3049	* Create the registered MMIO range record for it.
3050	*/
3051	PPGMREGMMIO2RANGE pNew;
3052	rc = pgmR3PhysMmio2Create(pVM, pDevIns, iSubDev, iRegion, cb, pszDesc, &pNew);
3053	if (RT_SUCCESS(rc))
3054	{
3055	if (phRegion)
3056	phRegion = idMmio2; / The ID of the first chunk. */
3057
3058	uint32_t iSrcPage = 0;
3059	uint8_t pbCurPages = (uint8_t )pvPages;
3060	for (PPGMREGMMIO2RANGE pCur = pNew; pCur; pCur = pCur->pNextR3)
3061	{
3062	pCur->pvR3 = pbCurPages;
3063	#if defined(VBOX_WITH_RAM_IN_KERNEL) && !defined(VBOX_WITH_LINEAR_HOST_PHYS_MEM)
3064	pCur->pvR0 = pvPagesR0 + (iSrcPage << PAGE_SHIFT);
3065	#endif
3066	pCur->RamRange.pvR3 = pbCurPages;
3067	pCur->idMmio2 = idMmio2;
3068	pCur->fFlags \|= PGMREGMMIO2RANGE_F_MMIO2;
3069
3070	uint32_t iDstPage = pCur->RamRange.cb >> X86_PAGE_SHIFT;
3071	while (iDstPage-- > 0)
3072	{
3073	PGM_PAGE_INIT(&pNew->RamRange.aPages[iDstPage],
3074	paPages[iDstPage + iSrcPage].Phys,
3075	PGM_MMIO2_PAGEID_MAKE(idMmio2, iDstPage),
3076	PGMPAGETYPE_MMIO2, PGM_PAGE_STATE_ALLOCATED);
3077	}
3078
3079	/* advance. */
3080	iSrcPage += pCur->RamRange.cb >> X86_PAGE_SHIFT;
3081	pbCurPages += pCur->RamRange.cb;
3082	idMmio2++;
3083	}
3084
3085	RTMemTmpFree(paPages);
3086
3087	/*
3088	* Update the page count stats, link the registration and we're done.
3089	*/
3090	pVM->pgm.s.cAllPages += cPages;
3091	pVM->pgm.s.cPrivatePages += cPages;
3092
3093	pgmR3PhysMmio2Link(pVM, pNew);
3094
3095	*ppv = pvPages;
3096	return VINF_SUCCESS;
3097	}
3098
3099	SUPR3PageFreeEx(pvPages, cPages);
3100	}
3101	}
3102	RTMemTmpFree(paPages);
3103	MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pszDesc);
3104	}
3105	if (pDevIns->iInstance > 0)
3106	MMR3HeapFree((void *)pszDesc);
3107	return rc;
3108	}
3109
3110
3111	/**
3112	* Deregisters and frees an MMIO2 region.
3113	*
3114	* Any physical access handlers registered for the region must be deregistered
3115	* before calling this function.
3116	*
3117	* @returns VBox status code.
3118	* @param pVM The cross context VM structure.
3119	* @param pDevIns The device instance owning the region.
3120	* @param hMmio2 The MMIO2 handle to deregister, or NIL if all
3121	* regions for the given device is to be deregistered.
3122	*/
3123	VMMR3_INT_DECL(int) PGMR3PhysMmio2Deregister(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3124	{
3125	/*
3126	* Validate input.
3127	*/
3128	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3129	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3130
3131	/*
3132	* The loop here scanning all registrations will make sure that multi-chunk ranges
3133	* get properly deregistered, though it's original purpose was the wildcard iRegion.
3134	*/
3135	pgmLock(pVM);
3136	int rc = VINF_SUCCESS;
3137	unsigned cFound = 0;
3138	PPGMREGMMIO2RANGE pPrev = NULL;
3139	PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3;
3140	while (pCur)
3141	{
3142	uint32_t const fFlags = pCur->fFlags;
3143	if ( pCur->pDevInsR3 == pDevIns
3144	&& ( hMmio2 == NIL_PGMMMIO2HANDLE
3145	\|\| pCur->idMmio2 == hMmio2))
3146	{
3147	Assert(fFlags & PGMREGMMIO2RANGE_F_MMIO2);
3148	cFound++;
3149
3150	/*
3151	* Unmap it if it's mapped.
3152	*/
3153	if (fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3154	{
3155	int rc2 = PGMR3PhysMmio2Unmap(pVM, pCur->pDevInsR3, pCur->idMmio2, pCur->RamRange.GCPhys);
3156	AssertRC(rc2);
3157	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3158	rc = rc2;
3159	}
3160
3161	/*
3162	* Unlink it
3163	*/
3164	PPGMREGMMIO2RANGE pNext = pCur->pNextR3;
3165	if (pPrev)
3166	pPrev->pNextR3 = pNext;
3167	else
3168	pVM->pgm.s.pRegMmioRangesR3 = pNext;
3169	pCur->pNextR3 = NULL;
3170
3171	uint8_t idMmio2 = pCur->idMmio2;
3172	if (idMmio2 != UINT8_MAX)
3173	{
3174	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == pCur);
3175	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = NULL;
3176	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = NIL_RTR0PTR;
3177	}
3178
3179	/*
3180	* Free the memory.
3181	*/
3182	const bool fIsMmio2 = RT_BOOL(fFlags & PGMREGMMIO2RANGE_F_MMIO2);
3183	uint32_t const cPages = pCur->cbReal >> PAGE_SHIFT;
3184	if (fIsMmio2)
3185	{
3186	int rc2 = SUPR3PageFreeEx(pCur->pvR3, cPages);
3187	AssertRC(rc2);
3188	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3189	rc = rc2;
3190
3191	rc2 = MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pCur->RamRange.pszDesc);
3192	AssertRC(rc2);
3193	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3194	rc = rc2;
3195	}
3196
3197	/* we're leaking hyper memory here if done at runtime. */
3198	#ifdef VBOX_STRICT
3199	VMSTATE const enmState = VMR3GetState(pVM);
3200	AssertMsg( enmState == VMSTATE_POWERING_OFF
3201	\|\| enmState == VMSTATE_POWERING_OFF_LS
3202	\|\| enmState == VMSTATE_OFF
3203	\|\| enmState == VMSTATE_OFF_LS
3204	\|\| enmState == VMSTATE_DESTROYING
3205	\|\| enmState == VMSTATE_TERMINATED
3206	\|\| enmState == VMSTATE_CREATING
3207	, ("%s\n", VMR3GetStateName(enmState)));
3208	#endif
3209
3210	if (pCur->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
3211	{
3212	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPages]);
3213	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
3214	SUPR3PageFreeEx(pCur, cChunkPages);
3215	}
3216	/*else
3217	{
3218	rc = MMHyperFree(pVM, pCur); - does not work, see the alloc call.
3219	AssertRCReturn(rc, rc);
3220	} */
3221
3222
3223	/* update page count stats */
3224	pVM->pgm.s.cAllPages -= cPages;
3225	if (fIsMmio2)
3226	pVM->pgm.s.cPrivatePages -= cPages;
3227	else
3228	pVM->pgm.s.cPureMmioPages -= cPages;
3229
3230	/* next */
3231	pCur = pNext;
3232	if (hMmio2 != NIL_PGMMMIO2HANDLE)
3233	{
3234	if (fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3235	break;
3236	hMmio2++;
3237	Assert(pCur->idMmio2 == hMmio2);
3238	Assert(pCur->pDevInsR3 == pDevIns);
3239	Assert(!(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK));
3240	}
3241	}
3242	else
3243	{
3244	pPrev = pCur;
3245	pCur = pCur->pNextR3;
3246	}
3247	}
3248	pgmPhysInvalidatePageMapTLB(pVM);
3249	pgmUnlock(pVM);
3250	return !cFound && hMmio2 != NIL_PGMMMIO2HANDLE ? VERR_NOT_FOUND : rc;
3251	}
3252
3253
3254	/**
3255	* Maps a MMIO2 region.
3256	*
3257	* This is typically done when a guest / the bios / state loading changes the
3258	* PCI config. The replacing of base memory has the same restrictions as during
3259	* registration, of course.
3260	*
3261	* @returns VBox status code.
3262	*
3263	* @param pVM The cross context VM structure.
3264	* @param pDevIns The device instance owning the region.
3265	* @param hMmio2 The handle of the region to map.
3266	* @param GCPhys The guest-physical address to be remapped.
3267	*/
3268	VMMR3_INT_DECL(int) PGMR3PhysMmio2Map(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS GCPhys)
3269	{
3270	/*
3271	* Validate input.
3272	*
3273	* Note! It's safe to walk the MMIO/MMIO2 list since registrations only
3274	* happens during VM construction.
3275	*/
3276	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3277	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3278	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
3279	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3280	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3281	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3282
3283	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3284	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3285	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3286
3287	PPGMREGMMIO2RANGE pLastMmio = pFirstMmio;
3288	RTGCPHYS cbRange = 0;
3289	for (;;)
3290	{
3291	AssertReturn(!(pLastMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED), VERR_WRONG_ORDER);
3292	Assert(pLastMmio->RamRange.GCPhys == NIL_RTGCPHYS);
3293	Assert(pLastMmio->RamRange.GCPhysLast == NIL_RTGCPHYS);
3294	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3295	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3296	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3297	cbRange += pLastMmio->RamRange.cb;
3298	if (pLastMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3299	break;
3300	pLastMmio = pLastMmio->pNextR3;
3301	}
3302
3303	RTGCPHYS GCPhysLast = GCPhys + cbRange - 1;
3304	AssertLogRelReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3305
3306	/*
3307	* Find our location in the ram range list, checking for restriction
3308	* we don't bother implementing yet (partially overlapping, multiple
3309	* ram ranges).
3310	*/
3311	pgmLock(pVM);
3312
3313	AssertReturnStmt(!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED), pgmUnlock(pVM), VERR_WRONG_ORDER);
3314
3315	bool fRamExists = false;
3316	PPGMRAMRANGE pRamPrev = NULL;
3317	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3318	while (pRam && GCPhysLast >= pRam->GCPhys)
3319	{
3320	if ( GCPhys <= pRam->GCPhysLast
3321	&& GCPhysLast >= pRam->GCPhys)
3322	{
3323	/* Completely within? */
3324	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
3325	&& GCPhysLast <= pRam->GCPhysLast,
3326	("%RGp-%RGp (MMIOEx/%s) falls partly outside %RGp-%RGp (%s)\n",
3327	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc,
3328	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3329	pgmUnlock(pVM),
3330	VERR_PGM_RAM_CONFLICT);
3331
3332	/* Check that all the pages are RAM pages. */
3333	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3334	uint32_t cPagesLeft = cbRange >> PAGE_SHIFT;
3335	while (cPagesLeft-- > 0)
3336	{
3337	AssertLogRelMsgReturnStmt(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
3338	("%RGp isn't a RAM page (%d) - mapping %RGp-%RGp (MMIO2/%s).\n",
3339	GCPhys, PGM_PAGE_GET_TYPE(pPage), GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc),
3340	pgmUnlock(pVM),
3341	VERR_PGM_RAM_CONFLICT);
3342	pPage++;
3343	}
3344
3345	/* There can only be one MMIO/MMIO2 chunk matching here! */
3346	AssertLogRelMsgReturnStmt(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK,
3347	("%RGp-%RGp (MMIOEx/%s, flags %#X) consists of multiple chunks whereas the RAM somehow doesn't!\n",
3348	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3349	pgmUnlock(pVM),
3350	VERR_PGM_PHYS_MMIO_EX_IPE);
3351
3352	fRamExists = true;
3353	break;
3354	}
3355
3356	/* next */
3357	pRamPrev = pRam;
3358	pRam = pRam->pNextR3;
3359	}
3360	Log(("PGMR3PhysMmio2Map: %RGp-%RGp fRamExists=%RTbool %s\n", GCPhys, GCPhysLast, fRamExists, pFirstMmio->RamRange.pszDesc));
3361
3362
3363	/*
3364	* Make the changes.
3365	*/
3366	RTGCPHYS GCPhysCur = GCPhys;
3367	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3368	{
3369	pCurMmio->RamRange.GCPhys = GCPhysCur;
3370	pCurMmio->RamRange.GCPhysLast = GCPhysCur + pCurMmio->RamRange.cb - 1;
3371	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3372	{
3373	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3374	break;
3375	}
3376	GCPhysCur += pCurMmio->RamRange.cb;
3377	}
3378
3379	if (fRamExists)
3380	{
3381	/*
3382	* Make all the pages in the range MMIO/ZERO pages, freeing any
3383	* RAM pages currently mapped here. This might not be 100% correct
3384	* for PCI memory, but we're doing the same thing for MMIO2 pages.
3385	*
3386	* We replace this MMIO/ZERO pages with real pages in the MMIO2 case.
3387	*/
3388	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK); /* Only one chunk */
3389
3390	int rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, PGMPAGETYPE_MMIO);
3391	AssertRCReturnStmt(rc, pgmUnlock(pVM), rc);
3392
3393	if (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MMIO2)
3394	{
3395	/* replace the pages, freeing all present RAM pages. */
3396	PPGMPAGE pPageSrc = &pFirstMmio->RamRange.aPages[0];
3397	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3398	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3399	while (cPagesLeft-- > 0)
3400	{
3401	Assert(PGM_PAGE_IS_MMIO(pPageDst));
3402
3403	RTHCPHYS const HCPhys = PGM_PAGE_GET_HCPHYS(pPageSrc);
3404	uint32_t const idPage = PGM_PAGE_GET_PAGEID(pPageSrc);
3405	PGM_PAGE_SET_PAGEID(pVM, pPageDst, idPage);
3406	PGM_PAGE_SET_HCPHYS(pVM, pPageDst, HCPhys);
3407	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO2);
3408	PGM_PAGE_SET_STATE(pVM, pPageDst, PGM_PAGE_STATE_ALLOCATED);
3409	PGM_PAGE_SET_PDE_TYPE(pVM, pPageDst, PGM_PAGE_PDE_TYPE_DONTCARE);
3410	PGM_PAGE_SET_PTE_INDEX(pVM, pPageDst, 0);
3411	PGM_PAGE_SET_TRACKING(pVM, pPageDst, 0);
3412	/* (We tell NEM at the end of the function.) */
3413
3414	pVM->pgm.s.cZeroPages--;
3415	GCPhys += PAGE_SIZE;
3416	pPageSrc++;
3417	pPageDst++;
3418	}
3419	}
3420
3421	/* Flush physical page map TLB. */
3422	pgmPhysInvalidatePageMapTLB(pVM);
3423
3424	/* Force a PGM pool flush as guest ram references have been changed. */
3425	/** @todo not entirely SMP safe; assuming for now the guest takes care of
3426	* this internally (not touch mapped mmio while changing the mapping). */
3427	PVMCPU pVCpu = VMMGetCpu(pVM);
3428	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3429	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3430	}
3431	else
3432	{
3433	/*
3434	* No RAM range, insert the ones prepared during registration.
3435	*/
3436	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3437	{
3438	/* Clear the tracking data of pages we're going to reactivate. */
3439	PPGMPAGE pPageSrc = &pCurMmio->RamRange.aPages[0];
3440	uint32_t cPagesLeft = pCurMmio->RamRange.cb >> PAGE_SHIFT;
3441	while (cPagesLeft-- > 0)
3442	{
3443	PGM_PAGE_SET_TRACKING(pVM, pPageSrc, 0);
3444	PGM_PAGE_SET_PTE_INDEX(pVM, pPageSrc, 0);
3445	pPageSrc++;
3446	}
3447
3448	/* link in the ram range */
3449	pgmR3PhysLinkRamRange(pVM, &pCurMmio->RamRange, pRamPrev);
3450
3451	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3452	{
3453	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3454	break;
3455	}
3456	pRamPrev = &pCurMmio->RamRange;
3457	}
3458	}
3459
3460	/*
3461	* Register the access handler if plain MMIO.
3462	*
3463	* We must register access handlers for each range since the access handler
3464	* code refuses to deal with multiple ranges (and we can).
3465	*/
3466	if (!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MMIO2))
3467	{
3468	AssertFailed();
3469	int rc = VINF_SUCCESS;
3470	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3471	{
3472	Assert(!(pCurMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED));
3473	rc = pgmHandlerPhysicalExRegister(pVM, pCurMmio->pPhysHandlerR3, pCurMmio->RamRange.GCPhys,
3474	pCurMmio->RamRange.GCPhysLast);
3475	if (RT_FAILURE(rc))
3476	break;
3477	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_MAPPED; /* Use this to mark that the handler is registered. */
3478	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3479	break;
3480	}
3481	if (RT_FAILURE(rc))
3482	{
3483	/* Almost impossible, but try clean up properly and get out of here. */
3484	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3485	{
3486	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3487	{
3488	pCurMmio->fFlags &= ~PGMREGMMIO2RANGE_F_MAPPED;
3489	pgmHandlerPhysicalExDeregister(pVM, pCurMmio->pPhysHandlerR3, fRamExists);
3490	}
3491
3492	if (!fRamExists)
3493	pgmR3PhysUnlinkRamRange(pVM, &pCurMmio->RamRange);
3494	else
3495	{
3496	Assert(pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK); /* Only one chunk */
3497
3498	uint32_t cPagesLeft = pCurMmio->RamRange.cb >> PAGE_SHIFT;
3499	PPGMPAGE pPageDst = &pRam->aPages[(pCurMmio->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3500	while (cPagesLeft-- > 0)
3501	{
3502	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
3503	pPageDst++;
3504	}
3505	}
3506
3507	pCurMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3508	pCurMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3509	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3510	break;
3511	}
3512
3513	pgmUnlock(pVM);
3514	return rc;
3515	}
3516	}
3517
3518	/*
3519	* We're good, set the flags and invalid the mapping TLB.
3520	*/
3521	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3522	{
3523	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_MAPPED;
3524	if (fRamExists)
3525	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_OVERLAPPING;
3526	else
3527	pCurMmio->fFlags &= ~PGMREGMMIO2RANGE_F_OVERLAPPING;
3528	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3529	break;
3530	}
3531	pgmPhysInvalidatePageMapTLB(pVM);
3532
3533	/*
3534	* Notify NEM while holding the lock (experimental) and REM without (like always).
3535	*/
3536	uint32_t const fNemNotify = (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MMIO2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0)
3537	\| (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_OVERLAPPING ? NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE : 0);
3538	int rc = NEMR3NotifyPhysMmioExMap(pVM, GCPhys, cbRange, fNemNotify, pFirstMmio->pvR3);
3539
3540	pgmUnlock(pVM);
3541
3542	return rc;
3543	}
3544
3545
3546	/**
3547	* Unmaps an MMIO2 region.
3548	*
3549	* This is typically done when a guest / the bios / state loading changes the
3550	* PCI config. The replacing of base memory has the same restrictions as during
3551	* registration, of course.
3552	*/
3553	VMMR3_INT_DECL(int) PGMR3PhysMmio2Unmap(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS GCPhys)
3554	{
3555	/*
3556	* Validate input
3557	*/
3558	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3559	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3560	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3561	if (GCPhys != NIL_RTGCPHYS)
3562	{
3563	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3564	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3565	}
3566
3567	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3568	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3569	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3570
3571	int rc = pgmLock(pVM);
3572	AssertRCReturn(rc, rc);
3573
3574	PPGMREGMMIO2RANGE pLastMmio = pFirstMmio;
3575	RTGCPHYS cbRange = 0;
3576	for (;;)
3577	{
3578	AssertReturnStmt(pLastMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED, pgmUnlock(pVM), VERR_WRONG_ORDER);
3579	AssertReturnStmt(pLastMmio->RamRange.GCPhys == GCPhys + cbRange \|\| GCPhys == NIL_RTGCPHYS, pgmUnlock(pVM), VERR_INVALID_PARAMETER);
3580	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3581	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3582	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3583	cbRange += pLastMmio->RamRange.cb;
3584	if (pLastMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3585	break;
3586	pLastMmio = pLastMmio->pNextR3;
3587	}
3588
3589	Log(("PGMR3PhysMmio2Unmap: %RGp-%RGp %s\n",
3590	pFirstMmio->RamRange.GCPhys, pLastMmio->RamRange.GCPhysLast, pFirstMmio->RamRange.pszDesc));
3591
3592	uint16_t const fOldFlags = pFirstMmio->fFlags;
3593	AssertReturnStmt(fOldFlags & PGMREGMMIO2RANGE_F_MAPPED, pgmUnlock(pVM), VERR_WRONG_ORDER);
3594
3595	/*
3596	* If plain MMIO, we must deregister the handlers first.
3597	*/
3598	if (!(fOldFlags & PGMREGMMIO2RANGE_F_MMIO2))
3599	{
3600	AssertFailed();
3601
3602	PPGMREGMMIO2RANGE pCurMmio = pFirstMmio;
3603	rc = pgmHandlerPhysicalExDeregister(pVM, pFirstMmio->pPhysHandlerR3, RT_BOOL(fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING));
3604	AssertRCReturnStmt(rc, pgmUnlock(pVM), rc);
3605	while (!(pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK))
3606	{
3607	pCurMmio = pCurMmio->pNextR3;
3608	rc = pgmHandlerPhysicalExDeregister(pVM, pCurMmio->pPhysHandlerR3, RT_BOOL(fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING));
3609	AssertRCReturnStmt(rc, pgmUnlock(pVM), VERR_PGM_PHYS_MMIO_EX_IPE);
3610	}
3611	}
3612
3613	/*
3614	* Unmap it.
3615	*/
3616	RTGCPHYS const GCPhysRangeNotify = pFirstMmio->RamRange.GCPhys;
3617	if (fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING)
3618	{
3619	/*
3620	* We've replaced RAM, replace with zero pages.
3621	*
3622	* Note! This is where we might differ a little from a real system, because
3623	* it's likely to just show the RAM pages as they were before the
3624	* MMIO/MMIO2 region was mapped here.
3625	*/
3626	/* Only one chunk allowed when overlapping! */
3627	Assert(fOldFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK);
3628
3629	/* Restore the RAM pages we've replaced. */
3630	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3631	while (pRam->GCPhys > pFirstMmio->RamRange.GCPhysLast)
3632	pRam = pRam->pNextR3;
3633
3634	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3635	if (fOldFlags & PGMREGMMIO2RANGE_F_MMIO2)
3636	pVM->pgm.s.cZeroPages += cPagesLeft;
3637
3638	PPGMPAGE pPageDst = &pRam->aPages[(pFirstMmio->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3639	while (cPagesLeft-- > 0)
3640	{
3641	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
3642	pPageDst++;
3643	}
3644
3645	/* Flush physical page map TLB. */
3646	pgmPhysInvalidatePageMapTLB(pVM);
3647
3648	/* Update range state. */
3649	pFirstMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3650	pFirstMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3651	pFirstMmio->fFlags &= ~(PGMREGMMIO2RANGE_F_OVERLAPPING \| PGMREGMMIO2RANGE_F_MAPPED);
3652	}
3653	else
3654	{
3655	/*
3656	* Unlink the chunks related to the MMIO/MMIO2 region.
3657	*/
3658	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3659	{
3660	pgmR3PhysUnlinkRamRange(pVM, &pCurMmio->RamRange);
3661	pCurMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3662	pCurMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3663	pCurMmio->fFlags &= ~(PGMREGMMIO2RANGE_F_OVERLAPPING \| PGMREGMMIO2RANGE_F_MAPPED);
3664	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3665	break;
3666	}
3667	}
3668
3669	/* Force a PGM pool flush as guest ram references have been changed. */
3670	/** @todo not entirely SMP safe; assuming for now the guest takes care
3671	* of this internally (not touch mapped mmio while changing the
3672	* mapping). */
3673	PVMCPU pVCpu = VMMGetCpu(pVM);
3674	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3675	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3676
3677	pgmPhysInvalidatePageMapTLB(pVM);
3678	pgmPhysInvalidRamRangeTlbs(pVM);
3679
3680	/*
3681	* Notify NEM while holding the lock (experimental) and REM without (like always).
3682	*/
3683	uint32_t const fNemFlags = (fOldFlags & PGMREGMMIO2RANGE_F_MMIO2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0)
3684	\| (fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING ? NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE : 0);
3685	rc = NEMR3NotifyPhysMmioExUnmap(pVM, GCPhysRangeNotify, cbRange, fNemFlags);
3686
3687	pgmUnlock(pVM);
3688	return rc;
3689	}
3690
3691
3692	/**
3693	* Reduces the mapping size of a MMIO2 region.
3694	*
3695	* This is mainly for dealing with old saved states after changing the default
3696	* size of a mapping region. See PGMDevHlpMMIOExReduce and
3697	* PDMPCIDEV::pfnRegionLoadChangeHookR3.
3698	*
3699	* The region must not currently be mapped when making this call. The VM state
3700	* must be state restore or VM construction.
3701	*
3702	* @returns VBox status code.
3703	* @param pVM The cross context VM structure.
3704	* @param pDevIns The device instance owning the region.
3705	* @param hMmio2 The handle of the region to reduce.
3706	* @param cbRegion The new mapping size.
3707	*/
3708	VMMR3_INT_DECL(int) PGMR3PhysMmio2Reduce(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS cbRegion)
3709	{
3710	/*
3711	* Validate input
3712	*/
3713	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3714	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3715	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3716	AssertReturn(cbRegion >= X86_PAGE_SIZE, VERR_INVALID_PARAMETER);
3717	AssertReturn(!(cbRegion & X86_PAGE_OFFSET_MASK), VERR_UNSUPPORTED_ALIGNMENT);
3718	VMSTATE enmVmState = VMR3GetState(pVM);
3719	AssertLogRelMsgReturn( enmVmState == VMSTATE_CREATING
3720	\|\| enmVmState == VMSTATE_LOADING,
3721	("enmVmState=%d (%s)\n", enmVmState, VMR3GetStateName(enmVmState)),
3722	VERR_VM_INVALID_VM_STATE);
3723
3724	int rc = pgmLock(pVM);
3725	AssertRCReturn(rc, rc);
3726
3727	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3728	if (pFirstMmio)
3729	{
3730	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3731	if (!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED))
3732	{
3733	/*
3734	* NOTE! Current implementation does not support multiple ranges.
3735	* Implement when there is a real world need and thus a testcase.
3736	*/
3737	AssertLogRelMsgStmt(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK,
3738	("%s: %#x\n", pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3739	rc = VERR_NOT_SUPPORTED);
3740	if (RT_SUCCESS(rc))
3741	{
3742	/*
3743	* Make the change.
3744	*/
3745	Log(("PGMR3PhysMmio2Reduce: %s changes from %RGp bytes (%RGp) to %RGp bytes.\n",
3746	pFirstMmio->RamRange.pszDesc, pFirstMmio->RamRange.cb, pFirstMmio->cbReal, cbRegion));
3747
3748	AssertLogRelMsgStmt(cbRegion <= pFirstMmio->cbReal,
3749	("%s: cbRegion=%#RGp cbReal=%#RGp\n", pFirstMmio->RamRange.pszDesc, cbRegion, pFirstMmio->cbReal),
3750	rc = VERR_OUT_OF_RANGE);
3751	if (RT_SUCCESS(rc))
3752	{
3753	pFirstMmio->RamRange.cb = cbRegion;
3754	}
3755	}
3756	}
3757	else
3758	rc = VERR_WRONG_ORDER;
3759	}
3760	else
3761	rc = VERR_NOT_FOUND;
3762
3763	pgmUnlock(pVM);
3764	return rc;
3765	}
3766
3767
3768	/**
3769	* Validates @a hMmio2, making sure it belongs to @a pDevIns.
3770	*
3771	* @returns VBox status code.
3772	* @param pVM The cross context VM structure.
3773	* @param pDevIns The device which allegedly owns @a hMmio2.
3774	* @param hMmio2 The handle to validate.
3775	*/
3776	VMMR3_INT_DECL(int) PGMR3PhysMmio2ValidateHandle(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3777	{
3778	/*
3779	* Validate input
3780	*/
3781	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3782	AssertPtrReturn(pDevIns, VERR_INVALID_POINTER);
3783
3784	/*
3785	* Just do this the simple way. No need for locking as this is only taken at
3786	*/
3787	pgmLock(pVM);
3788	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3789	pgmUnlock(pVM);
3790	AssertReturn(pFirstMmio, VERR_INVALID_HANDLE);
3791	AssertReturn(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MMIO2, VERR_INVALID_HANDLE);
3792	AssertReturn(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, VERR_INVALID_HANDLE);
3793	return VINF_SUCCESS;
3794	}
3795
3796
3797	#ifndef PGM_WITHOUT_MAPPINGS
3798	/**
3799	* Gets the HC physical address of a page in the MMIO2 region.
3800	*
3801	* This is API is intended for MMHyper and shouldn't be called
3802	* by anyone else...
3803	*
3804	* @returns VBox status code.
3805	* @param pVM The cross context VM structure.
3806	* @param pDevIns The owner of the memory, optional.
3807	* @param iSubDev Sub-device number.
3808	* @param iRegion The region.
3809	* @param off The page expressed an offset into the MMIO2 region.
3810	* @param pHCPhys Where to store the result.
3811	*/
3812	VMMR3_INT_DECL(int) PGMR3PhysMMIO2GetHCPhys(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion,
3813	RTGCPHYS off, PRTHCPHYS pHCPhys)
3814	{
3815	/*
3816	* Validate input
3817	*/
3818	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3819	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3820	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3821	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3822
3823	pgmLock(pVM);
3824	PPGMREGMMIO2RANGE pCurMmio = pgmR3PhysMmio2Find(pVM, pDevIns, iSubDev, iRegion, NIL_PGMMMIO2HANDLE);
3825	AssertReturn(pCurMmio, VERR_NOT_FOUND);
3826	AssertReturn(pCurMmio->fFlags & (PGMREGMMIO2RANGE_F_MMIO2 \| PGMREGMMIO2RANGE_F_FIRST_CHUNK), VERR_WRONG_TYPE);
3827
3828	while ( off >= pCurMmio->RamRange.cb
3829	&& !(pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK))
3830	{
3831	off -= pCurMmio->RamRange.cb;
3832	pCurMmio = pCurMmio->pNextR3;
3833	}
3834	AssertReturn(off < pCurMmio->RamRange.cb, VERR_INVALID_PARAMETER);
3835
3836	PCPGMPAGE pPage = &pCurMmio->RamRange.aPages[off >> PAGE_SHIFT];
3837	*pHCPhys = PGM_PAGE_GET_HCPHYS(pPage);
3838	pgmUnlock(pVM);
3839	return VINF_SUCCESS;
3840	}
3841	#endif /* !PGM_WITHOUT_MAPPINGS */
3842
3843
3844	/**
3845	* Gets the mapping address of an MMIO2 region.
3846	*
3847	* @returns Mapping address, NIL_RTGCPHYS if not mapped or invalid handle.
3848	*
3849	* @param pVM The cross context VM structure.
3850	* @param pDevIns The device owning the MMIO2 handle.
3851	* @param hMmio2 The region handle.
3852	*/
3853	VMMR3_INT_DECL(RTGCPHYS) PGMR3PhysMmio2GetMappingAddress(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3854	{
3855	AssertPtrReturn(pDevIns, NIL_RTGCPHYS);
3856
3857	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3858	AssertReturn(pFirstRegMmio, NIL_RTGCPHYS);
3859
3860	if (pFirstRegMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3861	return pFirstRegMmio->RamRange.GCPhys;
3862	return NIL_RTGCPHYS;
3863	}
3864
3865	/**
3866	* Changes the region number of an MMIO2 region.
3867	*
3868	* This is only for dealing with save state issues, nothing else.
3869	*
3870	* @return VBox status code.
3871	*
3872	* @param pVM The cross context VM structure.
3873	* @param pDevIns The device owning the MMIO2 memory.
3874	* @param hMmio2 The handle of the region.
3875	* @param iNewRegion The new region index.
3876	*
3877	* @thread EMT(0)
3878	* @sa @bugref{9359}
3879	*/
3880	VMMR3_INT_DECL(int) PGMR3PhysMmio2ChangeRegionNo(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, uint32_t iNewRegion)
3881	{
3882	/*
3883	* Validate input.
3884	*/
3885	VM_ASSERT_EMT0_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3886	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_LOADING, VERR_VM_INVALID_VM_STATE);
3887	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3888	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3889	AssertReturn(iNewRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3890
3891	AssertReturn(pVM->enmVMState == VMSTATE_LOADING, VERR_INVALID_STATE);
3892
3893	int rc = pgmLock(pVM);
3894	AssertRCReturn(rc, rc);
3895
3896	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3897	AssertReturnStmt(pFirstRegMmio, pgmUnlock(pVM), VERR_NOT_FOUND);
3898	AssertReturnStmt(pgmR3PhysMmio2Find(pVM, pDevIns, pFirstRegMmio->iSubDev, iNewRegion, NIL_PGMMMIO2HANDLE) == NULL,
3899	pgmUnlock(pVM), VERR_RESOURCE_IN_USE);
3900
3901	/*
3902	* Make the change.
3903	*/
3904	pFirstRegMmio->iRegion = (uint8_t)iNewRegion;
3905
3906	pgmUnlock(pVM);
3907	return VINF_SUCCESS;
3908	}
3909
3910
3911	/**
3912	* Worker for PGMR3PhysRomRegister.
3913	*
3914	* This is here to simplify lock management, i.e. the caller does all the
3915	* locking and we can simply return without needing to remember to unlock
3916	* anything first.
3917	*
3918	* @returns VBox status code.
3919	* @param pVM The cross context VM structure.
3920	* @param pDevIns The device instance owning the ROM.
3921	* @param GCPhys First physical address in the range.
3922	* Must be page aligned!
3923	* @param cb The size of the range (in bytes).
3924	* Must be page aligned!
3925	* @param pvBinary Pointer to the binary data backing the ROM image.
3926	* @param cbBinary The size of the binary data pvBinary points to.
3927	* This must be less or equal to @a cb.
3928	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
3929	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
3930	* @param pszDesc Pointer to description string. This must not be freed.
3931	*/
3932	static int pgmR3PhysRomRegisterLocked(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
3933	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
3934	{
3935	/*
3936	* Validate input.
3937	*/
3938	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3939	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
3940	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
3941	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
3942	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3943	AssertPtrReturn(pvBinary, VERR_INVALID_PARAMETER);
3944	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
3945	AssertReturn(!(fFlags & ~PGMPHYS_ROM_FLAGS_VALID_MASK), VERR_INVALID_PARAMETER);
3946	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_CREATING, VERR_VM_INVALID_VM_STATE);
3947
3948	const uint32_t cPages = cb >> PAGE_SHIFT;
3949
3950	/*
3951	* Find the ROM location in the ROM list first.
3952	*/
3953	PPGMROMRANGE pRomPrev = NULL;
3954	PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3;
3955	while (pRom && GCPhysLast >= pRom->GCPhys)
3956	{
3957	if ( GCPhys <= pRom->GCPhysLast
3958	&& GCPhysLast >= pRom->GCPhys)
3959	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
3960	GCPhys, GCPhysLast, pszDesc,
3961	pRom->GCPhys, pRom->GCPhysLast, pRom->pszDesc),
3962	VERR_PGM_RAM_CONFLICT);
3963	/* next */
3964	pRomPrev = pRom;
3965	pRom = pRom->pNextR3;
3966	}
3967
3968	/*
3969	* Find the RAM location and check for conflicts.
3970	*
3971	* Conflict detection is a bit different than for RAM
3972	* registration since a ROM can be located within a RAM
3973	* range. So, what we have to check for is other memory
3974	* types (other than RAM that is) and that we don't span
3975	* more than one RAM range (layz).
3976	*/
3977	bool fRamExists = false;
3978	PPGMRAMRANGE pRamPrev = NULL;
3979	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3980	while (pRam && GCPhysLast >= pRam->GCPhys)
3981	{
3982	if ( GCPhys <= pRam->GCPhysLast
3983	&& GCPhysLast >= pRam->GCPhys)
3984	{
3985	/* completely within? */
3986	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
3987	&& GCPhysLast <= pRam->GCPhysLast,
3988	("%RGp-%RGp (%s) falls partly outside %RGp-%RGp (%s)\n",
3989	GCPhys, GCPhysLast, pszDesc,
3990	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3991	VERR_PGM_RAM_CONFLICT);
3992	fRamExists = true;
3993	break;
3994	}
3995
3996	/* next */
3997	pRamPrev = pRam;
3998	pRam = pRam->pNextR3;
3999	}
4000	if (fRamExists)
4001	{
4002	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
4003	uint32_t cPagesLeft = cPages;
4004	while (cPagesLeft-- > 0)
4005	{
4006	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
4007	("%RGp (%R[pgmpage]) isn't a RAM page - registering %RGp-%RGp (%s).\n",
4008	pRam->GCPhys + ((RTGCPHYS)(uintptr_t)(pPage - &pRam->aPages[0]) << PAGE_SHIFT),
4009	pPage, GCPhys, GCPhysLast, pszDesc), VERR_PGM_RAM_CONFLICT);
4010	Assert(PGM_PAGE_IS_ZERO(pPage));
4011	pPage++;
4012	}
4013	}
4014
4015	/*
4016	* Update the base memory reservation if necessary.
4017	*/
4018	uint32_t cExtraBaseCost = fRamExists ? 0 : cPages;
4019	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4020	cExtraBaseCost += cPages;
4021	if (cExtraBaseCost)
4022	{
4023	int rc = MMR3IncreaseBaseReservation(pVM, cExtraBaseCost);
4024	if (RT_FAILURE(rc))
4025	return rc;
4026	}
4027
4028	/*
4029	* Allocate memory for the virgin copy of the RAM.
4030	*/
4031	PGMMALLOCATEPAGESREQ pReq;
4032	int rc = GMMR3AllocatePagesPrepare(pVM, &pReq, cPages, GMMACCOUNT_BASE);
4033	AssertRCReturn(rc, rc);
4034
4035	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4036	{
4037	pReq->aPages[iPage].HCPhysGCPhys = GCPhys + (iPage << PAGE_SHIFT);
4038	pReq->aPages[iPage].idPage = NIL_GMM_PAGEID;
4039	pReq->aPages[iPage].idSharedPage = NIL_GMM_PAGEID;
4040	}
4041
4042	rc = GMMR3AllocatePagesPerform(pVM, pReq);
4043	if (RT_FAILURE(rc))
4044	{
4045	GMMR3AllocatePagesCleanup(pReq);
4046	return rc;
4047	}
4048
4049	/*
4050	* Allocate the new ROM range and RAM range (if necessary).
4051	*/
4052	PPGMROMRANGE pRomNew;
4053	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMROMRANGE, aPages[cPages]), 0, MM_TAG_PGM_PHYS, (void **)&pRomNew);
4054	if (RT_SUCCESS(rc))
4055	{
4056	PPGMRAMRANGE pRamNew = NULL;
4057	if (!fRamExists)
4058	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), sizeof(PGMPAGE), MM_TAG_PGM_PHYS, (void **)&pRamNew);
4059	if (RT_SUCCESS(rc))
4060	{
4061	/*
4062	* Initialize and insert the RAM range (if required).
4063	*/
4064	PPGMROMPAGE pRomPage = &pRomNew->aPages[0];
4065	if (!fRamExists)
4066	{
4067	pRamNew->pSelfR0 = MMHyperCCToR0(pVM, pRamNew);
4068	pRamNew->GCPhys = GCPhys;
4069	pRamNew->GCPhysLast = GCPhysLast;
4070	pRamNew->cb = cb;
4071	pRamNew->pszDesc = pszDesc;
4072	pRamNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_ROM;
4073	pRamNew->pvR3 = NULL;
4074	pRamNew->paLSPages = NULL;
4075
4076	PPGMPAGE pPage = &pRamNew->aPages[0];
4077	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
4078	{
4079	PGM_PAGE_INIT(pPage,
4080	pReq->aPages[iPage].HCPhysGCPhys,
4081	pReq->aPages[iPage].idPage,
4082	PGMPAGETYPE_ROM,
4083	PGM_PAGE_STATE_ALLOCATED);
4084
4085	pRomPage->Virgin = *pPage;
4086	}
4087
4088	pVM->pgm.s.cAllPages += cPages;
4089	pgmR3PhysLinkRamRange(pVM, pRamNew, pRamPrev);
4090	}
4091	else
4092	{
4093	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
4094	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
4095	{
4096	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_ROM);
4097	PGM_PAGE_SET_HCPHYS(pVM, pPage, pReq->aPages[iPage].HCPhysGCPhys);
4098	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
4099	PGM_PAGE_SET_PAGEID(pVM, pPage, pReq->aPages[iPage].idPage);
4100	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4101	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
4102	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
4103
4104	pRomPage->Virgin = *pPage;
4105	}
4106
4107	pRamNew = pRam;
4108
4109	pVM->pgm.s.cZeroPages -= cPages;
4110	}
4111	pVM->pgm.s.cPrivatePages += cPages;
4112
4113	/* Flush physical page map TLB. */
4114	pgmPhysInvalidatePageMapTLB(pVM);
4115
4116
4117	/* Notify NEM before we register handlers. */
4118	uint32_t const fNemNotify = (fRamExists ? NEM_NOTIFY_PHYS_ROM_F_REPLACE : 0)
4119	\| (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED ? NEM_NOTIFY_PHYS_ROM_F_SHADOW : 0);
4120	rc = NEMR3NotifyPhysRomRegisterEarly(pVM, GCPhys, cb, fNemNotify);
4121
4122	/* Register the ROM access handler. */
4123	if (RT_SUCCESS(rc))
4124	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, pVM->pgm.s.hRomPhysHandlerType,
4125	pRomNew, MMHyperCCToR0(pVM, pRomNew), MMHyperCCToRC(pVM, pRomNew),
4126	pszDesc);
4127	if (RT_SUCCESS(rc))
4128	{
4129	/*
4130	* Copy the image over to the virgin pages.
4131	* This must be done after linking in the RAM range.
4132	*/
4133	size_t cbBinaryLeft = cbBinary;
4134	PPGMPAGE pRamPage = &pRamNew->aPages[(GCPhys - pRamNew->GCPhys) >> PAGE_SHIFT];
4135	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
4136	{
4137	void *pvDstPage;
4138	rc = pgmPhysPageMap(pVM, pRamPage, GCPhys + (iPage << PAGE_SHIFT), &pvDstPage);
4139	if (RT_FAILURE(rc))
4140	{
4141	VMSetError(pVM, rc, RT_SRC_POS, "Failed to map virgin ROM page at %RGp", GCPhys);
4142	break;
4143	}
4144	if (cbBinaryLeft >= PAGE_SIZE)
4145	{
4146	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), PAGE_SIZE);
4147	cbBinaryLeft -= PAGE_SIZE;
4148	}
4149	else
4150	{
4151	ASMMemZeroPage(pvDstPage); /* (shouldn't be necessary, but can't hurt either) */
4152	if (cbBinaryLeft > 0)
4153	{
4154	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), cbBinaryLeft);
4155	cbBinaryLeft = 0;
4156	}
4157	}
4158	}
4159	if (RT_SUCCESS(rc))
4160	{
4161	/*
4162	* Initialize the ROM range.
4163	* Note that the Virgin member of the pages has already been initialized above.
4164	*/
4165	pRomNew->GCPhys = GCPhys;
4166	pRomNew->GCPhysLast = GCPhysLast;
4167	pRomNew->cb = cb;
4168	pRomNew->fFlags = fFlags;
4169	pRomNew->idSavedState = UINT8_MAX;
4170	pRomNew->cbOriginal = cbBinary;
4171	pRomNew->pszDesc = pszDesc;
4172	pRomNew->pvOriginal = fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY
4173	? pvBinary : RTMemDup(pvBinary, cbBinary);
4174	if (pRomNew->pvOriginal)
4175	{
4176	for (unsigned iPage = 0; iPage < cPages; iPage++)
4177	{
4178	PPGMROMPAGE pPage = &pRomNew->aPages[iPage];
4179	pPage->enmProt = PGMROMPROT_READ_ROM_WRITE_IGNORE;
4180	PGM_PAGE_INIT_ZERO(&pPage->Shadow, pVM, PGMPAGETYPE_ROM_SHADOW);
4181	}
4182
4183	/* update the page count stats for the shadow pages. */
4184	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4185	{
4186	pVM->pgm.s.cZeroPages += cPages;
4187	pVM->pgm.s.cAllPages += cPages;
4188	}
4189
4190	/*
4191	* Insert the ROM range, tell REM and return successfully.
4192	*/
4193	pRomNew->pNextR3 = pRom;
4194	pRomNew->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4195
4196	if (pRomPrev)
4197	{
4198	pRomPrev->pNextR3 = pRomNew;
4199	pRomPrev->pNextR0 = MMHyperCCToR0(pVM, pRomNew);
4200	}
4201	else
4202	{
4203	pVM->pgm.s.pRomRangesR3 = pRomNew;
4204	pVM->pgm.s.pRomRangesR0 = MMHyperCCToR0(pVM, pRomNew);
4205	}
4206
4207	pgmPhysInvalidatePageMapTLB(pVM);
4208	GMMR3AllocatePagesCleanup(pReq);
4209
4210	/* Notify NEM again. */
4211	return NEMR3NotifyPhysRomRegisterLate(pVM, GCPhys, cb, fNemNotify);
4212	}
4213
4214	/* bail out */
4215	rc = VERR_NO_MEMORY;
4216	}
4217
4218	int rc2 = PGMHandlerPhysicalDeregister(pVM, GCPhys);
4219	AssertRC(rc2);
4220	}
4221
4222	if (!fRamExists)
4223	{
4224	pgmR3PhysUnlinkRamRange2(pVM, pRamNew, pRamPrev);
4225	MMHyperFree(pVM, pRamNew);
4226	}
4227	}
4228	MMHyperFree(pVM, pRomNew);
4229	}
4230
4231	/** @todo Purge the mapping cache or something... */
4232	GMMR3FreeAllocatedPages(pVM, pReq);
4233	GMMR3AllocatePagesCleanup(pReq);
4234	return rc;
4235	}
4236
4237
4238	/**
4239	* Registers a ROM image.
4240	*
4241	* Shadowed ROM images requires double the amount of backing memory, so,
4242	* don't use that unless you have to. Shadowing of ROM images is process
4243	* where we can select where the reads go and where the writes go. On real
4244	* hardware the chipset provides means to configure this. We provide
4245	* PGMR3PhysProtectROM() for this purpose.
4246	*
4247	* A read-only copy of the ROM image will always be kept around while we
4248	* will allocate RAM pages for the changes on demand (unless all memory
4249	* is configured to be preallocated).
4250	*
4251	* @returns VBox status code.
4252	* @param pVM The cross context VM structure.
4253	* @param pDevIns The device instance owning the ROM.
4254	* @param GCPhys First physical address in the range.
4255	* Must be page aligned!
4256	* @param cb The size of the range (in bytes).
4257	* Must be page aligned!
4258	* @param pvBinary Pointer to the binary data backing the ROM image.
4259	* @param cbBinary The size of the binary data pvBinary points to.
4260	* This must be less or equal to @a cb.
4261	* @param fFlags Mask of flags, PGMPHYS_ROM_FLAGS_XXX.
4262	* @param pszDesc Pointer to description string. This must not be freed.
4263	*
4264	* @remark There is no way to remove the rom, automatically on device cleanup or
4265	* manually from the device yet. This isn't difficult in any way, it's
4266	* just not something we expect to be necessary for a while.
4267	*/
4268	VMMR3DECL(int) PGMR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
4269	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
4270	{
4271	Log(("PGMR3PhysRomRegister: pDevIns=%p GCPhys=%RGp(-%RGp) cb=%RGp pvBinary=%p cbBinary=%#x fFlags=%#x pszDesc=%s\n",
4272	pDevIns, GCPhys, GCPhys + cb, cb, pvBinary, cbBinary, fFlags, pszDesc));
4273	pgmLock(pVM);
4274	int rc = pgmR3PhysRomRegisterLocked(pVM, pDevIns, GCPhys, cb, pvBinary, cbBinary, fFlags, pszDesc);
4275	pgmUnlock(pVM);
4276	return rc;
4277	}
4278
4279
4280	/**
4281	* Called by PGMR3MemSetup to reset the shadow, switch to the virgin, and verify
4282	* that the virgin part is untouched.
4283	*
4284	* This is done after the normal memory has been cleared.
4285	*
4286	* ASSUMES that the caller owns the PGM lock.
4287	*
4288	* @param pVM The cross context VM structure.
4289	*/
4290	int pgmR3PhysRomReset(PVM pVM)
4291	{
4292	PGM_LOCK_ASSERT_OWNER(pVM);
4293	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4294	{
4295	const uint32_t cPages = pRom->cb >> PAGE_SHIFT;
4296
4297	if (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4298	{
4299	/*
4300	* Reset the physical handler.
4301	*/
4302	int rc = PGMR3PhysRomProtect(pVM, pRom->GCPhys, pRom->cb, PGMROMPROT_READ_ROM_WRITE_IGNORE);
4303	AssertRCReturn(rc, rc);
4304
4305	/*
4306	* What we do with the shadow pages depends on the memory
4307	* preallocation option. If not enabled, we'll just throw
4308	* out all the dirty pages and replace them by the zero page.
4309	*/
4310	if (!pVM->pgm.s.fRamPreAlloc)
4311	{
4312	/* Free the dirty pages. */
4313	uint32_t cPendingPages = 0;
4314	PGMMFREEPAGESREQ pReq;
4315	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
4316	AssertRCReturn(rc, rc);
4317
4318	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4319	if ( !PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow)
4320	&& !PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow))
4321	{
4322	Assert(PGM_PAGE_GET_STATE(&pRom->aPages[iPage].Shadow) == PGM_PAGE_STATE_ALLOCATED);
4323	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, &pRom->aPages[iPage].Shadow,
4324	pRom->GCPhys + (iPage << PAGE_SHIFT),
4325	(PGMPAGETYPE)PGM_PAGE_GET_TYPE(&pRom->aPages[iPage].Shadow));
4326	AssertLogRelRCReturn(rc, rc);
4327	}
4328
4329	if (cPendingPages)
4330	{
4331	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
4332	AssertLogRelRCReturn(rc, rc);
4333	}
4334	GMMR3FreePagesCleanup(pReq);
4335	}
4336	else
4337	{
4338	/* clear all the shadow pages. */
4339	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4340	{
4341	if (PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow))
4342	continue;
4343	Assert(!PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow));
4344	void *pvDstPage;
4345	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4346	rc = pgmPhysPageMakeWritableAndMap(pVM, &pRom->aPages[iPage].Shadow, GCPhys, &pvDstPage);
4347	if (RT_FAILURE(rc))
4348	break;
4349	ASMMemZeroPage(pvDstPage);
4350	}
4351	AssertRCReturn(rc, rc);
4352	}
4353	}
4354
4355	/*
4356	* Restore the original ROM pages after a saved state load.
4357	* Also, in strict builds check that ROM pages remain unmodified.
4358	*/
4359	#ifndef VBOX_STRICT
4360	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4361	#endif
4362	{
4363	size_t cbSrcLeft = pRom->cbOriginal;
4364	uint8_t const pbSrcPage = (uint8_t const )pRom->pvOriginal;
4365	uint32_t cRestored = 0;
4366	for (uint32_t iPage = 0; iPage < cPages && cbSrcLeft > 0; iPage++, pbSrcPage += PAGE_SIZE)
4367	{
4368	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4369	void const *pvDstPage;
4370	int rc = pgmPhysPageMapReadOnly(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPage);
4371	if (RT_FAILURE(rc))
4372	break;
4373
4374	if (memcmp(pvDstPage, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE)))
4375	{
4376	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4377	{
4378	void *pvDstPageW;
4379	rc = pgmPhysPageMap(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPageW);
4380	AssertLogRelRCReturn(rc, rc);
4381	memcpy(pvDstPageW, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE));
4382	cRestored++;
4383	}
4384	else
4385	LogRel(("pgmR3PhysRomReset: %RGp: ROM page changed (%s)\n", GCPhys, pRom->pszDesc));
4386	}
4387	cbSrcLeft -= RT_MIN(cbSrcLeft, PAGE_SIZE);
4388	}
4389	if (cRestored > 0)
4390	LogRel(("PGM: ROM \"%s\": Reloaded %u of %u pages.\n", pRom->pszDesc, cRestored, cPages));
4391	}
4392	}
4393
4394	/* Clear the ROM restore flag now as we only need to do this once after
4395	loading saved state. */
4396	pVM->pgm.s.fRestoreRomPagesOnReset = false;
4397
4398	return VINF_SUCCESS;
4399	}
4400
4401
4402	/**
4403	* Called by PGMR3Term to free resources.
4404	*
4405	* ASSUMES that the caller owns the PGM lock.
4406	*
4407	* @param pVM The cross context VM structure.
4408	*/
4409	void pgmR3PhysRomTerm(PVM pVM)
4410	{
4411	/*
4412	* Free the heap copy of the original bits.
4413	*/
4414	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4415	{
4416	if ( pRom->pvOriginal
4417	&& !(pRom->fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY))
4418	{
4419	RTMemFree((void *)pRom->pvOriginal);
4420	pRom->pvOriginal = NULL;
4421	}
4422	}
4423	}
4424
4425
4426	/**
4427	* Change the shadowing of a range of ROM pages.
4428	*
4429	* This is intended for implementing chipset specific memory registers
4430	* and will not be very strict about the input. It will silently ignore
4431	* any pages that are not the part of a shadowed ROM.
4432	*
4433	* @returns VBox status code.
4434	* @retval VINF_PGM_SYNC_CR3
4435	*
4436	* @param pVM The cross context VM structure.
4437	* @param GCPhys Where to start. Page aligned.
4438	* @param cb How much to change. Page aligned.
4439	* @param enmProt The new ROM protection.
4440	*/
4441	VMMR3DECL(int) PGMR3PhysRomProtect(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMROMPROT enmProt)
4442	{
4443	/*
4444	* Check input
4445	*/
4446	if (!cb)
4447	return VINF_SUCCESS;
4448	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4449	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4450	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
4451	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
4452	AssertReturn(enmProt >= PGMROMPROT_INVALID && enmProt <= PGMROMPROT_END, VERR_INVALID_PARAMETER);
4453
4454	/*
4455	* Process the request.
4456	*/
4457	pgmLock(pVM);
4458	int rc = VINF_SUCCESS;
4459	bool fFlushTLB = false;
4460	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4461	{
4462	if ( GCPhys <= pRom->GCPhysLast
4463	&& GCPhysLast >= pRom->GCPhys
4464	&& (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED))
4465	{
4466	/*
4467	* Iterate the relevant pages and make necessary the changes.
4468	*/
4469	bool fChanges = false;
4470	uint32_t const cPages = pRom->GCPhysLast <= GCPhysLast
4471	? pRom->cb >> PAGE_SHIFT
4472	: (GCPhysLast - pRom->GCPhys + 1) >> PAGE_SHIFT;
4473	for (uint32_t iPage = (GCPhys - pRom->GCPhys) >> PAGE_SHIFT;
4474	iPage < cPages;
4475	iPage++)
4476	{
4477	PPGMROMPAGE pRomPage = &pRom->aPages[iPage];
4478	if (PGMROMPROT_IS_ROM(pRomPage->enmProt) != PGMROMPROT_IS_ROM(enmProt))
4479	{
4480	fChanges = true;
4481
4482	/* flush references to the page. */
4483	PPGMPAGE pRamPage = pgmPhysGetPage(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT));
4484	int rc2 = pgmPoolTrackUpdateGCPhys(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT), pRamPage,
4485	true /fFlushPTEs/, &fFlushTLB);
4486	if (rc2 != VINF_SUCCESS && (rc == VINF_SUCCESS \|\| RT_FAILURE(rc2)))
4487	rc = rc2;
4488	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pRamPage);
4489
4490	PPGMPAGE pOld = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Virgin : &pRomPage->Shadow;
4491	PPGMPAGE pNew = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Shadow : &pRomPage->Virgin;
4492
4493	pOld = pRamPage;
4494	pRamPage = pNew;
4495	/** @todo preserve the volatile flags (handlers) when these have been moved out of HCPhys! */
4496
4497	/* Tell NEM about the backing and protection change. */
4498	if (VM_IS_NEM_ENABLED(pVM))
4499	{
4500	PGMPAGETYPE enmType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pNew);
4501	NEMHCNotifyPhysPageChanged(pVM, GCPhys, PGM_PAGE_GET_HCPHYS(pOld), PGM_PAGE_GET_HCPHYS(pNew),
4502	pgmPhysPageCalcNemProtection(pRamPage, enmType), enmType, &u2State);
4503	PGM_PAGE_SET_NEM_STATE(pRamPage, u2State);
4504	}
4505	}
4506	pRomPage->enmProt = enmProt;
4507	}
4508
4509	/*
4510	* Reset the access handler if we made changes, no need
4511	* to optimize this.
4512	*/
4513	if (fChanges)
4514	{
4515	int rc2 = PGMHandlerPhysicalReset(pVM, pRom->GCPhys);
4516	if (RT_FAILURE(rc2))
4517	{
4518	pgmUnlock(pVM);
4519	AssertRC(rc);
4520	return rc2;
4521	}
4522	}
4523
4524	/* Advance - cb isn't updated. */
4525	GCPhys = pRom->GCPhys + (cPages << PAGE_SHIFT);
4526	}
4527	}
4528	pgmUnlock(pVM);
4529	if (fFlushTLB)
4530	PGM_INVL_ALL_VCPU_TLBS(pVM);
4531
4532	return rc;
4533	}
4534
4535
4536	/**
4537	* Sets the Address Gate 20 state.
4538	*
4539	* @param pVCpu The cross context virtual CPU structure.
4540	* @param fEnable True if the gate should be enabled.
4541	* False if the gate should be disabled.
4542	*/
4543	VMMDECL(void) PGMR3PhysSetA20(PVMCPU pVCpu, bool fEnable)
4544	{
4545	LogFlow(("PGMR3PhysSetA20 %d (was %d)\n", fEnable, pVCpu->pgm.s.fA20Enabled));
4546	if (pVCpu->pgm.s.fA20Enabled != fEnable)
4547	{
4548	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX
4549	PCCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
4550	if ( CPUMIsGuestInVmxRootMode(pCtx)
4551	&& !fEnable)
4552	{
4553	Log(("Cannot enter A20M mode while in VMX root mode\n"));
4554	return;
4555	}
4556	#endif
4557	pVCpu->pgm.s.fA20Enabled = fEnable;
4558	pVCpu->pgm.s.GCPhysA20Mask = ~((RTGCPHYS)!fEnable << 20);
4559	NEMR3NotifySetA20(pVCpu, fEnable);
4560	#ifdef PGM_WITH_A20
4561	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
4562	pgmR3RefreshShadowModeAfterA20Change(pVCpu);
4563	HMFlushTlb(pVCpu);
4564	#endif
4565	IEMTlbInvalidateAllPhysical(pVCpu);
4566	STAM_REL_COUNTER_INC(&pVCpu->pgm.s.cA20Changes);
4567	}
4568	}
4569
4570
4571	/**
4572	* Tree enumeration callback for dealing with age rollover.
4573	* It will perform a simple compression of the current age.
4574	*/
4575	static DECLCALLBACK(int) pgmR3PhysChunkAgeingRolloverCallback(PAVLU32NODECORE pNode, void *pvUser)
4576	{
4577	/* Age compression - ASSUMES iNow == 4. */
4578	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
4579	if (pChunk->iLastUsed >= UINT32_C(0xffffff00))
4580	pChunk->iLastUsed = 3;
4581	else if (pChunk->iLastUsed >= UINT32_C(0xfffff000))
4582	pChunk->iLastUsed = 2;
4583	else if (pChunk->iLastUsed)
4584	pChunk->iLastUsed = 1;
4585	else /* iLastUsed = 0 */
4586	pChunk->iLastUsed = 4;
4587
4588	NOREF(pvUser);
4589	return 0;
4590	}
4591
4592
4593	/**
4594	* The structure passed in the pvUser argument of pgmR3PhysChunkUnmapCandidateCallback().
4595	*/
4596	typedef struct PGMR3PHYSCHUNKUNMAPCB
4597	{
4598	PVM pVM; /*< Pointer to the VM. /
4599	PPGMCHUNKR3MAP pChunk; /*< The chunk to unmap. /
4600	} PGMR3PHYSCHUNKUNMAPCB, *PPGMR3PHYSCHUNKUNMAPCB;
4601
4602
4603	/**
4604	* Callback used to find the mapping that's been unused for
4605	* the longest time.
4606	*/
4607	static DECLCALLBACK(int) pgmR3PhysChunkUnmapCandidateCallback(PAVLU32NODECORE pNode, void *pvUser)
4608	{
4609	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
4610	PPGMR3PHYSCHUNKUNMAPCB pArg = (PPGMR3PHYSCHUNKUNMAPCB)pvUser;
4611
4612	/*
4613	* Check for locks and compare when last used.
4614	*/
4615	if (pChunk->cRefs)
4616	return 0;
4617	if (pChunk->cPermRefs)
4618	return 0;
4619	if ( pArg->pChunk
4620	&& pChunk->iLastUsed >= pArg->pChunk->iLastUsed)
4621	return 0;
4622
4623	/*
4624	* Check that it's not in any of the TLBs.
4625	*/
4626	PVM pVM = pArg->pVM;
4627	if ( pVM->pgm.s.ChunkR3Map.Tlb.aEntries[PGM_CHUNKR3MAPTLB_IDX(pChunk->Core.Key)].idChunk
4628	== pChunk->Core.Key)
4629	{
4630	pChunk = NULL;
4631	return 0;
4632	}
4633	#ifdef VBOX_STRICT
4634	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
4635	{
4636	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk != pChunk);
4637	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk != pChunk->Core.Key);
4638	}
4639	#endif
4640
4641	#ifndef VBOX_WITH_RAM_IN_KERNEL
4642	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.PhysTlbR0.aEntries); i++)
4643	if (pVM->pgm.s.PhysTlbR0.aEntries[i].pMap == pChunk)
4644	return 0;
4645	#endif
4646	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.PhysTlbR3.aEntries); i++)
4647	if (pVM->pgm.s.PhysTlbR3.aEntries[i].pMap == pChunk)
4648	return 0;
4649
4650	pArg->pChunk = pChunk;
4651	return 0;
4652	}
4653
4654
4655	/**
4656	* Finds a good candidate for unmapping when the ring-3 mapping cache is full.
4657	*
4658	* The candidate will not be part of any TLBs, so no need to flush
4659	* anything afterwards.
4660	*
4661	* @returns Chunk id.
4662	* @param pVM The cross context VM structure.
4663	*/
4664	static int32_t pgmR3PhysChunkFindUnmapCandidate(PVM pVM)
4665	{
4666	PGM_LOCK_ASSERT_OWNER(pVM);
4667
4668	/*
4669	* Enumerate the age tree starting with the left most node.
4670	*/
4671	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
4672	PGMR3PHYSCHUNKUNMAPCB Args;
4673	Args.pVM = pVM;
4674	Args.pChunk = NULL;
4675	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkUnmapCandidateCallback, &Args);
4676	Assert(Args.pChunk);
4677	if (Args.pChunk)
4678	{
4679	Assert(Args.pChunk->cRefs == 0);
4680	Assert(Args.pChunk->cPermRefs == 0);
4681	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
4682	return Args.pChunk->Core.Key;
4683	}
4684
4685	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
4686	return INT32_MAX;
4687	}
4688
4689
4690	/**
4691	* Rendezvous callback used by pgmR3PhysUnmapChunk that unmaps a chunk
4692	*
4693	* This is only called on one of the EMTs while the other ones are waiting for
4694	* it to complete this function.
4695	*
4696	* @returns VINF_SUCCESS (VBox strict status code).
4697	* @param pVM The cross context VM structure.
4698	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
4699	* @param pvUser User pointer. Unused
4700	*
4701	*/
4702	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysUnmapChunkRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
4703	{
4704	int rc = VINF_SUCCESS;
4705	pgmLock(pVM);
4706	NOREF(pVCpu); NOREF(pvUser);
4707
4708	if (pVM->pgm.s.ChunkR3Map.c >= pVM->pgm.s.ChunkR3Map.cMax)
4709	{
4710	/* Flush the pgm pool cache; call the internal rendezvous handler as we're already in a rendezvous handler here. */
4711	/** @todo also not really efficient to unmap a chunk that contains PD
4712	* or PT pages. */
4713	pgmR3PoolClearAllRendezvous(pVM, pVM->apCpusR3[0], NULL /* no need to flush the REM TLB as we already did that above */);
4714
4715	/*
4716	* Request the ring-0 part to unmap a chunk to make space in the mapping cache.
4717	*/
4718	GMMMAPUNMAPCHUNKREQ Req;
4719	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
4720	Req.Hdr.cbReq = sizeof(Req);
4721	Req.pvR3 = NULL;
4722	Req.idChunkMap = NIL_GMM_CHUNKID;
4723	Req.idChunkUnmap = pgmR3PhysChunkFindUnmapCandidate(pVM);
4724	if (Req.idChunkUnmap != INT32_MAX)
4725	{
4726	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkUnmap, a);
4727	rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
4728	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkUnmap, a);
4729	if (RT_SUCCESS(rc))
4730	{
4731	/*
4732	* Remove the unmapped one.
4733	*/
4734	PPGMCHUNKR3MAP pUnmappedChunk = (PPGMCHUNKR3MAP)RTAvlU32Remove(&pVM->pgm.s.ChunkR3Map.pTree, Req.idChunkUnmap);
4735	AssertRelease(pUnmappedChunk);
4736	AssertRelease(!pUnmappedChunk->cRefs);
4737	AssertRelease(!pUnmappedChunk->cPermRefs);
4738	pUnmappedChunk->pv = NULL;
4739	pUnmappedChunk->Core.Key = UINT32_MAX;
4740	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
4741	MMR3HeapFree(pUnmappedChunk);
4742	#else
4743	MMR3UkHeapFree(pVM, pUnmappedChunk, MM_TAG_PGM_CHUNK_MAPPING);
4744	#endif
4745	pVM->pgm.s.ChunkR3Map.c--;
4746	pVM->pgm.s.cUnmappedChunks++;
4747
4748	/*
4749	* Flush dangling PGM pointers (R3 & R0 ptrs to GC physical addresses).
4750	*/
4751	/** @todo We should not flush chunks which include cr3 mappings. */
4752	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
4753	{
4754	PPGMCPU pPGM = &pVM->apCpusR3[idCpu]->pgm.s;
4755
4756	pPGM->pGst32BitPdR3 = NULL;
4757	pPGM->pGstPaePdptR3 = NULL;
4758	pPGM->pGstAmd64Pml4R3 = NULL;
4759	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4760	pPGM->pGst32BitPdR0 = NIL_RTR0PTR;
4761	pPGM->pGstPaePdptR0 = NIL_RTR0PTR;
4762	pPGM->pGstAmd64Pml4R0 = NIL_RTR0PTR;
4763	#endif
4764	for (unsigned i = 0; i < RT_ELEMENTS(pPGM->apGstPaePDsR3); i++)
4765	{
4766	pPGM->apGstPaePDsR3[i] = NULL;
4767	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4768	pPGM->apGstPaePDsR0[i] = NIL_RTR0PTR;
4769	#endif
4770	}
4771
4772	/* Flush REM TLBs. */
4773	CPUMSetChangedFlags(pVM->apCpusR3[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
4774	}
4775	}
4776	}
4777	}
4778	pgmUnlock(pVM);
4779	return rc;
4780	}
4781
4782	/**
4783	* Unmap a chunk to free up virtual address space (request packet handler for pgmR3PhysChunkMap)
4784	*
4785	* @returns VBox status code.
4786	* @param pVM The cross context VM structure.
4787	*/
4788	void pgmR3PhysUnmapChunk(PVM pVM)
4789	{
4790	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysUnmapChunkRendezvous, NULL);
4791	AssertRC(rc);
4792	}
4793
4794
4795	/**
4796	* Maps the given chunk into the ring-3 mapping cache.
4797	*
4798	* This will call ring-0.
4799	*
4800	* @returns VBox status code.
4801	* @param pVM The cross context VM structure.
4802	* @param idChunk The chunk in question.
4803	* @param ppChunk Where to store the chunk tracking structure.
4804	*
4805	* @remarks Called from within the PGM critical section.
4806	* @remarks Can be called from any thread!
4807	*/
4808	int pgmR3PhysChunkMap(PVM pVM, uint32_t idChunk, PPPGMCHUNKR3MAP ppChunk)
4809	{
4810	int rc;
4811
4812	PGM_LOCK_ASSERT_OWNER(pVM);
4813
4814	/*
4815	* Move the chunk time forward.
4816	*/
4817	pVM->pgm.s.ChunkR3Map.iNow++;
4818	if (pVM->pgm.s.ChunkR3Map.iNow == 0)
4819	{
4820	pVM->pgm.s.ChunkR3Map.iNow = 4;
4821	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkAgeingRolloverCallback, NULL);
4822	}
4823
4824	/*
4825	* Allocate a new tracking structure first.
4826	*/
4827	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
4828	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3HeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk));
4829	#else
4830	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3UkHeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk), NULL);
4831	#endif
4832	AssertReturn(pChunk, VERR_NO_MEMORY);
4833	pChunk->Core.Key = idChunk;
4834	pChunk->iLastUsed = pVM->pgm.s.ChunkR3Map.iNow;
4835
4836	/*
4837	* Request the ring-0 part to map the chunk in question.
4838	*/
4839	GMMMAPUNMAPCHUNKREQ Req;
4840	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
4841	Req.Hdr.cbReq = sizeof(Req);
4842	Req.pvR3 = NULL;
4843	Req.idChunkMap = idChunk;
4844	Req.idChunkUnmap = NIL_GMM_CHUNKID;
4845
4846	/* Must be callable from any thread, so can't use VMMR3CallR0. */
4847	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkMap, a);
4848	rc = SUPR3CallVMMR0Ex(VMCC_GET_VMR0_FOR_CALL(pVM), NIL_VMCPUID, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
4849	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkMap, a);
4850	if (RT_SUCCESS(rc))
4851	{
4852	pChunk->pv = Req.pvR3;
4853
4854	/*
4855	* If we're running out of virtual address space, then we should
4856	* unmap another chunk.
4857	*
4858	* Currently, an unmap operation requires that all other virtual CPUs
4859	* are idling and not by chance making use of the memory we're
4860	* unmapping. So, we create an async unmap operation here.
4861	*
4862	* Now, when creating or restoring a saved state this wont work very
4863	* well since we may want to restore all guest RAM + a little something.
4864	* So, we have to do the unmap synchronously. Fortunately for us
4865	* though, during these operations the other virtual CPUs are inactive
4866	* and it should be safe to do this.
4867	*/
4868	/** @todo Eventually we should lock all memory when used and do
4869	* map+unmap as one kernel call without any rendezvous or
4870	* other precautions. */
4871	if (pVM->pgm.s.ChunkR3Map.c + 1 >= pVM->pgm.s.ChunkR3Map.cMax)
4872	{
4873	switch (VMR3GetState(pVM))
4874	{
4875	case VMSTATE_LOADING:
4876	case VMSTATE_SAVING:
4877	{
4878	PVMCPU pVCpu = VMMGetCpu(pVM);
4879	if ( pVCpu
4880	&& pVM->pgm.s.cDeprecatedPageLocks == 0)
4881	{
4882	pgmR3PhysUnmapChunkRendezvous(pVM, pVCpu, NULL);
4883	break;
4884	}
4885	}
4886	RT_FALL_THRU();
4887	default:
4888	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysUnmapChunk, 1, pVM);
4889	AssertRC(rc);
4890	break;
4891	}
4892	}
4893
4894	/*
4895	* Update the tree. We must do this after any unmapping to make sure
4896	* the chunk we're going to return isn't unmapped by accident.
4897	*/
4898	AssertPtr(Req.pvR3);
4899	bool fRc = RTAvlU32Insert(&pVM->pgm.s.ChunkR3Map.pTree, &pChunk->Core);
4900	AssertRelease(fRc);
4901	pVM->pgm.s.ChunkR3Map.c++;
4902	pVM->pgm.s.cMappedChunks++;
4903	}
4904	else
4905	{
4906	/** @todo this may fail because of /proc/sys/vm/max_map_count, so we
4907	* should probably restrict ourselves on linux. */
4908	AssertRC(rc);
4909	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
4910	MMR3HeapFree(pChunk);
4911	#else
4912	MMR3UkHeapFree(pVM, pChunk, MM_TAG_PGM_CHUNK_MAPPING);
4913	#endif
4914	pChunk = NULL;
4915	}
4916
4917	*ppChunk = pChunk;
4918	return rc;
4919	}
4920
4921
4922	/**
4923	* For VMMCALLRING3_PGM_MAP_CHUNK, considered internal.
4924	*
4925	* @returns see pgmR3PhysChunkMap.
4926	* @param pVM The cross context VM structure.
4927	* @param idChunk The chunk to map.
4928	*/
4929	VMMR3DECL(int) PGMR3PhysChunkMap(PVM pVM, uint32_t idChunk)
4930	{
4931	PPGMCHUNKR3MAP pChunk;
4932	int rc;
4933
4934	pgmLock(pVM);
4935	rc = pgmR3PhysChunkMap(pVM, idChunk, &pChunk);
4936	pgmUnlock(pVM);
4937	return rc;
4938	}
4939
4940
4941	/**
4942	* Invalidates the TLB for the ring-3 mapping cache.
4943	*
4944	* @param pVM The cross context VM structure.
4945	*/
4946	VMMR3DECL(void) PGMR3PhysChunkInvalidateTLB(PVM pVM)
4947	{
4948	pgmLock(pVM);
4949	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
4950	{
4951	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk = NIL_GMM_CHUNKID;
4952	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk = NULL;
4953	}
4954	/* The page map TLB references chunks, so invalidate that one too. */
4955	pgmPhysInvalidatePageMapTLB(pVM);
4956	pgmUnlock(pVM);
4957	}
4958
4959
4960	/**
4961	* Response to VMMCALLRING3_PGM_ALLOCATE_LARGE_HANDY_PAGE to allocate a large
4962	* (2MB) page for use with a nested paging PDE.
4963	*
4964	* @returns The following VBox status codes.
4965	* @retval VINF_SUCCESS on success.
4966	* @retval VINF_EM_NO_MEMORY if we're out of memory.
4967	*
4968	* @param pVM The cross context VM structure.
4969	* @param GCPhys GC physical start address of the 2 MB range
4970	*/
4971	VMMR3DECL(int) PGMR3PhysAllocateLargeHandyPage(PVM pVM, RTGCPHYS GCPhys)
4972	{
4973	#ifdef PGM_WITH_LARGE_PAGES
4974	uint64_t u64TimeStamp1, u64TimeStamp2;
4975
4976	pgmLock(pVM);
4977
4978	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatAllocLargePage, a);
4979	u64TimeStamp1 = RTTimeMilliTS();
4980	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_LARGE_HANDY_PAGE, 0, NULL);
4981	u64TimeStamp2 = RTTimeMilliTS();
4982	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatAllocLargePage, a);
4983	if (RT_SUCCESS(rc))
4984	{
4985	Assert(pVM->pgm.s.cLargeHandyPages == 1);
4986
4987	uint32_t idPage = pVM->pgm.s.aLargeHandyPage[0].idPage;
4988	RTHCPHYS HCPhys = pVM->pgm.s.aLargeHandyPage[0].HCPhysGCPhys;
4989
4990	void *pv;
4991
4992	/* Map the large page into our address space.
4993	*
4994	* Note: assuming that within the 2 MB range:
4995	* - GCPhys + PAGE_SIZE = HCPhys + PAGE_SIZE (whole point of this exercise)
4996	* - user space mapping is continuous as well
4997	* - page id (GCPhys) + 1 = page id (GCPhys + PAGE_SIZE)
4998	*/
4999	rc = pgmPhysPageMapByPageID(pVM, idPage, HCPhys, &pv);
5000	AssertLogRelMsg(RT_SUCCESS(rc), ("idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n", idPage, HCPhys, rc));
5001
5002	if (RT_SUCCESS(rc))
5003	{
5004	/*
5005	* Clear the pages.
5006	*/
5007	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatClearLargePage, b);
5008	for (unsigned i = 0; i < _2M/PAGE_SIZE; i++)
5009	{
5010	ASMMemZeroPage(pv);
5011
5012	PPGMPAGE pPage;
5013	rc = pgmPhysGetPageEx(pVM, GCPhys, &pPage);
5014	AssertRC(rc);
5015
5016	Assert(PGM_PAGE_IS_ZERO(pPage));
5017	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatRZPageReplaceZero);
5018	pVM->pgm.s.cZeroPages--;
5019
5020	/*
5021	* Do the PGMPAGE modifications.
5022	*/
5023	pVM->pgm.s.cPrivatePages++;
5024	PGM_PAGE_SET_HCPHYS(pVM, pPage, HCPhys);
5025	PGM_PAGE_SET_PAGEID(pVM, pPage, idPage);
5026	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
5027	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_PDE);
5028	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
5029	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
5030
5031	/* Somewhat dirty assumption that page ids are increasing. */
5032	idPage++;
5033
5034	HCPhys += PAGE_SIZE;
5035	GCPhys += PAGE_SIZE;
5036
5037	pv = (void *)((uintptr_t)pv + PAGE_SIZE);
5038
5039	Log3(("PGMR3PhysAllocateLargePage: idPage=%#x HCPhys=%RGp\n", idPage, HCPhys));
5040	}
5041	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatClearLargePage, b);
5042
5043	/* Flush all TLBs. */
5044	PGM_INVL_ALL_VCPU_TLBS(pVM);
5045	pgmPhysInvalidatePageMapTLB(pVM);
5046	}
5047	pVM->pgm.s.cLargeHandyPages = 0;
5048	}
5049
5050	if (RT_SUCCESS(rc))
5051	{
5052	static uint32_t cTimeOut = 0;
5053	uint64_t u64TimeStampDelta = u64TimeStamp2 - u64TimeStamp1;
5054
5055	if (u64TimeStampDelta > 100)
5056	{
5057	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatLargePageOverflow);
5058	if ( ++cTimeOut > 10
5059	\|\| u64TimeStampDelta > 1000 /* more than one second forces an early retirement from allocating large pages. */)
5060	{
5061	/* If repeated attempts to allocate a large page takes more than 100 ms, then we fall back to normal 4k pages.
5062	* E.g. Vista 64 tries to move memory around, which takes a huge amount of time.
5063	*/
5064	LogRel(("PGMR3PhysAllocateLargePage: allocating large pages takes too long (last attempt %d ms; nr of timeouts %d); DISABLE\n", u64TimeStampDelta, cTimeOut));
5065	PGMSetLargePageUsage(pVM, false);
5066	}
5067	}
5068	else
5069	if (cTimeOut > 0)
5070	cTimeOut--;
5071	}
5072
5073	pgmUnlock(pVM);
5074	return rc;
5075	#else
5076	RT_NOREF(pVM, GCPhys);
5077	return VERR_NOT_IMPLEMENTED;
5078	#endif /* PGM_WITH_LARGE_PAGES */
5079	}
5080
5081
5082	/**
5083	* Response to VM_FF_PGM_NEED_HANDY_PAGES and VMMCALLRING3_PGM_ALLOCATE_HANDY_PAGES.
5084	*
5085	* This function will also work the VM_FF_PGM_NO_MEMORY force action flag, to
5086	* signal and clear the out of memory condition. When contracted, this API is
5087	* used to try clear the condition when the user wants to resume.
5088	*
5089	* @returns The following VBox status codes.
5090	* @retval VINF_SUCCESS on success. FFs cleared.
5091	* @retval VINF_EM_NO_MEMORY if we're out of memory. The FF is not cleared in
5092	* this case and it gets accompanied by VM_FF_PGM_NO_MEMORY.
5093	*
5094	* @param pVM The cross context VM structure.
5095	*
5096	* @remarks The VINF_EM_NO_MEMORY status is for the benefit of the FF processing
5097	* in EM.cpp and shouldn't be propagated outside TRPM, HM, EM and
5098	* pgmPhysEnsureHandyPage. There is one exception to this in the \#PF
5099	* handler.
5100	*/
5101	VMMR3DECL(int) PGMR3PhysAllocateHandyPages(PVM pVM)
5102	{
5103	pgmLock(pVM);
5104
5105	/*
5106	* Allocate more pages, noting down the index of the first new page.
5107	*/
5108	uint32_t iClear = pVM->pgm.s.cHandyPages;
5109	AssertMsgReturn(iClear <= RT_ELEMENTS(pVM->pgm.s.aHandyPages), ("%d", iClear), VERR_PGM_HANDY_PAGE_IPE);
5110	Log(("PGMR3PhysAllocateHandyPages: %d -> %d\n", iClear, RT_ELEMENTS(pVM->pgm.s.aHandyPages)));
5111	int rcAlloc = VINF_SUCCESS;
5112	int rcSeed = VINF_SUCCESS;
5113	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
5114	while (rc == VERR_GMM_SEED_ME)
5115	{
5116	void *pvChunk;
5117	rcAlloc = rc = SUPR3PageAlloc(GMM_CHUNK_SIZE >> PAGE_SHIFT, &pvChunk);
5118	if (RT_SUCCESS(rc))
5119	{
5120	rcSeed = rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_SEED_CHUNK, (uintptr_t)pvChunk, NULL);
5121	if (RT_FAILURE(rc))
5122	SUPR3PageFree(pvChunk, GMM_CHUNK_SIZE >> PAGE_SHIFT);
5123	}
5124	if (RT_SUCCESS(rc))
5125	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
5126	}
5127
5128	/** @todo we should split this up into an allocate and flush operation. sometimes you want to flush and not allocate more (which will trigger the vm account limit error) */
5129	if ( rc == VERR_GMM_HIT_VM_ACCOUNT_LIMIT
5130	&& pVM->pgm.s.cHandyPages > 0)
5131	{
5132	/* Still handy pages left, so don't panic. */
5133	rc = VINF_SUCCESS;
5134	}
5135
5136	if (RT_SUCCESS(rc))
5137	{
5138	AssertMsg(rc == VINF_SUCCESS, ("%Rrc\n", rc));
5139	Assert(pVM->pgm.s.cHandyPages > 0);
5140	VM_FF_CLEAR(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
5141	VM_FF_CLEAR(pVM, VM_FF_PGM_NO_MEMORY);
5142
5143	#ifdef VBOX_STRICT
5144	uint32_t i;
5145	for (i = iClear; i < pVM->pgm.s.cHandyPages; i++)
5146	if ( pVM->pgm.s.aHandyPages[i].idPage == NIL_GMM_PAGEID
5147	\|\| pVM->pgm.s.aHandyPages[i].idSharedPage != NIL_GMM_PAGEID
5148	\|\| (pVM->pgm.s.aHandyPages[i].HCPhysGCPhys & PAGE_OFFSET_MASK))
5149	break;
5150	if (i != pVM->pgm.s.cHandyPages)
5151	{
5152	RTAssertMsg1Weak(NULL, __LINE__, __FILE__, __FUNCTION__);
5153	RTAssertMsg2Weak("i=%d iClear=%d cHandyPages=%d\n", i, iClear, pVM->pgm.s.cHandyPages);
5154	for (uint32_t j = iClear; j < pVM->pgm.s.cHandyPages; j++)
5155	RTAssertMsg2Add("%03d: idPage=%d HCPhysGCPhys=%RHp idSharedPage=%d%\n", j,
5156	pVM->pgm.s.aHandyPages[j].idPage,
5157	pVM->pgm.s.aHandyPages[j].HCPhysGCPhys,
5158	pVM->pgm.s.aHandyPages[j].idSharedPage,
5159	j == i ? " <---" : "");
5160	RTAssertPanic();
5161	}
5162	#endif
5163	/*
5164	* Clear the pages.
5165	*/
5166	while (iClear < pVM->pgm.s.cHandyPages)
5167	{
5168	PGMMPAGEDESC pPage = &pVM->pgm.s.aHandyPages[iClear];
5169	void *pv;
5170	rc = pgmPhysPageMapByPageID(pVM, pPage->idPage, pPage->HCPhysGCPhys, &pv);
5171	AssertLogRelMsgBreak(RT_SUCCESS(rc),
5172	("%u/%u: idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n",
5173	iClear, pVM->pgm.s.cHandyPages, pPage->idPage, pPage->HCPhysGCPhys, rc));
5174	ASMMemZeroPage(pv);
5175	iClear++;
5176	Log3(("PGMR3PhysAllocateHandyPages: idPage=%#x HCPhys=%RGp\n", pPage->idPage, pPage->HCPhysGCPhys));
5177	}
5178	}
5179	else
5180	{
5181	uint64_t cAllocPages, cMaxPages, cBalloonPages;
5182
5183	/*
5184	* We should never get here unless there is a genuine shortage of
5185	* memory (or some internal error). Flag the error so the VM can be
5186	* suspended ASAP and the user informed. If we're totally out of
5187	* handy pages we will return failure.
5188	*/
5189	/* Report the failure. */
5190	LogRel(("PGM: Failed to procure handy pages; rc=%Rrc rcAlloc=%Rrc rcSeed=%Rrc cHandyPages=%#x\n"
5191	" cAllPages=%#x cPrivatePages=%#x cSharedPages=%#x cZeroPages=%#x\n",
5192	rc, rcAlloc, rcSeed,
5193	pVM->pgm.s.cHandyPages,
5194	pVM->pgm.s.cAllPages,
5195	pVM->pgm.s.cPrivatePages,
5196	pVM->pgm.s.cSharedPages,
5197	pVM->pgm.s.cZeroPages));
5198
5199	if (GMMR3QueryMemoryStats(pVM, &cAllocPages, &cMaxPages, &cBalloonPages) == VINF_SUCCESS)
5200	{
5201	LogRel(("GMM: Statistics:\n"
5202	" Allocated pages: %RX64\n"
5203	" Maximum pages: %RX64\n"
5204	" Ballooned pages: %RX64\n", cAllocPages, cMaxPages, cBalloonPages));
5205	}
5206
5207	if ( rc != VERR_NO_MEMORY
5208	&& rc != VERR_NO_PHYS_MEMORY
5209	&& rc != VERR_LOCK_FAILED)
5210	{
5211	for (uint32_t i = 0; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
5212	{
5213	LogRel(("PGM: aHandyPages[#%#04x] = {.HCPhysGCPhys=%RHp, .idPage=%#08x, .idSharedPage=%#08x}\n",
5214	i, pVM->pgm.s.aHandyPages[i].HCPhysGCPhys, pVM->pgm.s.aHandyPages[i].idPage,
5215	pVM->pgm.s.aHandyPages[i].idSharedPage));
5216	uint32_t const idPage = pVM->pgm.s.aHandyPages[i].idPage;
5217	if (idPage != NIL_GMM_PAGEID)
5218	{
5219	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
5220	pRam;
5221	pRam = pRam->pNextR3)
5222	{
5223	uint32_t const cPages = pRam->cb >> PAGE_SHIFT;
5224	for (uint32_t iPage = 0; iPage < cPages; iPage++)
5225	if (PGM_PAGE_GET_PAGEID(&pRam->aPages[iPage]) == idPage)
5226	LogRel(("PGM: Used by %RGp %R[pgmpage] (%s)\n",
5227	pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pRam->aPages[iPage], pRam->pszDesc));
5228	}
5229	}
5230	}
5231	}
5232
5233	if (rc == VERR_NO_MEMORY)
5234	{
5235	uint64_t cbHostRamAvail = 0;
5236	int rc2 = RTSystemQueryAvailableRam(&cbHostRamAvail);
5237	if (RT_SUCCESS(rc2))
5238	LogRel(("Host RAM: %RU64MB available\n", cbHostRamAvail / _1M));
5239	else
5240	LogRel(("Cannot determine the amount of available host memory\n"));
5241	}
5242
5243	/* Set the FFs and adjust rc. */
5244	VM_FF_SET(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
5245	VM_FF_SET(pVM, VM_FF_PGM_NO_MEMORY);
5246	if ( rc == VERR_NO_MEMORY
5247	\|\| rc == VERR_NO_PHYS_MEMORY
5248	\|\| rc == VERR_LOCK_FAILED)
5249	rc = VINF_EM_NO_MEMORY;
5250	}
5251
5252	pgmUnlock(pVM);
5253	return rc;
5254	}
5255
5256
5257	/**
5258	* Frees the specified RAM page and replaces it with the ZERO page.
5259	*
5260	* This is used by ballooning, remapping MMIO2, RAM reset and state loading.
5261	*
5262	* @param pVM The cross context VM structure.
5263	* @param pReq Pointer to the request.
5264	* @param pcPendingPages Where the number of pages waiting to be freed are
5265	* kept. This will normally be incremented.
5266	* @param pPage Pointer to the page structure.
5267	* @param GCPhys The guest physical address of the page, if applicable.
5268	* @param enmNewType New page type for NEM notification, since several
5269	* callers will change the type upon successful return.
5270	*
5271	* @remarks The caller must own the PGM lock.
5272	*/
5273	int pgmPhysFreePage(PVM pVM, PGMMFREEPAGESREQ pReq, uint32_t *pcPendingPages, PPGMPAGE pPage, RTGCPHYS GCPhys,
5274	PGMPAGETYPE enmNewType)
5275	{
5276	/*
5277	* Assert sanity.
5278	*/
5279	PGM_LOCK_ASSERT_OWNER(pVM);
5280	if (RT_UNLIKELY( PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM
5281	&& PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_ROM_SHADOW))
5282	{
5283	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
5284	return VMSetError(pVM, VERR_PGM_PHYS_NOT_RAM, RT_SRC_POS, "GCPhys=%RGp type=%d", GCPhys, PGM_PAGE_GET_TYPE(pPage));
5285	}
5286
5287	/** @todo What about ballooning of large pages??! */
5288	Assert( PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE
5289	&& PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE_DISABLED);
5290
5291	if ( PGM_PAGE_IS_ZERO(pPage)
5292	\|\| PGM_PAGE_IS_BALLOONED(pPage))
5293	return VINF_SUCCESS;
5294
5295	const uint32_t idPage = PGM_PAGE_GET_PAGEID(pPage);
5296	Log3(("pgmPhysFreePage: idPage=%#x GCPhys=%RGp pPage=%R[pgmpage]\n", idPage, GCPhys, pPage));
5297	if (RT_UNLIKELY( idPage == NIL_GMM_PAGEID
5298	\|\| idPage > GMM_PAGEID_LAST
5299	\|\| PGM_PAGE_GET_CHUNKID(pPage) == NIL_GMM_CHUNKID))
5300	{
5301	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
5302	return VMSetError(pVM, VERR_PGM_PHYS_INVALID_PAGE_ID, RT_SRC_POS, "GCPhys=%RGp idPage=%#x", GCPhys, pPage);
5303	}
5304	const RTHCPHYS HCPhysPrev = PGM_PAGE_GET_HCPHYS(pPage);
5305
5306	/* update page count stats. */
5307	if (PGM_PAGE_IS_SHARED(pPage))
5308	pVM->pgm.s.cSharedPages--;
5309	else
5310	pVM->pgm.s.cPrivatePages--;
5311	pVM->pgm.s.cZeroPages++;
5312
5313	/* Deal with write monitored pages. */
5314	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
5315	{
5316	PGM_PAGE_SET_WRITTEN_TO(pVM, pPage);
5317	pVM->pgm.s.cWrittenToPages++;
5318	}
5319
5320	/*
5321	* pPage = ZERO page.
5322	*/
5323	PGM_PAGE_SET_HCPHYS(pVM, pPage, pVM->pgm.s.HCPhysZeroPg);
5324	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
5325	PGM_PAGE_SET_PAGEID(pVM, pPage, NIL_GMM_PAGEID);
5326	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
5327	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
5328	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
5329
5330	/* Flush physical page map TLB entry. */
5331	pgmPhysInvalidatePageMapTLBEntry(pVM, GCPhys);
5332
5333	/* Notify NEM. */
5334	/** @todo consider doing batch NEM notifications. */
5335	if (VM_IS_NEM_ENABLED(pVM))
5336	{
5337	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pPage);
5338	NEMHCNotifyPhysPageChanged(pVM, GCPhys, HCPhysPrev, pVM->pgm.s.HCPhysZeroPg,
5339	pgmPhysPageCalcNemProtection(pPage, enmNewType), enmNewType, &u2State);
5340	PGM_PAGE_SET_NEM_STATE(pPage, u2State);
5341	}
5342
5343	/*
5344	* Make sure it's not in the handy page array.
5345	*/
5346	for (uint32_t i = pVM->pgm.s.cHandyPages; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
5347	{
5348	if (pVM->pgm.s.aHandyPages[i].idPage == idPage)
5349	{
5350	pVM->pgm.s.aHandyPages[i].idPage = NIL_GMM_PAGEID;
5351	break;
5352	}
5353	if (pVM->pgm.s.aHandyPages[i].idSharedPage == idPage)
5354	{
5355	pVM->pgm.s.aHandyPages[i].idSharedPage = NIL_GMM_PAGEID;
5356	break;
5357	}
5358	}
5359
5360	/*
5361	* Push it onto the page array.
5362	*/
5363	uint32_t iPage = *pcPendingPages;
5364	Assert(iPage < PGMPHYS_FREE_PAGE_BATCH_SIZE);
5365	*pcPendingPages += 1;
5366
5367	pReq->aPages[iPage].idPage = idPage;
5368
5369	if (iPage + 1 < PGMPHYS_FREE_PAGE_BATCH_SIZE)
5370	return VINF_SUCCESS;
5371
5372	/*
5373	* Flush the pages.
5374	*/
5375	int rc = GMMR3FreePagesPerform(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE);
5376	if (RT_SUCCESS(rc))
5377	{
5378	GMMR3FreePagesRePrep(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
5379	*pcPendingPages = 0;
5380	}
5381	return rc;
5382	}
5383
5384
5385	/**
5386	* Converts a GC physical address to a HC ring-3 pointer, with some
5387	* additional checks.
5388	*
5389	* @returns VBox status code.
5390	* @retval VINF_SUCCESS on success.
5391	* @retval VINF_PGM_PHYS_TLB_CATCH_WRITE and *ppv set if the page has a write
5392	* access handler of some kind.
5393	* @retval VERR_PGM_PHYS_TLB_CATCH_ALL if the page has a handler catching all
5394	* accesses or is odd in any way.
5395	* @retval VERR_PGM_PHYS_TLB_UNASSIGNED if the page doesn't exist.
5396	*
5397	* @param pVM The cross context VM structure.
5398	* @param GCPhys The GC physical address to convert. Since this is only
5399	* used for filling the REM TLB, the A20 mask must be
5400	* applied before calling this API.
5401	* @param fWritable Whether write access is required.
5402	* @param ppv Where to store the pointer corresponding to GCPhys on
5403	* success.
5404	*/
5405	VMMR3DECL(int) PGMR3PhysTlbGCPhys2Ptr(PVM pVM, RTGCPHYS GCPhys, bool fWritable, void **ppv)
5406	{
5407	pgmLock(pVM);
5408	PGM_A20_ASSERT_MASKED(VMMGetCpu(pVM), GCPhys);
5409
5410	PPGMRAMRANGE pRam;
5411	PPGMPAGE pPage;
5412	int rc = pgmPhysGetPageAndRangeEx(pVM, GCPhys, &pPage, &pRam);
5413	if (RT_SUCCESS(rc))
5414	{
5415	if (PGM_PAGE_IS_BALLOONED(pPage))
5416	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5417	else if (!PGM_PAGE_HAS_ANY_HANDLERS(pPage))
5418	rc = VINF_SUCCESS;
5419	else
5420	{
5421	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)) /* catches MMIO */
5422	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
5423	else if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
5424	{
5425	/** @todo Handle TLB loads of virtual handlers so ./test.sh can be made to work
5426	* in -norawr0 mode. */
5427	if (fWritable)
5428	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5429	}
5430	else
5431	{
5432	/* Temporarily disabled physical handler(s), since the recompiler
5433	doesn't get notified when it's reset we'll have to pretend it's
5434	operating normally. */
5435	if (pgmHandlerPhysicalIsAll(pVM, GCPhys))
5436	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
5437	else
5438	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5439	}
5440	}
5441	if (RT_SUCCESS(rc))
5442	{
5443	int rc2;
5444
5445	/* Make sure what we return is writable. */
5446	if (fWritable)
5447	switch (PGM_PAGE_GET_STATE(pPage))
5448	{
5449	case PGM_PAGE_STATE_ALLOCATED:
5450	break;
5451	case PGM_PAGE_STATE_BALLOONED:
5452	AssertFailed();
5453	break;
5454	case PGM_PAGE_STATE_ZERO:
5455	case PGM_PAGE_STATE_SHARED:
5456	if (rc == VINF_PGM_PHYS_TLB_CATCH_WRITE)
5457	break;
5458	RT_FALL_THRU();
5459	case PGM_PAGE_STATE_WRITE_MONITORED:
5460	rc2 = pgmPhysPageMakeWritable(pVM, pPage, GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK);
5461	AssertLogRelRCReturn(rc2, rc2);
5462	break;
5463	}
5464
5465	/* Get a ring-3 mapping of the address. */
5466	PPGMPAGER3MAPTLBE pTlbe;
5467	rc2 = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
5468	AssertLogRelRCReturn(rc2, rc2);
5469	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
5470	/** @todo mapping/locking hell; this isn't horribly efficient since
5471	* pgmPhysPageLoadIntoTlb will repeat the lookup we've done here. */
5472
5473	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage] ppv=%p\n", GCPhys, rc, pPage, ppv));
5474	}
5475	else
5476	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage]\n", GCPhys, rc, pPage));
5477
5478	/* else: handler catching all access, no pointer returned. */
5479	}
5480	else
5481	rc = VERR_PGM_PHYS_TLB_UNASSIGNED;
5482
5483	pgmUnlock(pVM);
5484	return rc;
5485	}
5486

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

Download in other formats:

© 2023 Oracle
Contact – Privacy policy – Terms of Use