PGMPhys.cpp@ 93444

Last change on this file since 93444 was 93210, checked in by vboxsync, 3 years ago
VMM/PGMR3PhysMmio2Register: Call MMR3AdjustFixedReservation even in NEM mode. bugref:10122
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 223.5 KB

Line
1	/* $Id: PGMPhys.cpp 93210 2022-01-13 02:08:52Z vboxsync $ */
2	/** @file
3	* PGM - Page Manager and Monitor, Physical Memory Addressing.
4	*/
5
6	/*
7	* Copyright (C) 2006-2022 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	#define VBOX_WITHOUT_PAGING_BIT_FIELDS /* 64-bit bitfields are just asking for trouble. See @bugref{9841} and others. */
24	#include <VBox/vmm/pgm.h>
25	#include <VBox/vmm/iem.h>
26	#include <VBox/vmm/iom.h>
27	#include <VBox/vmm/mm.h>
28	#include <VBox/vmm/nem.h>
29	#include <VBox/vmm/stam.h>
30	#include <VBox/vmm/pdmdev.h>
31	#include "PGMInternal.h"
32	#include <VBox/vmm/vmcc.h>
33
34	#include "PGMInline.h"
35
36	#include <VBox/sup.h>
37	#include <VBox/param.h>
38	#include <VBox/err.h>
39	#include <VBox/log.h>
40	#include <iprt/assert.h>
41	#include <iprt/alloc.h>
42	#include <iprt/asm.h>
43	#ifdef VBOX_STRICT
44	# include <iprt/crc.h>
45	#endif
46	#include <iprt/thread.h>
47	#include <iprt/string.h>
48	#include <iprt/system.h>
49
50
51	/*********************************************************************************************************************************
52	* Defined Constants And Macros *
53	*********************************************************************************************************************************/
54	/** The number of pages to free in one batch. */
55	#define PGMPHYS_FREE_PAGE_BATCH_SIZE 128
56
57
58
59	/*********************************************************************************************************************************
60	* Reading and Writing Guest Pysical Memory *
61	*********************************************************************************************************************************/
62
63	/*
64	* PGMR3PhysReadU8-64
65	* PGMR3PhysWriteU8-64
66	*/
67	#define PGMPHYSFN_READNAME PGMR3PhysReadU8
68	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU8
69	#define PGMPHYS_DATASIZE 1
70	#define PGMPHYS_DATATYPE uint8_t
71	#include "PGMPhysRWTmpl.h"
72
73	#define PGMPHYSFN_READNAME PGMR3PhysReadU16
74	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU16
75	#define PGMPHYS_DATASIZE 2
76	#define PGMPHYS_DATATYPE uint16_t
77	#include "PGMPhysRWTmpl.h"
78
79	#define PGMPHYSFN_READNAME PGMR3PhysReadU32
80	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU32
81	#define PGMPHYS_DATASIZE 4
82	#define PGMPHYS_DATATYPE uint32_t
83	#include "PGMPhysRWTmpl.h"
84
85	#define PGMPHYSFN_READNAME PGMR3PhysReadU64
86	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU64
87	#define PGMPHYS_DATASIZE 8
88	#define PGMPHYS_DATATYPE uint64_t
89	#include "PGMPhysRWTmpl.h"
90
91
92	/**
93	* EMT worker for PGMR3PhysReadExternal.
94	*/
95	static DECLCALLBACK(int) pgmR3PhysReadExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, void *pvBuf, size_t cbRead,
96	PGMACCESSORIGIN enmOrigin)
97	{
98	VBOXSTRICTRC rcStrict = PGMPhysRead(pVM, *pGCPhys, pvBuf, cbRead, enmOrigin);
99	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
100	return VINF_SUCCESS;
101	}
102
103
104	/**
105	* Read from physical memory, external users.
106	*
107	* @returns VBox status code.
108	* @retval VINF_SUCCESS.
109	*
110	* @param pVM The cross context VM structure.
111	* @param GCPhys Physical address to read from.
112	* @param pvBuf Where to read into.
113	* @param cbRead How many bytes to read.
114	* @param enmOrigin Who is calling.
115	*
116	* @thread Any but EMTs.
117	*/
118	VMMR3DECL(int) PGMR3PhysReadExternal(PVM pVM, RTGCPHYS GCPhys, void *pvBuf, size_t cbRead, PGMACCESSORIGIN enmOrigin)
119	{
120	VM_ASSERT_OTHER_THREAD(pVM);
121
122	AssertMsgReturn(cbRead > 0, ("don't even think about reading zero bytes!\n"), VINF_SUCCESS);
123	LogFlow(("PGMR3PhysReadExternal: %RGp %d\n", GCPhys, cbRead));
124
125	PGM_LOCK_VOID(pVM);
126
127	/*
128	* Copy loop on ram ranges.
129	*/
130	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
131	for (;;)
132	{
133	/* Inside range or not? */
134	if (pRam && GCPhys >= pRam->GCPhys)
135	{
136	/*
137	* Must work our way thru this page by page.
138	*/
139	RTGCPHYS off = GCPhys - pRam->GCPhys;
140	while (off < pRam->cb)
141	{
142	unsigned iPage = off >> PAGE_SHIFT;
143	PPGMPAGE pPage = &pRam->aPages[iPage];
144
145	/*
146	* If the page has an ALL access handler, we'll have to
147	* delegate the job to EMT.
148	*/
149	if ( PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
150	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
151	{
152	PGM_UNLOCK(pVM);
153
154	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysReadExternalEMT, 5,
155	pVM, &GCPhys, pvBuf, cbRead, enmOrigin);
156	}
157	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
158
159	/*
160	* Simple stuff, go ahead.
161	*/
162	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
163	if (cb > cbRead)
164	cb = cbRead;
165	PGMPAGEMAPLOCK PgMpLck;
166	const void *pvSrc;
167	int rc = pgmPhysGCPhys2CCPtrInternalReadOnly(pVM, pPage, pRam->GCPhys + off, &pvSrc, &PgMpLck);
168	if (RT_SUCCESS(rc))
169	{
170	memcpy(pvBuf, pvSrc, cb);
171	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
172	}
173	else
174	{
175	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternalReadOnly failed on %RGp / %R[pgmpage] -> %Rrc\n",
176	pRam->GCPhys + off, pPage, rc));
177	memset(pvBuf, 0xff, cb);
178	}
179
180	/* next page */
181	if (cb >= cbRead)
182	{
183	PGM_UNLOCK(pVM);
184	return VINF_SUCCESS;
185	}
186	cbRead -= cb;
187	off += cb;
188	GCPhys += cb;
189	pvBuf = (char *)pvBuf + cb;
190	} /* walk pages in ram range. */
191	}
192	else
193	{
194	LogFlow(("PGMPhysRead: Unassigned %RGp size=%u\n", GCPhys, cbRead));
195
196	/*
197	* Unassigned address space.
198	*/
199	size_t cb = pRam ? pRam->GCPhys - GCPhys : ~(size_t)0;
200	if (cb >= cbRead)
201	{
202	memset(pvBuf, 0xff, cbRead);
203	break;
204	}
205	memset(pvBuf, 0xff, cb);
206
207	cbRead -= cb;
208	pvBuf = (char *)pvBuf + cb;
209	GCPhys += cb;
210	}
211
212	/* Advance range if necessary. */
213	while (pRam && GCPhys > pRam->GCPhysLast)
214	pRam = pRam->CTX_SUFF(pNext);
215	} /* Ram range walk */
216
217	PGM_UNLOCK(pVM);
218
219	return VINF_SUCCESS;
220	}
221
222
223	/**
224	* EMT worker for PGMR3PhysWriteExternal.
225	*/
226	static DECLCALLBACK(int) pgmR3PhysWriteExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, const void *pvBuf, size_t cbWrite,
227	PGMACCESSORIGIN enmOrigin)
228	{
229	/** @todo VERR_EM_NO_MEMORY */
230	VBOXSTRICTRC rcStrict = PGMPhysWrite(pVM, *pGCPhys, pvBuf, cbWrite, enmOrigin);
231	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
232	return VINF_SUCCESS;
233	}
234
235
236	/**
237	* Write to physical memory, external users.
238	*
239	* @returns VBox status code.
240	* @retval VINF_SUCCESS.
241	* @retval VERR_EM_NO_MEMORY.
242	*
243	* @param pVM The cross context VM structure.
244	* @param GCPhys Physical address to write to.
245	* @param pvBuf What to write.
246	* @param cbWrite How many bytes to write.
247	* @param enmOrigin Who is calling.
248	*
249	* @thread Any but EMTs.
250	*/
251	VMMDECL(int) PGMR3PhysWriteExternal(PVM pVM, RTGCPHYS GCPhys, const void *pvBuf, size_t cbWrite, PGMACCESSORIGIN enmOrigin)
252	{
253	VM_ASSERT_OTHER_THREAD(pVM);
254
255	AssertMsg(!pVM->pgm.s.fNoMorePhysWrites,
256	("Calling PGMR3PhysWriteExternal after pgmR3Save()! GCPhys=%RGp cbWrite=%#x enmOrigin=%d\n",
257	GCPhys, cbWrite, enmOrigin));
258	AssertMsgReturn(cbWrite > 0, ("don't even think about writing zero bytes!\n"), VINF_SUCCESS);
259	LogFlow(("PGMR3PhysWriteExternal: %RGp %d\n", GCPhys, cbWrite));
260
261	PGM_LOCK_VOID(pVM);
262
263	/*
264	* Copy loop on ram ranges, stop when we hit something difficult.
265	*/
266	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
267	for (;;)
268	{
269	/* Inside range or not? */
270	if (pRam && GCPhys >= pRam->GCPhys)
271	{
272	/*
273	* Must work our way thru this page by page.
274	*/
275	RTGCPTR off = GCPhys - pRam->GCPhys;
276	while (off < pRam->cb)
277	{
278	RTGCPTR iPage = off >> PAGE_SHIFT;
279	PPGMPAGE pPage = &pRam->aPages[iPage];
280
281	/*
282	* Is the page problematic, we have to do the work on the EMT.
283	*
284	* Allocating writable pages and access handlers are
285	* problematic, write monitored pages are simple and can be
286	* dealt with here.
287	*/
288	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
289	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
290	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
291	{
292	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
293	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
294	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, GCPhys);
295	else
296	{
297	PGM_UNLOCK(pVM);
298
299	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysWriteExternalEMT, 5,
300	pVM, &GCPhys, pvBuf, cbWrite, enmOrigin);
301	}
302	}
303	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
304
305	/*
306	* Simple stuff, go ahead.
307	*/
308	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
309	if (cb > cbWrite)
310	cb = cbWrite;
311	PGMPAGEMAPLOCK PgMpLck;
312	void *pvDst;
313	int rc = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, pRam->GCPhys + off, &pvDst, &PgMpLck);
314	if (RT_SUCCESS(rc))
315	{
316	memcpy(pvDst, pvBuf, cb);
317	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
318	}
319	else
320	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternal failed on %RGp / %R[pgmpage] -> %Rrc\n",
321	pRam->GCPhys + off, pPage, rc));
322
323	/* next page */
324	if (cb >= cbWrite)
325	{
326	PGM_UNLOCK(pVM);
327	return VINF_SUCCESS;
328	}
329
330	cbWrite -= cb;
331	off += cb;
332	GCPhys += cb;
333	pvBuf = (const char *)pvBuf + cb;
334	} /* walk pages in ram range */
335	}
336	else
337	{
338	/*
339	* Unassigned address space, skip it.
340	*/
341	if (!pRam)
342	break;
343	size_t cb = pRam->GCPhys - GCPhys;
344	if (cb >= cbWrite)
345	break;
346	cbWrite -= cb;
347	pvBuf = (const char *)pvBuf + cb;
348	GCPhys += cb;
349	}
350
351	/* Advance range if necessary. */
352	while (pRam && GCPhys > pRam->GCPhysLast)
353	pRam = pRam->CTX_SUFF(pNext);
354	} /* Ram range walk */
355
356	PGM_UNLOCK(pVM);
357	return VINF_SUCCESS;
358	}
359
360
361	/*********************************************************************************************************************************
362	* Mapping Guest Physical Memory *
363	*********************************************************************************************************************************/
364
365	/**
366	* VMR3ReqCall worker for PGMR3PhysGCPhys2CCPtrExternal to make pages writable.
367	*
368	* @returns see PGMR3PhysGCPhys2CCPtrExternal
369	* @param pVM The cross context VM structure.
370	* @param pGCPhys Pointer to the guest physical address.
371	* @param ppv Where to store the mapping address.
372	* @param pLock Where to store the lock.
373	*/
374	static DECLCALLBACK(int) pgmR3PhysGCPhys2CCPtrDelegated(PVM pVM, PRTGCPHYS pGCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
375	{
376	/*
377	* Just hand it to PGMPhysGCPhys2CCPtr and check that it's not a page with
378	* an access handler after it succeeds.
379	*/
380	int rc = PGM_LOCK(pVM);
381	AssertRCReturn(rc, rc);
382
383	rc = PGMPhysGCPhys2CCPtr(pVM, *pGCPhys, ppv, pLock);
384	if (RT_SUCCESS(rc))
385	{
386	PPGMPAGEMAPTLBE pTlbe;
387	int rc2 = pgmPhysPageQueryTlbe(pVM, *pGCPhys, &pTlbe);
388	AssertFatalRC(rc2);
389	PPGMPAGE pPage = pTlbe->pPage;
390	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
391	{
392	PGMPhysReleasePageMappingLock(pVM, pLock);
393	rc = VERR_PGM_PHYS_PAGE_RESERVED;
394	}
395	else if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
396	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
397	\|\| pgmPoolIsDirtyPage(pVM, *pGCPhys)
398	#endif
399	)
400	{
401	/* We must flush any corresponding pgm pool page here, otherwise we'll
402	* not be informed about writes and keep bogus gst->shw mappings around.
403	*/
404	pgmPoolFlushPageByGCPhys(pVM, *pGCPhys);
405	Assert(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage));
406	/** @todo r=bird: return VERR_PGM_PHYS_PAGE_RESERVED here if it still has
407	* active handlers, see the PGMR3PhysGCPhys2CCPtrExternal docs. */
408	}
409	}
410
411	PGM_UNLOCK(pVM);
412	return rc;
413	}
414
415
416	/**
417	* Requests the mapping of a guest page into ring-3, external threads.
418	*
419	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
420	* release it.
421	*
422	* This API will assume your intention is to write to the page, and will
423	* therefore replace shared and zero pages. If you do not intend to modify the
424	* page, use the PGMR3PhysGCPhys2CCPtrReadOnlyExternal() API.
425	*
426	* @returns VBox status code.
427	* @retval VINF_SUCCESS on success.
428	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
429	* backing or if the page has any active access handlers. The caller
430	* must fall back on using PGMR3PhysWriteExternal.
431	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
432	*
433	* @param pVM The cross context VM structure.
434	* @param GCPhys The guest physical address of the page that should be mapped.
435	* @param ppv Where to store the address corresponding to GCPhys.
436	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
437	*
438	* @remark Avoid calling this API from within critical sections (other than the
439	* PGM one) because of the deadlock risk when we have to delegating the
440	* task to an EMT.
441	* @thread Any.
442	*/
443	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrExternal(PVM pVM, RTGCPHYS GCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
444	{
445	AssertPtr(ppv);
446	AssertPtr(pLock);
447
448	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
449
450	int rc = PGM_LOCK(pVM);
451	AssertRCReturn(rc, rc);
452
453	/*
454	* Query the Physical TLB entry for the page (may fail).
455	*/
456	PPGMPAGEMAPTLBE pTlbe;
457	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
458	if (RT_SUCCESS(rc))
459	{
460	PPGMPAGE pPage = pTlbe->pPage;
461	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
462	rc = VERR_PGM_PHYS_PAGE_RESERVED;
463	else
464	{
465	/*
466	* If the page is shared, the zero page, or being write monitored
467	* it must be converted to an page that's writable if possible.
468	* We can only deal with write monitored pages here, the rest have
469	* to be on an EMT.
470	*/
471	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
472	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
473	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
474	\|\| pgmPoolIsDirtyPage(pVM, GCPhys)
475	#endif
476	)
477	{
478	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
479	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
480	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
481	&& !pgmPoolIsDirtyPage(pVM, GCPhys) /** @todo we're very likely doing this twice. */
482	#endif
483	)
484	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, GCPhys);
485	else
486	{
487	PGM_UNLOCK(pVM);
488
489	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
490	pVM, &GCPhys, ppv, pLock);
491	}
492	}
493
494	/*
495	* Now, just perform the locking and calculate the return address.
496	*/
497	PPGMPAGEMAP pMap = pTlbe->pMap;
498	if (pMap)
499	pMap->cRefs++;
500
501	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
502	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
503	{
504	if (cLocks == 0)
505	pVM->pgm.s.cWriteLockedPages++;
506	PGM_PAGE_INC_WRITE_LOCKS(pPage);
507	}
508	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
509	{
510	PGM_PAGE_INC_WRITE_LOCKS(pPage);
511	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", GCPhys, pPage));
512	if (pMap)
513	pMap->cRefs++; /* Extra ref to prevent it from going away. */
514	}
515
516	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
517	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
518	pLock->pvMap = pMap;
519	}
520	}
521
522	PGM_UNLOCK(pVM);
523	return rc;
524	}
525
526
527	/**
528	* Requests the mapping of a guest page into ring-3, external threads.
529	*
530	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
531	* release it.
532	*
533	* @returns VBox status code.
534	* @retval VINF_SUCCESS on success.
535	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
536	* backing or if the page as an active ALL access handler. The caller
537	* must fall back on using PGMPhysRead.
538	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
539	*
540	* @param pVM The cross context VM structure.
541	* @param GCPhys The guest physical address of the page that should be mapped.
542	* @param ppv Where to store the address corresponding to GCPhys.
543	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
544	*
545	* @remark Avoid calling this API from within critical sections (other than
546	* the PGM one) because of the deadlock risk.
547	* @thread Any.
548	*/
549	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrReadOnlyExternal(PVM pVM, RTGCPHYS GCPhys, void const **ppv, PPGMPAGEMAPLOCK pLock)
550	{
551	int rc = PGM_LOCK(pVM);
552	AssertRCReturn(rc, rc);
553
554	/*
555	* Query the Physical TLB entry for the page (may fail).
556	*/
557	PPGMPAGEMAPTLBE pTlbe;
558	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
559	if (RT_SUCCESS(rc))
560	{
561	PPGMPAGE pPage = pTlbe->pPage;
562	#if 1
563	/* MMIO pages doesn't have any readable backing. */
564	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
565	rc = VERR_PGM_PHYS_PAGE_RESERVED;
566	#else
567	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
568	rc = VERR_PGM_PHYS_PAGE_RESERVED;
569	#endif
570	else
571	{
572	/*
573	* Now, just perform the locking and calculate the return address.
574	*/
575	PPGMPAGEMAP pMap = pTlbe->pMap;
576	if (pMap)
577	pMap->cRefs++;
578
579	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
580	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
581	{
582	if (cLocks == 0)
583	pVM->pgm.s.cReadLockedPages++;
584	PGM_PAGE_INC_READ_LOCKS(pPage);
585	}
586	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
587	{
588	PGM_PAGE_INC_READ_LOCKS(pPage);
589	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", GCPhys, pPage));
590	if (pMap)
591	pMap->cRefs++; /* Extra ref to prevent it from going away. */
592	}
593
594	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
595	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
596	pLock->pvMap = pMap;
597	}
598	}
599
600	PGM_UNLOCK(pVM);
601	return rc;
602	}
603
604
605	/**
606	* Requests the mapping of multiple guest page into ring-3, external threads.
607	*
608	* When you're done with the pages, call PGMPhysBulkReleasePageMappingLock()
609	* ASAP to release them.
610	*
611	* This API will assume your intention is to write to the pages, and will
612	* therefore replace shared and zero pages. If you do not intend to modify the
613	* pages, use the PGMR3PhysBulkGCPhys2CCPtrReadOnlyExternal() API.
614	*
615	* @returns VBox status code.
616	* @retval VINF_SUCCESS on success.
617	* @retval VERR_PGM_PHYS_PAGE_RESERVED if any of the pages has no physical
618	* backing or if any of the pages the page has any active access
619	* handlers. The caller must fall back on using PGMR3PhysWriteExternal.
620	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if @a paGCPhysPages contains
621	* an invalid physical address.
622	*
623	* @param pVM The cross context VM structure.
624	* @param cPages Number of pages to lock.
625	* @param paGCPhysPages The guest physical address of the pages that
626	* should be mapped (@a cPages entries).
627	* @param papvPages Where to store the ring-3 mapping addresses
628	* corresponding to @a paGCPhysPages.
629	* @param paLocks Where to store the locking information that
630	* pfnPhysBulkReleasePageMappingLock needs (@a cPages
631	* in length).
632	*
633	* @remark Avoid calling this API from within critical sections (other than the
634	* PGM one) because of the deadlock risk when we have to delegating the
635	* task to an EMT.
636	* @thread Any.
637	*/
638	VMMR3DECL(int) PGMR3PhysBulkGCPhys2CCPtrExternal(PVM pVM, uint32_t cPages, PCRTGCPHYS paGCPhysPages,
639	void **papvPages, PPGMPAGEMAPLOCK paLocks)
640	{
641	Assert(cPages > 0);
642	AssertPtr(papvPages);
643	AssertPtr(paLocks);
644
645	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
646
647	int rc = PGM_LOCK(pVM);
648	AssertRCReturn(rc, rc);
649
650	/*
651	* Lock the pages one by one.
652	* The loop body is similar to PGMR3PhysGCPhys2CCPtrExternal.
653	*/
654	int32_t cNextYield = 128;
655	uint32_t iPage;
656	for (iPage = 0; iPage < cPages; iPage++)
657	{
658	if (--cNextYield > 0)
659	{ /* likely */ }
660	else
661	{
662	PGM_UNLOCK(pVM);
663	ASMNopPause();
664	PGM_LOCK_VOID(pVM);
665	cNextYield = 128;
666	}
667
668	/*
669	* Query the Physical TLB entry for the page (may fail).
670	*/
671	PPGMPAGEMAPTLBE pTlbe;
672	rc = pgmPhysPageQueryTlbe(pVM, paGCPhysPages[iPage], &pTlbe);
673	if (RT_SUCCESS(rc))
674	{ }
675	else
676	break;
677	PPGMPAGE pPage = pTlbe->pPage;
678
679	/*
680	* No MMIO or active access handlers.
681	*/
682	if ( !PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage)
683	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
684	{ }
685	else
686	{
687	rc = VERR_PGM_PHYS_PAGE_RESERVED;
688	break;
689	}
690
691	/*
692	* The page must be in the allocated state and not be a dirty pool page.
693	* We can handle converting a write monitored page to an allocated one, but
694	* anything more complicated must be delegated to an EMT.
695	*/
696	bool fDelegateToEmt = false;
697	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_ALLOCATED)
698	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
699	fDelegateToEmt = pgmPoolIsDirtyPage(pVM, paGCPhysPages[iPage]);
700	#else
701	fDelegateToEmt = false;
702	#endif
703	else if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
704	{
705	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
706	if (!pgmPoolIsDirtyPage(pVM, paGCPhysPages[iPage]))
707	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, paGCPhysPages[iPage]);
708	else
709	fDelegateToEmt = true;
710	#endif
711	}
712	else
713	fDelegateToEmt = true;
714	if (!fDelegateToEmt)
715	{ }
716	else
717	{
718	/* We could do this delegation in bulk, but considered too much work vs gain. */
719	PGM_UNLOCK(pVM);
720	rc = VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
721	pVM, &paGCPhysPages[iPage], &papvPages[iPage], &paLocks[iPage]);
722	PGM_LOCK_VOID(pVM);
723	if (RT_FAILURE(rc))
724	break;
725	cNextYield = 128;
726	}
727
728	/*
729	* Now, just perform the locking and address calculation.
730	*/
731	PPGMPAGEMAP pMap = pTlbe->pMap;
732	if (pMap)
733	pMap->cRefs++;
734
735	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
736	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
737	{
738	if (cLocks == 0)
739	pVM->pgm.s.cWriteLockedPages++;
740	PGM_PAGE_INC_WRITE_LOCKS(pPage);
741	}
742	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
743	{
744	PGM_PAGE_INC_WRITE_LOCKS(pPage);
745	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", paGCPhysPages[iPage], pPage));
746	if (pMap)
747	pMap->cRefs++; /* Extra ref to prevent it from going away. */
748	}
749
750	papvPages[iPage] = (void *)((uintptr_t)pTlbe->pv \| (uintptr_t)(paGCPhysPages[iPage] & PAGE_OFFSET_MASK));
751	paLocks[iPage].uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
752	paLocks[iPage].pvMap = pMap;
753	}
754
755	PGM_UNLOCK(pVM);
756
757	/*
758	* On failure we must unlock any pages we managed to get already.
759	*/
760	if (RT_FAILURE(rc) && iPage > 0)
761	PGMPhysBulkReleasePageMappingLocks(pVM, iPage, paLocks);
762
763	return rc;
764	}
765
766
767	/**
768	* Requests the mapping of multiple guest page into ring-3, for reading only,
769	* external threads.
770	*
771	* When you're done with the pages, call PGMPhysReleasePageMappingLock() ASAP
772	* to release them.
773	*
774	* @returns VBox status code.
775	* @retval VINF_SUCCESS on success.
776	* @retval VERR_PGM_PHYS_PAGE_RESERVED if any of the pages has no physical
777	* backing or if any of the pages the page has an active ALL access
778	* handler. The caller must fall back on using PGMR3PhysWriteExternal.
779	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if @a paGCPhysPages contains
780	* an invalid physical address.
781	*
782	* @param pVM The cross context VM structure.
783	* @param cPages Number of pages to lock.
784	* @param paGCPhysPages The guest physical address of the pages that
785	* should be mapped (@a cPages entries).
786	* @param papvPages Where to store the ring-3 mapping addresses
787	* corresponding to @a paGCPhysPages.
788	* @param paLocks Where to store the lock information that
789	* pfnPhysReleasePageMappingLock needs (@a cPages
790	* in length).
791	*
792	* @remark Avoid calling this API from within critical sections (other than
793	* the PGM one) because of the deadlock risk.
794	* @thread Any.
795	*/
796	VMMR3DECL(int) PGMR3PhysBulkGCPhys2CCPtrReadOnlyExternal(PVM pVM, uint32_t cPages, PCRTGCPHYS paGCPhysPages,
797	void const **papvPages, PPGMPAGEMAPLOCK paLocks)
798	{
799	Assert(cPages > 0);
800	AssertPtr(papvPages);
801	AssertPtr(paLocks);
802
803	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
804
805	int rc = PGM_LOCK(pVM);
806	AssertRCReturn(rc, rc);
807
808	/*
809	* Lock the pages one by one.
810	* The loop body is similar to PGMR3PhysGCPhys2CCPtrReadOnlyExternal.
811	*/
812	int32_t cNextYield = 256;
813	uint32_t iPage;
814	for (iPage = 0; iPage < cPages; iPage++)
815	{
816	if (--cNextYield > 0)
817	{ /* likely */ }
818	else
819	{
820	PGM_UNLOCK(pVM);
821	ASMNopPause();
822	PGM_LOCK_VOID(pVM);
823	cNextYield = 256;
824	}
825
826	/*
827	* Query the Physical TLB entry for the page (may fail).
828	*/
829	PPGMPAGEMAPTLBE pTlbe;
830	rc = pgmPhysPageQueryTlbe(pVM, paGCPhysPages[iPage], &pTlbe);
831	if (RT_SUCCESS(rc))
832	{ }
833	else
834	break;
835	PPGMPAGE pPage = pTlbe->pPage;
836
837	/*
838	* No MMIO or active all access handlers, everything else can be accessed.
839	*/
840	if ( !PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage)
841	&& !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
842	{ }
843	else
844	{
845	rc = VERR_PGM_PHYS_PAGE_RESERVED;
846	break;
847	}
848
849	/*
850	* Now, just perform the locking and address calculation.
851	*/
852	PPGMPAGEMAP pMap = pTlbe->pMap;
853	if (pMap)
854	pMap->cRefs++;
855
856	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
857	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
858	{
859	if (cLocks == 0)
860	pVM->pgm.s.cReadLockedPages++;
861	PGM_PAGE_INC_READ_LOCKS(pPage);
862	}
863	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
864	{
865	PGM_PAGE_INC_READ_LOCKS(pPage);
866	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", paGCPhysPages[iPage], pPage));
867	if (pMap)
868	pMap->cRefs++; /* Extra ref to prevent it from going away. */
869	}
870
871	papvPages[iPage] = (void *)((uintptr_t)pTlbe->pv \| (uintptr_t)(paGCPhysPages[iPage] & PAGE_OFFSET_MASK));
872	paLocks[iPage].uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
873	paLocks[iPage].pvMap = pMap;
874	}
875
876	PGM_UNLOCK(pVM);
877
878	/*
879	* On failure we must unlock any pages we managed to get already.
880	*/
881	if (RT_FAILURE(rc) && iPage > 0)
882	PGMPhysBulkReleasePageMappingLocks(pVM, iPage, paLocks);
883
884	return rc;
885	}
886
887
888	/**
889	* Converts a GC physical address to a HC ring-3 pointer, with some
890	* additional checks.
891	*
892	* @returns VBox status code.
893	* @retval VINF_SUCCESS on success.
894	* @retval VINF_PGM_PHYS_TLB_CATCH_WRITE and *ppv set if the page has a write
895	* access handler of some kind.
896	* @retval VERR_PGM_PHYS_TLB_CATCH_ALL if the page has a handler catching all
897	* accesses or is odd in any way.
898	* @retval VERR_PGM_PHYS_TLB_UNASSIGNED if the page doesn't exist.
899	*
900	* @param pVM The cross context VM structure.
901	* @param GCPhys The GC physical address to convert. Since this is only
902	* used for filling the REM TLB, the A20 mask must be
903	* applied before calling this API.
904	* @param fWritable Whether write access is required.
905	* @param ppv Where to store the pointer corresponding to GCPhys on
906	* success.
907	*/
908	VMMR3DECL(int) PGMR3PhysTlbGCPhys2Ptr(PVM pVM, RTGCPHYS GCPhys, bool fWritable, void **ppv)
909	{
910	PGM_LOCK_VOID(pVM);
911	PGM_A20_ASSERT_MASKED(VMMGetCpu(pVM), GCPhys);
912
913	PPGMRAMRANGE pRam;
914	PPGMPAGE pPage;
915	int rc = pgmPhysGetPageAndRangeEx(pVM, GCPhys, &pPage, &pRam);
916	if (RT_SUCCESS(rc))
917	{
918	if (PGM_PAGE_IS_BALLOONED(pPage))
919	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
920	else if (!PGM_PAGE_HAS_ANY_HANDLERS(pPage))
921	rc = VINF_SUCCESS;
922	else
923	{
924	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)) /* catches MMIO */
925	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
926	else if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
927	{
928	/** @todo Handle TLB loads of virtual handlers so ./test.sh can be made to work
929	* in -norawr0 mode. */
930	if (fWritable)
931	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
932	}
933	else
934	{
935	/* Temporarily disabled physical handler(s), since the recompiler
936	doesn't get notified when it's reset we'll have to pretend it's
937	operating normally. */
938	if (pgmHandlerPhysicalIsAll(pVM, GCPhys))
939	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
940	else
941	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
942	}
943	}
944	if (RT_SUCCESS(rc))
945	{
946	int rc2;
947
948	/* Make sure what we return is writable. */
949	if (fWritable)
950	switch (PGM_PAGE_GET_STATE(pPage))
951	{
952	case PGM_PAGE_STATE_ALLOCATED:
953	break;
954	case PGM_PAGE_STATE_BALLOONED:
955	AssertFailed();
956	break;
957	case PGM_PAGE_STATE_ZERO:
958	case PGM_PAGE_STATE_SHARED:
959	if (rc == VINF_PGM_PHYS_TLB_CATCH_WRITE)
960	break;
961	RT_FALL_THRU();
962	case PGM_PAGE_STATE_WRITE_MONITORED:
963	rc2 = pgmPhysPageMakeWritable(pVM, pPage, GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK);
964	AssertLogRelRCReturn(rc2, rc2);
965	break;
966	}
967
968	/* Get a ring-3 mapping of the address. */
969	PPGMPAGER3MAPTLBE pTlbe;
970	rc2 = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
971	AssertLogRelRCReturn(rc2, rc2);
972	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
973	/** @todo mapping/locking hell; this isn't horribly efficient since
974	* pgmPhysPageLoadIntoTlb will repeat the lookup we've done here. */
975
976	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage] ppv=%p\n", GCPhys, rc, pPage, ppv));
977	}
978	else
979	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage]\n", GCPhys, rc, pPage));
980
981	/* else: handler catching all access, no pointer returned. */
982	}
983	else
984	rc = VERR_PGM_PHYS_TLB_UNASSIGNED;
985
986	PGM_UNLOCK(pVM);
987	return rc;
988	}
989
990
991
992	/*********************************************************************************************************************************
993	* RAM Range Management *
994	*********************************************************************************************************************************/
995
996	#define MAKE_LEAF(a_pNode) \
997	do { \
998	(a_pNode)->pLeftR3 = NIL_RTR3PTR; \
999	(a_pNode)->pRightR3 = NIL_RTR3PTR; \
1000	(a_pNode)->pLeftR0 = NIL_RTR0PTR; \
1001	(a_pNode)->pRightR0 = NIL_RTR0PTR; \
1002	} while (0)
1003
1004	#define INSERT_LEFT(a_pParent, a_pNode) \
1005	do { \
1006	(a_pParent)->pLeftR3 = (a_pNode); \
1007	(a_pParent)->pLeftR0 = (a_pNode)->pSelfR0; \
1008	} while (0)
1009	#define INSERT_RIGHT(a_pParent, a_pNode) \
1010	do { \
1011	(a_pParent)->pRightR3 = (a_pNode); \
1012	(a_pParent)->pRightR0 = (a_pNode)->pSelfR0; \
1013	} while (0)
1014
1015
1016	/**
1017	* Recursive tree builder.
1018	*
1019	* @param ppRam Pointer to the iterator variable.
1020	* @param iDepth The current depth. Inserts a leaf node if 0.
1021	*/
1022	static PPGMRAMRANGE pgmR3PhysRebuildRamRangeSearchTreesRecursively(PPGMRAMRANGE *ppRam, int iDepth)
1023	{
1024	PPGMRAMRANGE pRam;
1025	if (iDepth <= 0)
1026	{
1027	/*
1028	* Leaf node.
1029	*/
1030	pRam = *ppRam;
1031	if (pRam)
1032	{
1033	*ppRam = pRam->pNextR3;
1034	MAKE_LEAF(pRam);
1035	}
1036	}
1037	else
1038	{
1039
1040	/*
1041	* Intermediate node.
1042	*/
1043	PPGMRAMRANGE pLeft = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
1044
1045	pRam = *ppRam;
1046	if (!pRam)
1047	return pLeft;
1048	*ppRam = pRam->pNextR3;
1049	MAKE_LEAF(pRam);
1050	INSERT_LEFT(pRam, pLeft);
1051
1052	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
1053	if (pRight)
1054	INSERT_RIGHT(pRam, pRight);
1055	}
1056	return pRam;
1057	}
1058
1059
1060	/**
1061	* Rebuilds the RAM range search trees.
1062	*
1063	* @param pVM The cross context VM structure.
1064	*/
1065	static void pgmR3PhysRebuildRamRangeSearchTrees(PVM pVM)
1066	{
1067
1068	/*
1069	* Create the reasonably balanced tree in a sequential fashion.
1070	* For simplicity (laziness) we use standard recursion here.
1071	*/
1072	int iDepth = 0;
1073	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
1074	PPGMRAMRANGE pRoot = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, 0);
1075	while (pRam)
1076	{
1077	PPGMRAMRANGE pLeft = pRoot;
1078
1079	pRoot = pRam;
1080	pRam = pRam->pNextR3;
1081	MAKE_LEAF(pRoot);
1082	INSERT_LEFT(pRoot, pLeft);
1083
1084	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, iDepth);
1085	if (pRight)
1086	INSERT_RIGHT(pRoot, pRight);
1087	/** @todo else: rotate the tree. */
1088
1089	iDepth++;
1090	}
1091
1092	pVM->pgm.s.pRamRangeTreeR3 = pRoot;
1093	pVM->pgm.s.pRamRangeTreeR0 = pRoot ? pRoot->pSelfR0 : NIL_RTR0PTR;
1094
1095	#ifdef VBOX_STRICT
1096	/*
1097	* Verify that the above code works.
1098	*/
1099	unsigned cRanges = 0;
1100	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1101	cRanges++;
1102	Assert(cRanges > 0);
1103
1104	unsigned cMaxDepth = ASMBitLastSetU32(cRanges);
1105	if ((1U << cMaxDepth) < cRanges)
1106	cMaxDepth++;
1107
1108	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1109	{
1110	unsigned cDepth = 0;
1111	PPGMRAMRANGE pRam2 = pVM->pgm.s.pRamRangeTreeR3;
1112	for (;;)
1113	{
1114	if (pRam == pRam2)
1115	break;
1116	Assert(pRam2);
1117	if (pRam->GCPhys < pRam2->GCPhys)
1118	pRam2 = pRam2->pLeftR3;
1119	else
1120	pRam2 = pRam2->pRightR3;
1121	}
1122	AssertMsg(cDepth <= cMaxDepth, ("cDepth=%d cMaxDepth=%d\n", cDepth, cMaxDepth));
1123	}
1124	#endif /* VBOX_STRICT */
1125	}
1126
1127	#undef MAKE_LEAF
1128	#undef INSERT_LEFT
1129	#undef INSERT_RIGHT
1130
1131	/**
1132	* Relinks the RAM ranges using the pSelfRC and pSelfR0 pointers.
1133	*
1134	* Called when anything was relocated.
1135	*
1136	* @param pVM The cross context VM structure.
1137	*/
1138	void pgmR3PhysRelinkRamRanges(PVM pVM)
1139	{
1140	PPGMRAMRANGE pCur;
1141
1142	#ifdef VBOX_STRICT
1143	for (pCur = pVM->pgm.s.pRamRangesXR3; pCur; pCur = pCur->pNextR3)
1144	{
1145	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfR0 == MMHyperCCToR0(pVM, pCur));
1146	Assert((pCur->GCPhys & PAGE_OFFSET_MASK) == 0);
1147	Assert((pCur->GCPhysLast & PAGE_OFFSET_MASK) == PAGE_OFFSET_MASK);
1148	Assert((pCur->cb & PAGE_OFFSET_MASK) == 0);
1149	Assert(pCur->cb == pCur->GCPhysLast - pCur->GCPhys + 1);
1150	for (PPGMRAMRANGE pCur2 = pVM->pgm.s.pRamRangesXR3; pCur2; pCur2 = pCur2->pNextR3)
1151	Assert( pCur2 == pCur
1152	\|\| strcmp(pCur2->pszDesc, pCur->pszDesc)); /** @todo fix MMIO ranges!! */
1153	}
1154	#endif
1155
1156	pCur = pVM->pgm.s.pRamRangesXR3;
1157	if (pCur)
1158	{
1159	pVM->pgm.s.pRamRangesXR0 = pCur->pSelfR0;
1160
1161	for (; pCur->pNextR3; pCur = pCur->pNextR3)
1162	pCur->pNextR0 = pCur->pNextR3->pSelfR0;
1163
1164	Assert(pCur->pNextR0 == NIL_RTR0PTR);
1165	}
1166	else
1167	{
1168	Assert(pVM->pgm.s.pRamRangesXR0 == NIL_RTR0PTR);
1169	}
1170	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1171
1172	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1173	}
1174
1175
1176	/**
1177	* Links a new RAM range into the list.
1178	*
1179	* @param pVM The cross context VM structure.
1180	* @param pNew Pointer to the new list entry.
1181	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
1182	*/
1183	static void pgmR3PhysLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, PPGMRAMRANGE pPrev)
1184	{
1185	AssertMsg(pNew->pszDesc, ("%RGp-%RGp\n", pNew->GCPhys, pNew->GCPhysLast));
1186	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfR0 == MMHyperCCToR0(pVM, pNew));
1187
1188	PGM_LOCK_VOID(pVM);
1189
1190	PPGMRAMRANGE pRam = pPrev ? pPrev->pNextR3 : pVM->pgm.s.pRamRangesXR3;
1191	pNew->pNextR3 = pRam;
1192	pNew->pNextR0 = pRam ? pRam->pSelfR0 : NIL_RTR0PTR;
1193
1194	if (pPrev)
1195	{
1196	pPrev->pNextR3 = pNew;
1197	pPrev->pNextR0 = pNew->pSelfR0;
1198	}
1199	else
1200	{
1201	pVM->pgm.s.pRamRangesXR3 = pNew;
1202	pVM->pgm.s.pRamRangesXR0 = pNew->pSelfR0;
1203	}
1204	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1205
1206	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1207	PGM_UNLOCK(pVM);
1208	}
1209
1210
1211	/**
1212	* Unlink an existing RAM range from the list.
1213	*
1214	* @param pVM The cross context VM structure.
1215	* @param pRam Pointer to the new list entry.
1216	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
1217	*/
1218	static void pgmR3PhysUnlinkRamRange2(PVM pVM, PPGMRAMRANGE pRam, PPGMRAMRANGE pPrev)
1219	{
1220	Assert(pPrev ? pPrev->pNextR3 == pRam : pVM->pgm.s.pRamRangesXR3 == pRam);
1221	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfR0 == MMHyperCCToR0(pVM, pRam));
1222
1223	PGM_LOCK_VOID(pVM);
1224
1225	PPGMRAMRANGE pNext = pRam->pNextR3;
1226	if (pPrev)
1227	{
1228	pPrev->pNextR3 = pNext;
1229	pPrev->pNextR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
1230	}
1231	else
1232	{
1233	Assert(pVM->pgm.s.pRamRangesXR3 == pRam);
1234	pVM->pgm.s.pRamRangesXR3 = pNext;
1235	pVM->pgm.s.pRamRangesXR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
1236	}
1237	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1238
1239	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1240	PGM_UNLOCK(pVM);
1241	}
1242
1243
1244	/**
1245	* Unlink an existing RAM range from the list.
1246	*
1247	* @param pVM The cross context VM structure.
1248	* @param pRam Pointer to the new list entry.
1249	*/
1250	static void pgmR3PhysUnlinkRamRange(PVM pVM, PPGMRAMRANGE pRam)
1251	{
1252	PGM_LOCK_VOID(pVM);
1253
1254	/* find prev. */
1255	PPGMRAMRANGE pPrev = NULL;
1256	PPGMRAMRANGE pCur = pVM->pgm.s.pRamRangesXR3;
1257	while (pCur != pRam)
1258	{
1259	pPrev = pCur;
1260	pCur = pCur->pNextR3;
1261	}
1262	AssertFatal(pCur);
1263
1264	pgmR3PhysUnlinkRamRange2(pVM, pRam, pPrev);
1265	PGM_UNLOCK(pVM);
1266	}
1267
1268
1269	/**
1270	* Gets the number of ram ranges.
1271	*
1272	* @returns Number of ram ranges. Returns UINT32_MAX if @a pVM is invalid.
1273	* @param pVM The cross context VM structure.
1274	*/
1275	VMMR3DECL(uint32_t) PGMR3PhysGetRamRangeCount(PVM pVM)
1276	{
1277	VM_ASSERT_VALID_EXT_RETURN(pVM, UINT32_MAX);
1278
1279	PGM_LOCK_VOID(pVM);
1280	uint32_t cRamRanges = 0;
1281	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext))
1282	cRamRanges++;
1283	PGM_UNLOCK(pVM);
1284	return cRamRanges;
1285	}
1286
1287
1288	/**
1289	* Get information about a range.
1290	*
1291	* @returns VINF_SUCCESS or VERR_OUT_OF_RANGE.
1292	* @param pVM The cross context VM structure.
1293	* @param iRange The ordinal of the range.
1294	* @param pGCPhysStart Where to return the start of the range. Optional.
1295	* @param pGCPhysLast Where to return the address of the last byte in the
1296	* range. Optional.
1297	* @param ppszDesc Where to return the range description. Optional.
1298	* @param pfIsMmio Where to indicate that this is a pure MMIO range.
1299	* Optional.
1300	*/
1301	VMMR3DECL(int) PGMR3PhysGetRange(PVM pVM, uint32_t iRange, PRTGCPHYS pGCPhysStart, PRTGCPHYS pGCPhysLast,
1302	const char *ppszDesc, bool pfIsMmio)
1303	{
1304	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1305
1306	PGM_LOCK_VOID(pVM);
1307	uint32_t iCurRange = 0;
1308	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext), iCurRange++)
1309	if (iCurRange == iRange)
1310	{
1311	if (pGCPhysStart)
1312	*pGCPhysStart = pCur->GCPhys;
1313	if (pGCPhysLast)
1314	*pGCPhysLast = pCur->GCPhysLast;
1315	if (ppszDesc)
1316	*ppszDesc = pCur->pszDesc;
1317	if (pfIsMmio)
1318	*pfIsMmio = !!(pCur->fFlags & PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO);
1319
1320	PGM_UNLOCK(pVM);
1321	return VINF_SUCCESS;
1322	}
1323	PGM_UNLOCK(pVM);
1324	return VERR_OUT_OF_RANGE;
1325	}
1326
1327
1328	/*********************************************************************************************************************************
1329	* RAM *
1330	*********************************************************************************************************************************/
1331
1332	/**
1333	* Frees the specified RAM page and replaces it with the ZERO page.
1334	*
1335	* This is used by ballooning, remapping MMIO2, RAM reset and state loading.
1336	*
1337	* @param pVM The cross context VM structure.
1338	* @param pReq Pointer to the request. This is NULL when doing a
1339	* bulk free in NEM memory mode.
1340	* @param pcPendingPages Where the number of pages waiting to be freed are
1341	* kept. This will normally be incremented. This is
1342	* NULL when doing a bulk free in NEM memory mode.
1343	* @param pPage Pointer to the page structure.
1344	* @param GCPhys The guest physical address of the page, if applicable.
1345	* @param enmNewType New page type for NEM notification, since several
1346	* callers will change the type upon successful return.
1347	*
1348	* @remarks The caller must own the PGM lock.
1349	*/
1350	int pgmPhysFreePage(PVM pVM, PGMMFREEPAGESREQ pReq, uint32_t *pcPendingPages, PPGMPAGE pPage, RTGCPHYS GCPhys,
1351	PGMPAGETYPE enmNewType)
1352	{
1353	/*
1354	* Assert sanity.
1355	*/
1356	PGM_LOCK_ASSERT_OWNER(pVM);
1357	if (RT_UNLIKELY( PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM
1358	&& PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_ROM_SHADOW))
1359	{
1360	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
1361	return VMSetError(pVM, VERR_PGM_PHYS_NOT_RAM, RT_SRC_POS, "GCPhys=%RGp type=%d", GCPhys, PGM_PAGE_GET_TYPE(pPage));
1362	}
1363
1364	/** @todo What about ballooning of large pages??! */
1365	Assert( PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE
1366	&& PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE_DISABLED);
1367
1368	if ( PGM_PAGE_IS_ZERO(pPage)
1369	\|\| PGM_PAGE_IS_BALLOONED(pPage))
1370	return VINF_SUCCESS;
1371
1372	const uint32_t idPage = PGM_PAGE_GET_PAGEID(pPage);
1373	Log3(("pgmPhysFreePage: idPage=%#x GCPhys=%RGp pPage=%R[pgmpage]\n", idPage, GCPhys, pPage));
1374	if (RT_UNLIKELY(!PGM_IS_IN_NEM_MODE(pVM)
1375	? idPage == NIL_GMM_PAGEID
1376	\|\| idPage > GMM_PAGEID_LAST
1377	\|\| PGM_PAGE_GET_CHUNKID(pPage) == NIL_GMM_CHUNKID
1378	: idPage != NIL_GMM_PAGEID))
1379	{
1380	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
1381	return VMSetError(pVM, VERR_PGM_PHYS_INVALID_PAGE_ID, RT_SRC_POS, "GCPhys=%RGp idPage=%#x", GCPhys, pPage);
1382	}
1383	#ifdef VBOX_WITH_NATIVE_NEM
1384	const RTHCPHYS HCPhysPrev = PGM_PAGE_GET_HCPHYS(pPage);
1385	#endif
1386
1387	/* update page count stats. */
1388	if (PGM_PAGE_IS_SHARED(pPage))
1389	pVM->pgm.s.cSharedPages--;
1390	else
1391	pVM->pgm.s.cPrivatePages--;
1392	pVM->pgm.s.cZeroPages++;
1393
1394	/* Deal with write monitored pages. */
1395	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
1396	{
1397	PGM_PAGE_SET_WRITTEN_TO(pVM, pPage);
1398	pVM->pgm.s.cWrittenToPages++;
1399	}
1400
1401	/*
1402	* pPage = ZERO page.
1403	*/
1404	PGM_PAGE_SET_HCPHYS(pVM, pPage, pVM->pgm.s.HCPhysZeroPg);
1405	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
1406	PGM_PAGE_SET_PAGEID(pVM, pPage, NIL_GMM_PAGEID);
1407	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
1408	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
1409	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
1410
1411	/* Flush physical page map TLB entry. */
1412	pgmPhysInvalidatePageMapTLBEntry(pVM, GCPhys);
1413
1414	#ifdef VBOX_WITH_PGM_NEM_MODE
1415	/*
1416	* Skip the rest if we're doing a bulk free in NEM memory mode.
1417	*/
1418	if (!pReq)
1419	return VINF_SUCCESS;
1420	AssertLogRelReturn(!pVM->pgm.s.fNemMode, VERR_PGM_NOT_SUPPORTED_FOR_NEM_MODE);
1421	#endif
1422
1423	#ifdef VBOX_WITH_NATIVE_NEM
1424	/* Notify NEM. */
1425	/** @todo Remove this one? */
1426	if (VM_IS_NEM_ENABLED(pVM))
1427	{
1428	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pPage);
1429	NEMHCNotifyPhysPageChanged(pVM, GCPhys, HCPhysPrev, pVM->pgm.s.HCPhysZeroPg, pVM->pgm.s.pvZeroPgR3,
1430	pgmPhysPageCalcNemProtection(pPage, enmNewType), enmNewType, &u2State);
1431	PGM_PAGE_SET_NEM_STATE(pPage, u2State);
1432	}
1433	#else
1434	RT_NOREF(enmNewType);
1435	#endif
1436
1437	/*
1438	* Make sure it's not in the handy page array.
1439	*/
1440	for (uint32_t i = pVM->pgm.s.cHandyPages; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
1441	{
1442	if (pVM->pgm.s.aHandyPages[i].idPage == idPage)
1443	{
1444	pVM->pgm.s.aHandyPages[i].HCPhysGCPhys = NIL_GMMPAGEDESC_PHYS;
1445	pVM->pgm.s.aHandyPages[i].fZeroed = false;
1446	pVM->pgm.s.aHandyPages[i].idPage = NIL_GMM_PAGEID;
1447	break;
1448	}
1449	if (pVM->pgm.s.aHandyPages[i].idSharedPage == idPage)
1450	{
1451	pVM->pgm.s.aHandyPages[i].idSharedPage = NIL_GMM_PAGEID;
1452	break;
1453	}
1454	}
1455
1456	/*
1457	* Push it onto the page array.
1458	*/
1459	uint32_t iPage = *pcPendingPages;
1460	Assert(iPage < PGMPHYS_FREE_PAGE_BATCH_SIZE);
1461	*pcPendingPages += 1;
1462
1463	pReq->aPages[iPage].idPage = idPage;
1464
1465	if (iPage + 1 < PGMPHYS_FREE_PAGE_BATCH_SIZE)
1466	return VINF_SUCCESS;
1467
1468	/*
1469	* Flush the pages.
1470	*/
1471	int rc = GMMR3FreePagesPerform(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE);
1472	if (RT_SUCCESS(rc))
1473	{
1474	GMMR3FreePagesRePrep(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1475	*pcPendingPages = 0;
1476	}
1477	return rc;
1478	}
1479
1480
1481	/**
1482	* Frees a range of pages, replacing them with ZERO pages of the specified type.
1483	*
1484	* @returns VBox status code.
1485	* @param pVM The cross context VM structure.
1486	* @param pRam The RAM range in which the pages resides.
1487	* @param GCPhys The address of the first page.
1488	* @param GCPhysLast The address of the last page.
1489	* @param pvMmio2 Pointer to the ring-3 mapping of any MMIO2 memory that
1490	* will replace the pages we're freeing up.
1491	*/
1492	static int pgmR3PhysFreePageRange(PVM pVM, PPGMRAMRANGE pRam, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast, void *pvMmio2)
1493	{
1494	PGM_LOCK_ASSERT_OWNER(pVM);
1495
1496	#ifdef VBOX_WITH_PGM_NEM_MODE
1497	/*
1498	* In simplified memory mode we don't actually free the memory,
1499	* we just unmap it and let NEM do any unlocking of it.
1500	*/
1501	if (pVM->pgm.s.fNemMode)
1502	{
1503	Assert(VM_IS_NEM_ENABLED(pVM));
1504	uint32_t const fNemNotify = (pvMmio2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0) \| NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE;
1505	uint8_t u2State = 0; /* (We don't support UINT8_MAX here.) */
1506	int rc = NEMR3NotifyPhysMmioExMapEarly(pVM, GCPhys, GCPhysLast - GCPhys + 1, fNemNotify,
1507	pRam->pvR3 ? (uint8_t *)pRam->pvR3 + GCPhys - pRam->GCPhys : NULL,
1508	pvMmio2, &u2State, NULL /puNemRange/);
1509	AssertLogRelRCReturn(rc, rc);
1510
1511	/* Iterate the pages. */
1512	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
1513	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
1514	while (cPagesLeft-- > 0)
1515	{
1516	rc = pgmPhysFreePage(pVM, NULL, NULL, pPageDst, GCPhys, PGMPAGETYPE_MMIO);
1517	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
1518
1519	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO);
1520	PGM_PAGE_SET_NEM_STATE(pPageDst, u2State);
1521
1522	GCPhys += PAGE_SIZE;
1523	pPageDst++;
1524	}
1525	return rc;
1526	}
1527	#else /* !VBOX_WITH_PGM_NEM_MODE */
1528	RT_NOREF(pvMmio2);
1529	#endif /* !VBOX_WITH_PGM_NEM_MODE */
1530
1531	/*
1532	* Regular mode.
1533	*/
1534	/* Prepare. */
1535	uint32_t cPendingPages = 0;
1536	PGMMFREEPAGESREQ pReq;
1537	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1538	AssertLogRelRCReturn(rc, rc);
1539
1540	#ifdef VBOX_WITH_NATIVE_NEM
1541	/* Tell NEM up-front. */
1542	uint8_t u2State = UINT8_MAX;
1543	if (VM_IS_NEM_ENABLED(pVM))
1544	{
1545	uint32_t const fNemNotify = (pvMmio2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0) \| NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE;
1546	rc = NEMR3NotifyPhysMmioExMapEarly(pVM, GCPhys, GCPhysLast - GCPhys + 1, fNemNotify, NULL, pvMmio2,
1547	&u2State, NULL /puNemRange/);
1548	AssertLogRelRCReturnStmt(rc, GMMR3FreePagesCleanup(pReq), rc);
1549	}
1550	#endif
1551
1552	/* Iterate the pages. */
1553	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
1554	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
1555	while (cPagesLeft-- > 0)
1556	{
1557	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys, PGMPAGETYPE_MMIO);
1558	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
1559
1560	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO);
1561	#ifdef VBOX_WITH_NATIVE_NEM
1562	if (u2State != UINT8_MAX)
1563	PGM_PAGE_SET_NEM_STATE(pPageDst, u2State);
1564	#endif
1565
1566	GCPhys += PAGE_SIZE;
1567	pPageDst++;
1568	}
1569
1570	/* Finish pending and cleanup. */
1571	if (cPendingPages)
1572	{
1573	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1574	AssertLogRelRCReturn(rc, rc);
1575	}
1576	GMMR3FreePagesCleanup(pReq);
1577
1578	return rc;
1579	}
1580
1581
1582	/**
1583	* PGMR3PhysRegisterRam worker that initializes and links a RAM range.
1584	*
1585	* In NEM mode, this will allocate the pages backing the RAM range and this may
1586	* fail. NEM registration may also fail. (In regular HM mode it won't fail.)
1587	*
1588	* @returns VBox status code.
1589	* @param pVM The cross context VM structure.
1590	* @param pNew The new RAM range.
1591	* @param GCPhys The address of the RAM range.
1592	* @param GCPhysLast The last address of the RAM range.
1593	* @param R0PtrNew Ditto for R0.
1594	* @param fFlags PGM_RAM_RANGE_FLAGS_FLOATING or zero.
1595	* @param pszDesc The description.
1596	* @param pPrev The previous RAM range (for linking).
1597	*/
1598	static int pgmR3PhysInitAndLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast,
1599	RTR0PTR R0PtrNew, uint32_t fFlags, const char *pszDesc, PPGMRAMRANGE pPrev)
1600	{
1601	/*
1602	* Initialize the range.
1603	*/
1604	pNew->pSelfR0 = R0PtrNew;
1605	pNew->GCPhys = GCPhys;
1606	pNew->GCPhysLast = GCPhysLast;
1607	pNew->cb = GCPhysLast - GCPhys + 1;
1608	pNew->pszDesc = pszDesc;
1609	pNew->fFlags = fFlags;
1610	pNew->uNemRange = UINT32_MAX;
1611	pNew->pvR3 = NULL;
1612	pNew->paLSPages = NULL;
1613
1614	uint32_t const cPages = pNew->cb >> PAGE_SHIFT;
1615	#ifdef VBOX_WITH_PGM_NEM_MODE
1616	if (!pVM->pgm.s.fNemMode)
1617	#endif
1618	{
1619	RTGCPHYS iPage = cPages;
1620	while (iPage-- > 0)
1621	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_RAM);
1622
1623	/* Update the page count stats. */
1624	pVM->pgm.s.cZeroPages += cPages;
1625	pVM->pgm.s.cAllPages += cPages;
1626	}
1627	#ifdef VBOX_WITH_PGM_NEM_MODE
1628	else
1629	{
1630	int rc = SUPR3PageAlloc(cPages, pVM->pgm.s.fUseLargePages ? SUP_PAGE_ALLOC_F_LARGE_PAGES : 0, &pNew->pvR3);
1631	if (RT_FAILURE(rc))
1632	return rc;
1633
1634	RTGCPHYS iPage = cPages;
1635	while (iPage-- > 0)
1636	PGM_PAGE_INIT(&pNew->aPages[iPage], UINT64_C(0x0000fffffffff000), NIL_GMM_PAGEID,
1637	PGMPAGETYPE_RAM, PGM_PAGE_STATE_ALLOCATED);
1638
1639	/* Update the page count stats. */
1640	pVM->pgm.s.cPrivatePages += cPages;
1641	pVM->pgm.s.cAllPages += cPages;
1642	}
1643	#endif
1644
1645	/*
1646	* Link it.
1647	*/
1648	pgmR3PhysLinkRamRange(pVM, pNew, pPrev);
1649
1650	#ifdef VBOX_WITH_NATIVE_NEM
1651	/*
1652	* Notify NEM now that it has been linked.
1653	*/
1654	if (VM_IS_NEM_ENABLED(pVM))
1655	{
1656	uint8_t u2State = UINT8_MAX;
1657	int rc = NEMR3NotifyPhysRamRegister(pVM, GCPhys, pNew->cb, pNew->pvR3, &u2State, &pNew->uNemRange);
1658	if (RT_SUCCESS(rc))
1659	{
1660	if (u2State != UINT8_MAX)
1661	pgmPhysSetNemStateForPages(&pNew->aPages[0], cPages, u2State);
1662	}
1663	else
1664	pgmR3PhysUnlinkRamRange2(pVM, pNew, pPrev);
1665	return rc;
1666	}
1667	#endif
1668	return VINF_SUCCESS;
1669	}
1670
1671
1672	/**
1673	* PGMR3PhysRegisterRam worker that registers a high chunk.
1674	*
1675	* @returns VBox status code.
1676	* @param pVM The cross context VM structure.
1677	* @param GCPhys The address of the RAM.
1678	* @param cRamPages The number of RAM pages to register.
1679	* @param iChunk The chunk number.
1680	* @param pszDesc The RAM range description.
1681	* @param ppPrev Previous RAM range pointer. In/Out.
1682	*/
1683	static int pgmR3PhysRegisterHighRamChunk(PVM pVM, RTGCPHYS GCPhys, uint32_t cRamPages, uint32_t iChunk,
1684	const char pszDesc, PPGMRAMRANGE ppPrev)
1685	{
1686	const char *pszDescChunk = iChunk == 0
1687	? pszDesc
1688	: MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s (#%u)", pszDesc, iChunk + 1);
1689	AssertReturn(pszDescChunk, VERR_NO_MEMORY);
1690
1691	/*
1692	* Allocate memory for the new chunk.
1693	*/
1694	size_t const cChunkPages = RT_ALIGN_Z(RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cRamPages]), PAGE_SIZE) >> PAGE_SHIFT;
1695	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
1696	AssertReturn(paChunkPages, VERR_NO_TMP_MEMORY);
1697	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
1698	void *pvChunk = NULL;
1699	int rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk, &R0PtrChunk, paChunkPages);
1700	if (RT_SUCCESS(rc))
1701	{
1702	Assert(R0PtrChunk != NIL_RTR0PTR \|\| PGM_IS_IN_NEM_MODE(pVM));
1703	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
1704
1705	PPGMRAMRANGE pNew = (PPGMRAMRANGE)pvChunk;
1706
1707	/*
1708	* Ok, init and link the range.
1709	*/
1710	rc = pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhys + ((RTGCPHYS)cRamPages << PAGE_SHIFT) - 1,
1711	R0PtrChunk, PGM_RAM_RANGE_FLAGS_FLOATING, pszDescChunk, *ppPrev);
1712	if (RT_SUCCESS(rc))
1713	*ppPrev = pNew;
1714
1715	if (RT_FAILURE(rc))
1716	SUPR3PageFreeEx(pvChunk, cChunkPages);
1717	}
1718
1719	RTMemTmpFree(paChunkPages);
1720	return rc;
1721	}
1722
1723
1724	/**
1725	* Sets up a range RAM.
1726	*
1727	* This will check for conflicting registrations, make a resource
1728	* reservation for the memory (with GMM), and setup the per-page
1729	* tracking structures (PGMPAGE).
1730	*
1731	* @returns VBox status code.
1732	* @param pVM The cross context VM structure.
1733	* @param GCPhys The physical address of the RAM.
1734	* @param cb The size of the RAM.
1735	* @param pszDesc The description - not copied, so, don't free or change it.
1736	*/
1737	VMMR3DECL(int) PGMR3PhysRegisterRam(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, const char *pszDesc)
1738	{
1739	/*
1740	* Validate input.
1741	*/
1742	Log(("PGMR3PhysRegisterRam: GCPhys=%RGp cb=%RGp pszDesc=%s\n", GCPhys, cb, pszDesc));
1743	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
1744	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
1745	AssertReturn(cb > 0, VERR_INVALID_PARAMETER);
1746	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
1747	AssertMsgReturn(GCPhysLast > GCPhys, ("The range wraps! GCPhys=%RGp cb=%RGp\n", GCPhys, cb), VERR_INVALID_PARAMETER);
1748	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
1749	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1750
1751	PGM_LOCK_VOID(pVM);
1752
1753	/*
1754	* Find range location and check for conflicts.
1755	*/
1756	PPGMRAMRANGE pPrev = NULL;
1757	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
1758	while (pRam && GCPhysLast >= pRam->GCPhys)
1759	{
1760	AssertLogRelMsgReturnStmt( GCPhysLast < pRam->GCPhys
1761	\|\| GCPhys > pRam->GCPhysLast,
1762	("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
1763	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
1764	PGM_UNLOCK(pVM), VERR_PGM_RAM_CONFLICT);
1765
1766	/* next */
1767	pPrev = pRam;
1768	pRam = pRam->pNextR3;
1769	}
1770
1771	/*
1772	* Register it with GMM (the API bitches).
1773	*/
1774	const RTGCPHYS cPages = cb >> PAGE_SHIFT;
1775	int rc = MMR3IncreaseBaseReservation(pVM, cPages);
1776	if (RT_FAILURE(rc))
1777	{
1778	PGM_UNLOCK(pVM);
1779	return rc;
1780	}
1781
1782	if ( GCPhys >= _4G
1783	&& cPages > 256)
1784	{
1785	/*
1786	* The PGMRAMRANGE structures for the high memory can get very big.
1787	* There used to be some limitations on SUPR3PageAllocEx allocation
1788	* sizes, so traditionally we limited this to 16MB chunks. These days
1789	* we do ~64 MB chunks each covering 16GB of guest RAM, making sure
1790	* each range is a multiple of 1GB to enable eager hosts to use 1GB
1791	* pages in NEM mode.
1792	*
1793	* See also pgmR3PhysMmio2CalcChunkCount.
1794	*/
1795	uint32_t const cPagesPerChunk = _4M;
1796	Assert(RT_ALIGN_32(cPagesPerChunk, X86_PD_PAE_SHIFT - X86_PAGE_SHIFT)); /* NEM large page requirement: 1GB pages. */
1797
1798	RTGCPHYS cPagesLeft = cPages;
1799	RTGCPHYS GCPhysChunk = GCPhys;
1800	uint32_t iChunk = 0;
1801	while (cPagesLeft > 0)
1802	{
1803	uint32_t cPagesInChunk = cPagesLeft;
1804	if (cPagesInChunk > cPagesPerChunk)
1805	cPagesInChunk = cPagesPerChunk;
1806
1807	rc = pgmR3PhysRegisterHighRamChunk(pVM, GCPhysChunk, cPagesInChunk, iChunk, pszDesc, &pPrev);
1808	AssertRCReturn(rc, rc);
1809
1810	/* advance */
1811	GCPhysChunk += (RTGCPHYS)cPagesInChunk << PAGE_SHIFT;
1812	cPagesLeft -= cPagesInChunk;
1813	iChunk++;
1814	}
1815	}
1816	else
1817	{
1818	/*
1819	* Allocate, initialize and link the new RAM range.
1820	*/
1821	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
1822	PPGMRAMRANGE pNew;
1823	rc = MMR3HyperAllocOnceNoRel(pVM, cbRamRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
1824	AssertLogRelMsgRCReturn(rc, ("rc=%Rrc cbRamRange=%zu\n", rc, cbRamRange), rc);
1825
1826	rc = pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhysLast, MMHyperCCToR0(pVM, pNew), 0 /fFlags/, pszDesc, pPrev);
1827	AssertLogRelMsgRCReturn(rc, ("rc=%Rrc cbRamRange=%zu\n", rc, cbRamRange), rc);
1828	}
1829	pgmPhysInvalidatePageMapTLB(pVM);
1830
1831	PGM_UNLOCK(pVM);
1832	return rc;
1833	}
1834
1835
1836	/**
1837	* Worker called by PGMR3InitFinalize if we're configured to pre-allocate RAM.
1838	*
1839	* We do this late in the init process so that all the ROM and MMIO ranges have
1840	* been registered already and we don't go wasting memory on them.
1841	*
1842	* @returns VBox status code.
1843	*
1844	* @param pVM The cross context VM structure.
1845	*/
1846	int pgmR3PhysRamPreAllocate(PVM pVM)
1847	{
1848	Assert(pVM->pgm.s.fRamPreAlloc);
1849	Log(("pgmR3PhysRamPreAllocate: enter\n"));
1850	#ifdef VBOX_WITH_PGM_NEM_MODE
1851	AssertLogRelReturn(!pVM->pgm.s.fNemMode, VERR_PGM_NOT_SUPPORTED_FOR_NEM_MODE);
1852	#endif
1853
1854	/*
1855	* Walk the RAM ranges and allocate all RAM pages, halt at
1856	* the first allocation error.
1857	*/
1858	uint64_t cPages = 0;
1859	uint64_t NanoTS = RTTimeNanoTS();
1860	PGM_LOCK_VOID(pVM);
1861	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1862	{
1863	PPGMPAGE pPage = &pRam->aPages[0];
1864	RTGCPHYS GCPhys = pRam->GCPhys;
1865	uint32_t cLeft = pRam->cb >> PAGE_SHIFT;
1866	while (cLeft-- > 0)
1867	{
1868	if (PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1869	{
1870	switch (PGM_PAGE_GET_STATE(pPage))
1871	{
1872	case PGM_PAGE_STATE_ZERO:
1873	{
1874	int rc = pgmPhysAllocPage(pVM, pPage, GCPhys);
1875	if (RT_FAILURE(rc))
1876	{
1877	LogRel(("PGM: RAM Pre-allocation failed at %RGp (in %s) with rc=%Rrc\n", GCPhys, pRam->pszDesc, rc));
1878	PGM_UNLOCK(pVM);
1879	return rc;
1880	}
1881	cPages++;
1882	break;
1883	}
1884
1885	case PGM_PAGE_STATE_BALLOONED:
1886	case PGM_PAGE_STATE_ALLOCATED:
1887	case PGM_PAGE_STATE_WRITE_MONITORED:
1888	case PGM_PAGE_STATE_SHARED:
1889	/* nothing to do here. */
1890	break;
1891	}
1892	}
1893
1894	/* next */
1895	pPage++;
1896	GCPhys += PAGE_SIZE;
1897	}
1898	}
1899	PGM_UNLOCK(pVM);
1900	NanoTS = RTTimeNanoTS() - NanoTS;
1901
1902	LogRel(("PGM: Pre-allocated %llu pages in %llu ms\n", cPages, NanoTS / 1000000));
1903	Log(("pgmR3PhysRamPreAllocate: returns VINF_SUCCESS\n"));
1904	return VINF_SUCCESS;
1905	}
1906
1907
1908	/**
1909	* Checks shared page checksums.
1910	*
1911	* @param pVM The cross context VM structure.
1912	*/
1913	void pgmR3PhysAssertSharedPageChecksums(PVM pVM)
1914	{
1915	#ifdef VBOX_STRICT
1916	PGM_LOCK_VOID(pVM);
1917
1918	if (pVM->pgm.s.cSharedPages > 0)
1919	{
1920	/*
1921	* Walk the ram ranges.
1922	*/
1923	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1924	{
1925	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
1926	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
1927
1928	while (iPage-- > 0)
1929	{
1930	PPGMPAGE pPage = &pRam->aPages[iPage];
1931	if (PGM_PAGE_IS_SHARED(pPage))
1932	{
1933	uint32_t u32Checksum = pPage->s.u2Unused0/* \| ((uint32_t)pPage->s.u2Unused1 << 8)*/;
1934	if (!u32Checksum)
1935	{
1936	RTGCPHYS GCPhysPage = pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT);
1937	void const *pvPage;
1938	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhysPage, &pvPage);
1939	if (RT_SUCCESS(rc))
1940	{
1941	uint32_t u32Checksum2 = RTCrc32(pvPage, PAGE_SIZE);
1942	# if 0
1943	AssertMsg((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum, ("GCPhysPage=%RGp\n", GCPhysPage));
1944	# else
1945	if ((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum)
1946	LogFlow(("shpg %#x @ %RGp %#x [OK]\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
1947	else
1948	AssertMsgFailed(("shpg %#x @ %RGp %#x\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
1949	# endif
1950	}
1951	else
1952	AssertRC(rc);
1953	}
1954	}
1955
1956	} /* for each page */
1957
1958	} /* for each ram range */
1959	}
1960
1961	PGM_UNLOCK(pVM);
1962	#endif /* VBOX_STRICT */
1963	NOREF(pVM);
1964	}
1965
1966
1967	/**
1968	* Resets the physical memory state.
1969	*
1970	* ASSUMES that the caller owns the PGM lock.
1971	*
1972	* @returns VBox status code.
1973	* @param pVM The cross context VM structure.
1974	*/
1975	int pgmR3PhysRamReset(PVM pVM)
1976	{
1977	PGM_LOCK_ASSERT_OWNER(pVM);
1978
1979	/* Reset the memory balloon. */
1980	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
1981	AssertRC(rc);
1982
1983	#ifdef VBOX_WITH_PAGE_SHARING
1984	/* Clear all registered shared modules. */
1985	pgmR3PhysAssertSharedPageChecksums(pVM);
1986	rc = GMMR3ResetSharedModules(pVM);
1987	AssertRC(rc);
1988	#endif
1989	/* Reset counters. */
1990	pVM->pgm.s.cReusedSharedPages = 0;
1991	pVM->pgm.s.cBalloonedPages = 0;
1992
1993	return VINF_SUCCESS;
1994	}
1995
1996
1997	/**
1998	* Resets (zeros) the RAM after all devices and components have been reset.
1999	*
2000	* ASSUMES that the caller owns the PGM lock.
2001	*
2002	* @returns VBox status code.
2003	* @param pVM The cross context VM structure.
2004	*/
2005	int pgmR3PhysRamZeroAll(PVM pVM)
2006	{
2007	PGM_LOCK_ASSERT_OWNER(pVM);
2008
2009	/*
2010	* We batch up pages that should be freed instead of calling GMM for
2011	* each and every one of them.
2012	*/
2013	uint32_t cPendingPages = 0;
2014	PGMMFREEPAGESREQ pReq;
2015	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2016	AssertLogRelRCReturn(rc, rc);
2017
2018	/*
2019	* Walk the ram ranges.
2020	*/
2021	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2022	{
2023	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2024	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2025
2026	if ( !pVM->pgm.s.fRamPreAlloc
2027	#ifdef VBOX_WITH_PGM_NEM_MODE
2028	&& !pVM->pgm.s.fNemMode
2029	#endif
2030	&& pVM->pgm.s.fZeroRamPagesOnReset)
2031	{
2032	/* Replace all RAM pages by ZERO pages. */
2033	while (iPage-- > 0)
2034	{
2035	PPGMPAGE pPage = &pRam->aPages[iPage];
2036	switch (PGM_PAGE_GET_TYPE(pPage))
2037	{
2038	case PGMPAGETYPE_RAM:
2039	/* Do not replace pages part of a 2 MB continuous range
2040	with zero pages, but zero them instead. */
2041	if ( PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE
2042	\|\| PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE_DISABLED)
2043	{
2044	void *pvPage;
2045	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2046	AssertLogRelRCReturn(rc, rc);
2047	ASMMemZeroPage(pvPage);
2048	}
2049	else if (PGM_PAGE_IS_BALLOONED(pPage))
2050	{
2051	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2052	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2053	}
2054	else if (!PGM_PAGE_IS_ZERO(pPage))
2055	{
2056	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2057	PGMPAGETYPE_RAM);
2058	AssertLogRelRCReturn(rc, rc);
2059	}
2060	break;
2061
2062	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2063	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2064	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2065	pRam, true /fDoAccounting/);
2066	break;
2067
2068	case PGMPAGETYPE_MMIO2:
2069	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2070	case PGMPAGETYPE_ROM:
2071	case PGMPAGETYPE_MMIO:
2072	break;
2073	default:
2074	AssertFailed();
2075	}
2076	} /* for each page */
2077	}
2078	else
2079	{
2080	/* Zero the memory. */
2081	while (iPage-- > 0)
2082	{
2083	PPGMPAGE pPage = &pRam->aPages[iPage];
2084	switch (PGM_PAGE_GET_TYPE(pPage))
2085	{
2086	case PGMPAGETYPE_RAM:
2087	switch (PGM_PAGE_GET_STATE(pPage))
2088	{
2089	case PGM_PAGE_STATE_ZERO:
2090	break;
2091
2092	case PGM_PAGE_STATE_BALLOONED:
2093	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2094	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2095	break;
2096
2097	case PGM_PAGE_STATE_SHARED:
2098	case PGM_PAGE_STATE_WRITE_MONITORED:
2099	rc = pgmPhysPageMakeWritable(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2100	AssertLogRelRCReturn(rc, rc);
2101	RT_FALL_THRU();
2102
2103	case PGM_PAGE_STATE_ALLOCATED:
2104	if (pVM->pgm.s.fZeroRamPagesOnReset)
2105	{
2106	void *pvPage;
2107	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2108	AssertLogRelRCReturn(rc, rc);
2109	ASMMemZeroPage(pvPage);
2110	}
2111	break;
2112	}
2113	break;
2114
2115	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2116	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2117	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2118	pRam, true /fDoAccounting/);
2119	break;
2120
2121	case PGMPAGETYPE_MMIO2:
2122	case PGMPAGETYPE_ROM_SHADOW:
2123	case PGMPAGETYPE_ROM:
2124	case PGMPAGETYPE_MMIO:
2125	break;
2126	default:
2127	AssertFailed();
2128
2129	}
2130	} /* for each page */
2131	}
2132
2133	}
2134
2135	/*
2136	* Finish off any pages pending freeing.
2137	*/
2138	if (cPendingPages)
2139	{
2140	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2141	AssertLogRelRCReturn(rc, rc);
2142	}
2143	GMMR3FreePagesCleanup(pReq);
2144	return VINF_SUCCESS;
2145	}
2146
2147
2148	/**
2149	* Frees all RAM during VM termination
2150	*
2151	* ASSUMES that the caller owns the PGM lock.
2152	*
2153	* @returns VBox status code.
2154	* @param pVM The cross context VM structure.
2155	*/
2156	int pgmR3PhysRamTerm(PVM pVM)
2157	{
2158	PGM_LOCK_ASSERT_OWNER(pVM);
2159
2160	/* Reset the memory balloon. */
2161	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2162	AssertRC(rc);
2163
2164	#ifdef VBOX_WITH_PAGE_SHARING
2165	/*
2166	* Clear all registered shared modules.
2167	*/
2168	pgmR3PhysAssertSharedPageChecksums(pVM);
2169	rc = GMMR3ResetSharedModules(pVM);
2170	AssertRC(rc);
2171
2172	/*
2173	* Flush the handy pages updates to make sure no shared pages are hiding
2174	* in there. (Not unlikely if the VM shuts down, apparently.)
2175	*/
2176	# ifdef VBOX_WITH_PGM_NEM_MODE
2177	if (!pVM->pgm.s.fNemMode)
2178	# endif
2179	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_FLUSH_HANDY_PAGES, 0, NULL);
2180	#endif
2181
2182	/*
2183	* We batch up pages that should be freed instead of calling GMM for
2184	* each and every one of them.
2185	*/
2186	uint32_t cPendingPages = 0;
2187	PGMMFREEPAGESREQ pReq;
2188	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2189	AssertLogRelRCReturn(rc, rc);
2190
2191	/*
2192	* Walk the ram ranges.
2193	*/
2194	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2195	{
2196	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2197	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2198
2199	while (iPage-- > 0)
2200	{
2201	PPGMPAGE pPage = &pRam->aPages[iPage];
2202	switch (PGM_PAGE_GET_TYPE(pPage))
2203	{
2204	case PGMPAGETYPE_RAM:
2205	/* Free all shared pages. Private pages are automatically freed during GMM VM cleanup. */
2206	/** @todo change this to explicitly free private pages here. */
2207	if (PGM_PAGE_IS_SHARED(pPage))
2208	{
2209	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2210	PGMPAGETYPE_RAM);
2211	AssertLogRelRCReturn(rc, rc);
2212	}
2213	break;
2214
2215	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2216	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO:
2217	case PGMPAGETYPE_MMIO2:
2218	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2219	case PGMPAGETYPE_ROM:
2220	case PGMPAGETYPE_MMIO:
2221	break;
2222	default:
2223	AssertFailed();
2224	}
2225	} /* for each page */
2226	}
2227
2228	/*
2229	* Finish off any pages pending freeing.
2230	*/
2231	if (cPendingPages)
2232	{
2233	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2234	AssertLogRelRCReturn(rc, rc);
2235	}
2236	GMMR3FreePagesCleanup(pReq);
2237	return VINF_SUCCESS;
2238	}
2239
2240
2241
2242	/*********************************************************************************************************************************
2243	* MMIO *
2244	*********************************************************************************************************************************/
2245
2246	/**
2247	* This is the interface IOM is using to register an MMIO region.
2248	*
2249	* It will check for conflicts and ensure that a RAM range structure
2250	* is present before calling the PGMR3HandlerPhysicalRegister API to
2251	* register the callbacks.
2252	*
2253	* @returns VBox status code.
2254	*
2255	* @param pVM The cross context VM structure.
2256	* @param GCPhys The start of the MMIO region.
2257	* @param cb The size of the MMIO region.
2258	* @param hType The physical access handler type registration.
2259	* @param pvUserR3 The user argument for R3.
2260	* @param pvUserR0 The user argument for R0.
2261	* @param pvUserRC The user argument for RC.
2262	* @param pszDesc The description of the MMIO region.
2263	*/
2264	VMMR3DECL(int) PGMR3PhysMMIORegister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMPHYSHANDLERTYPE hType,
2265	RTR3PTR pvUserR3, RTR0PTR pvUserR0, RTRCPTR pvUserRC, const char *pszDesc)
2266	{
2267	/*
2268	* Assert on some assumption.
2269	*/
2270	VM_ASSERT_EMT(pVM);
2271	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2272	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2273	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2274	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2275	Assert(((PPGMPHYSHANDLERTYPEINT)MMHyperHeapOffsetToPtr(pVM, hType))->enmKind == PGMPHYSHANDLERKIND_MMIO);
2276
2277	int rc = PGM_LOCK(pVM);
2278	AssertRCReturn(rc, rc);
2279
2280	/*
2281	* Make sure there's a RAM range structure for the region.
2282	*/
2283	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2284	bool fRamExists = false;
2285	PPGMRAMRANGE pRamPrev = NULL;
2286	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2287	while (pRam && GCPhysLast >= pRam->GCPhys)
2288	{
2289	if ( GCPhysLast >= pRam->GCPhys
2290	&& GCPhys <= pRam->GCPhysLast)
2291	{
2292	/* Simplification: all within the same range. */
2293	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
2294	&& GCPhysLast <= pRam->GCPhysLast,
2295	("%RGp-%RGp (MMIO/%s) falls partly outside %RGp-%RGp (%s)\n",
2296	GCPhys, GCPhysLast, pszDesc,
2297	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2298	PGM_UNLOCK(pVM),
2299	VERR_PGM_RAM_CONFLICT);
2300
2301	/* Check that it's all RAM or MMIO pages. */
2302	PCPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2303	uint32_t cLeft = cb >> PAGE_SHIFT;
2304	while (cLeft-- > 0)
2305	{
2306	AssertLogRelMsgReturnStmt( PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM
2307	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO,
2308	("%RGp-%RGp (MMIO/%s): %RGp is not a RAM or MMIO page - type=%d desc=%s\n",
2309	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, PGM_PAGE_GET_TYPE(pPage), pRam->pszDesc),
2310	PGM_UNLOCK(pVM),
2311	VERR_PGM_RAM_CONFLICT);
2312	pPage++;
2313	}
2314
2315	/* Looks good. */
2316	fRamExists = true;
2317	break;
2318	}
2319
2320	/* next */
2321	pRamPrev = pRam;
2322	pRam = pRam->pNextR3;
2323	}
2324	PPGMRAMRANGE pNew;
2325	if (fRamExists)
2326	{
2327	pNew = NULL;
2328
2329	/*
2330	* Make all the pages in the range MMIO/ZERO pages, freeing any
2331	* RAM pages currently mapped here. This might not be 100% correct
2332	* for PCI memory, but we're doing the same thing for MMIO2 pages.
2333	*/
2334	rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, NULL);
2335	AssertRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
2336
2337	/* Force a PGM pool flush as guest ram references have been changed. */
2338	/** @todo not entirely SMP safe; assuming for now the guest takes
2339	* care of this internally (not touch mapped mmio while changing the
2340	* mapping). */
2341	PVMCPU pVCpu = VMMGetCpu(pVM);
2342	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2343	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2344	}
2345	else
2346	{
2347	/*
2348	* No RAM range, insert an ad hoc one.
2349	*
2350	* Note that we don't have to tell REM about this range because
2351	* PGMHandlerPhysicalRegisterEx will do that for us.
2352	*/
2353	Log(("PGMR3PhysMMIORegister: Adding ad hoc MMIO range for %RGp-%RGp %s\n", GCPhys, GCPhysLast, pszDesc));
2354
2355	/* Alloc. */
2356	const uint32_t cPages = cb >> PAGE_SHIFT;
2357	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
2358	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), 16, MM_TAG_PGM_PHYS, (void **)&pNew);
2359	AssertLogRelMsgRCReturnStmt(rc, ("cbRamRange=%zu\n", cbRamRange), PGM_UNLOCK(pVM), rc);
2360
2361	#ifdef VBOX_WITH_NATIVE_NEM
2362	/* Notify NEM. */
2363	uint8_t u2State = 0; /* (must have valid state as there can't be anything to preserve) */
2364	if (VM_IS_NEM_ENABLED(pVM))
2365	{
2366	rc = NEMR3NotifyPhysMmioExMapEarly(pVM, GCPhys, cPages << PAGE_SHIFT, 0 /fFlags/, NULL, NULL,
2367	&u2State, &pNew->uNemRange);
2368	AssertLogRelRCReturnStmt(rc, MMHyperFree(pVM, pNew), rc);
2369	}
2370	#endif
2371
2372	/* Initialize the range. */
2373	pNew->pSelfR0 = MMHyperCCToR0(pVM, pNew);
2374	pNew->GCPhys = GCPhys;
2375	pNew->GCPhysLast = GCPhysLast;
2376	pNew->cb = cb;
2377	pNew->pszDesc = pszDesc;
2378	pNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO;
2379	pNew->pvR3 = NULL;
2380	pNew->paLSPages = NULL;
2381
2382	uint32_t iPage = cPages;
2383	while (iPage-- > 0)
2384	{
2385	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_MMIO);
2386	#ifdef VBOX_WITH_NATIVE_NEM
2387	PGM_PAGE_SET_NEM_STATE(&pNew->aPages[iPage], u2State);
2388	#endif
2389	}
2390	Assert(PGM_PAGE_GET_TYPE(&pNew->aPages[0]) == PGMPAGETYPE_MMIO);
2391
2392	/* update the page count stats. */
2393	pVM->pgm.s.cPureMmioPages += cPages;
2394	pVM->pgm.s.cAllPages += cPages;
2395
2396	/* link it */
2397	pgmR3PhysLinkRamRange(pVM, pNew, pRamPrev);
2398	}
2399
2400	/*
2401	* Register the access handler.
2402	*/
2403	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, hType, pvUserR3, pvUserR0, pvUserRC, pszDesc);
2404	if (RT_SUCCESS(rc))
2405	{
2406	#ifdef VBOX_WITH_NATIVE_NEM
2407	/* Late NEM notification. */
2408	if (VM_IS_NEM_ENABLED(pVM))
2409	{
2410	uint32_t const fNemNotify = (fRamExists ? NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE : 0);
2411	rc = NEMR3NotifyPhysMmioExMapLate(pVM, GCPhys, GCPhysLast - GCPhys + 1, fNemNotify,
2412	fRamExists ? (uint8_t *)pRam->pvR3 + (uintptr_t)(GCPhys - pRam->GCPhys) : NULL,
2413	NULL, !fRamExists ? &pRam->uNemRange : NULL);
2414	AssertLogRelRCReturn(rc, rc);
2415	}
2416	#endif
2417	}
2418	/** @todo the phys handler failure handling isn't complete, esp. wrt NEM. */
2419	else if (!fRamExists)
2420	{
2421	pVM->pgm.s.cPureMmioPages -= cb >> PAGE_SHIFT;
2422	pVM->pgm.s.cAllPages -= cb >> PAGE_SHIFT;
2423
2424	/* remove the ad hoc range. */
2425	pgmR3PhysUnlinkRamRange2(pVM, pNew, pRamPrev);
2426	pNew->cb = pNew->GCPhys = pNew->GCPhysLast = NIL_RTGCPHYS;
2427	MMHyperFree(pVM, pRam);
2428	}
2429	pgmPhysInvalidatePageMapTLB(pVM);
2430
2431	PGM_UNLOCK(pVM);
2432	return rc;
2433	}
2434
2435
2436	/**
2437	* This is the interface IOM is using to register an MMIO region.
2438	*
2439	* It will take care of calling PGMHandlerPhysicalDeregister and clean up
2440	* any ad hoc PGMRAMRANGE left behind.
2441	*
2442	* @returns VBox status code.
2443	* @param pVM The cross context VM structure.
2444	* @param GCPhys The start of the MMIO region.
2445	* @param cb The size of the MMIO region.
2446	*/
2447	VMMR3DECL(int) PGMR3PhysMMIODeregister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb)
2448	{
2449	VM_ASSERT_EMT(pVM);
2450
2451	int rc = PGM_LOCK(pVM);
2452	AssertRCReturn(rc, rc);
2453
2454	/*
2455	* First deregister the handler, then check if we should remove the ram range.
2456	*/
2457	rc = PGMHandlerPhysicalDeregister(pVM, GCPhys);
2458	if (RT_SUCCESS(rc))
2459	{
2460	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2461	PPGMRAMRANGE pRamPrev = NULL;
2462	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2463	while (pRam && GCPhysLast >= pRam->GCPhys)
2464	{
2465	/** @todo We're being a bit too careful here. rewrite. */
2466	if ( GCPhysLast == pRam->GCPhysLast
2467	&& GCPhys == pRam->GCPhys)
2468	{
2469	Assert(pRam->cb == cb);
2470
2471	/*
2472	* See if all the pages are dead MMIO pages.
2473	*/
2474	uint32_t const cPages = cb >> PAGE_SHIFT;
2475	bool fAllMMIO = true;
2476	uint32_t iPage = 0;
2477	uint32_t cLeft = cPages;
2478	while (cLeft-- > 0)
2479	{
2480	PPGMPAGE pPage = &pRam->aPages[iPage];
2481	if ( !PGM_PAGE_IS_MMIO_OR_ALIAS(pPage)
2482	/\|\| not-out-of-action later /)
2483	{
2484	fAllMMIO = false;
2485	AssertMsgFailed(("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2486	break;
2487	}
2488	Assert( PGM_PAGE_IS_ZERO(pPage)
2489	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2490	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO);
2491	pPage++;
2492	}
2493	if (fAllMMIO)
2494	{
2495	/*
2496	* Ad-hoc range, unlink and free it.
2497	*/
2498	Log(("PGMR3PhysMMIODeregister: Freeing ad hoc MMIO range for %RGp-%RGp %s\n",
2499	GCPhys, GCPhysLast, pRam->pszDesc));
2500	/** @todo check the ad-hoc flags? */
2501
2502	#ifdef VBOX_WITH_NATIVE_NEM
2503	if (VM_IS_NEM_ENABLED(pVM)) /* Notify REM before we unlink the range. */
2504	{
2505	rc = NEMR3NotifyPhysMmioExUnmap(pVM, GCPhys, GCPhysLast - GCPhys + 1, 0 /fFlags/,
2506	NULL, NULL, NULL, &pRam->uNemRange);
2507	AssertLogRelRCReturn(rc, rc);
2508	}
2509	#endif
2510
2511	pVM->pgm.s.cAllPages -= cPages;
2512	pVM->pgm.s.cPureMmioPages -= cPages;
2513
2514	pgmR3PhysUnlinkRamRange2(pVM, pRam, pRamPrev);
2515	pRam->cb = pRam->GCPhys = pRam->GCPhysLast = NIL_RTGCPHYS;
2516	MMHyperFree(pVM, pRam);
2517	break;
2518	}
2519	}
2520
2521	/*
2522	* Range match? It will all be within one range (see PGMAllHandler.cpp).
2523	*/
2524	if ( GCPhysLast >= pRam->GCPhys
2525	&& GCPhys <= pRam->GCPhysLast)
2526	{
2527	Assert(GCPhys >= pRam->GCPhys);
2528	Assert(GCPhysLast <= pRam->GCPhysLast);
2529
2530	/*
2531	* Turn the pages back into RAM pages.
2532	*/
2533	uint32_t iPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2534	uint32_t cLeft = cb >> PAGE_SHIFT;
2535	while (cLeft--)
2536	{
2537	PPGMPAGE pPage = &pRam->aPages[iPage];
2538	AssertMsg( (PGM_PAGE_IS_MMIO(pPage) && PGM_PAGE_IS_ZERO(pPage))
2539	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2540	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO,
2541	("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2542	if (PGM_PAGE_IS_MMIO_OR_ALIAS(pPage))
2543	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_RAM);
2544	iPage++;
2545	}
2546
2547	#ifdef VBOX_WITH_NATIVE_NEM
2548	/* Notify REM (failure will probably leave things in a non-working state). */
2549	if (VM_IS_NEM_ENABLED(pVM))
2550	{
2551	uint8_t u2State = UINT8_MAX;
2552	rc = NEMR3NotifyPhysMmioExUnmap(pVM, GCPhys, GCPhysLast - GCPhys + 1, NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE,
2553	pRam->pvR3 ? (uint8_t *)pRam->pvR3 + GCPhys - pRam->GCPhys : NULL,
2554	NULL, &u2State, &pRam->uNemRange);
2555	AssertLogRelRCReturn(rc, rc);
2556	if (u2State != UINT8_MAX)
2557	pgmPhysSetNemStateForPages(&pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT],
2558	cb >> PAGE_SHIFT, u2State);
2559	}
2560	#endif
2561	break;
2562	}
2563
2564	/* next */
2565	pRamPrev = pRam;
2566	pRam = pRam->pNextR3;
2567	}
2568	}
2569
2570	/* Force a PGM pool flush as guest ram references have been changed. */
2571	/** @todo Not entirely SMP safe; assuming for now the guest takes care of
2572	* this internally (not touch mapped mmio while changing the mapping). */
2573	PVMCPU pVCpu = VMMGetCpu(pVM);
2574	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2575	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2576
2577	pgmPhysInvalidatePageMapTLB(pVM);
2578	pgmPhysInvalidRamRangeTlbs(pVM);
2579	PGM_UNLOCK(pVM);
2580	return rc;
2581	}
2582
2583
2584
2585	/*********************************************************************************************************************************
2586	* MMIO2 *
2587	*********************************************************************************************************************************/
2588
2589	/**
2590	* Locate a MMIO2 range.
2591	*
2592	* @returns Pointer to the MMIO2 range.
2593	* @param pVM The cross context VM structure.
2594	* @param pDevIns The device instance owning the region.
2595	* @param iSubDev The sub-device number.
2596	* @param iRegion The region.
2597	* @param hMmio2 Handle to look up. If NIL, use the @a iSubDev and
2598	* @a iRegion.
2599	*/
2600	DECLINLINE(PPGMREGMMIO2RANGE) pgmR3PhysMmio2Find(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev,
2601	uint32_t iRegion, PGMMMIO2HANDLE hMmio2)
2602	{
2603	if (hMmio2 != NIL_PGMMMIO2HANDLE)
2604	{
2605	if (hMmio2 <= RT_ELEMENTS(pVM->pgm.s.apMmio2RangesR3) && hMmio2 != 0)
2606	{
2607	PPGMREGMMIO2RANGE pCur = pVM->pgm.s.apMmio2RangesR3[hMmio2 - 1];
2608	if (pCur && pCur->pDevInsR3 == pDevIns)
2609	{
2610	Assert(pCur->idMmio2 == hMmio2);
2611	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, NULL);
2612	return pCur;
2613	}
2614	Assert(!pCur);
2615	}
2616	for (PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2617	if (pCur->idMmio2 == hMmio2)
2618	{
2619	AssertBreak(pCur->pDevInsR3 == pDevIns);
2620	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, NULL);
2621	return pCur;
2622	}
2623	}
2624	else
2625	{
2626	/*
2627	* Search the list. There shouldn't be many entries.
2628	*/
2629	/** @todo Optimize this lookup! There may now be many entries and it'll
2630	* become really slow when doing MMR3HyperMapMMIO2 and similar. */
2631	for (PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2632	if ( pCur->pDevInsR3 == pDevIns
2633	&& pCur->iRegion == iRegion
2634	&& pCur->iSubDev == iSubDev)
2635	return pCur;
2636	}
2637	return NULL;
2638	}
2639
2640
2641	/**
2642	* Worker for PGMR3PhysMmio2ControlDirtyPageTracking and PGMR3PhysMmio2Map.
2643	*/
2644	static int pgmR3PhysMmio2EnableDirtyPageTracing(PVM pVM, PPGMREGMMIO2RANGE pFirstMmio2)
2645	{
2646	int rc = VINF_SUCCESS;
2647	for (PPGMREGMMIO2RANGE pCurMmio2 = pFirstMmio2; pCurMmio2; pCurMmio2 = pCurMmio2->pNextR3)
2648	{
2649	Assert(!(pCurMmio2->fFlags & PGMREGMMIO2RANGE_F_IS_TRACKING));
2650	int rc2 = pgmHandlerPhysicalExRegister(pVM, pCurMmio2->pPhysHandlerR3, pCurMmio2->RamRange.GCPhys,
2651	pCurMmio2->RamRange.GCPhysLast);
2652	AssertLogRelMsgRC(rc2, ("%#RGp-%#RGp %s failed -> %Rrc\n", pCurMmio2->RamRange.GCPhys, pCurMmio2->RamRange.GCPhysLast,
2653	pCurMmio2->RamRange.pszDesc, rc2));
2654	if (RT_SUCCESS(rc2))
2655	pCurMmio2->fFlags \|= PGMREGMMIO2RANGE_F_IS_TRACKING;
2656	else if (RT_SUCCESS(rc))
2657	rc = rc2;
2658	if (pCurMmio2->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2659	return rc;
2660	}
2661	AssertFailed();
2662	return rc;
2663	}
2664
2665
2666	/**
2667	* Worker for PGMR3PhysMmio2ControlDirtyPageTracking and PGMR3PhysMmio2Unmap.
2668	*/
2669	static int pgmR3PhysMmio2DisableDirtyPageTracing(PVM pVM, PPGMREGMMIO2RANGE pFirstMmio2)
2670	{
2671	for (PPGMREGMMIO2RANGE pCurMmio2 = pFirstMmio2; pCurMmio2; pCurMmio2 = pCurMmio2->pNextR3)
2672	{
2673	if (pCurMmio2->fFlags & PGMREGMMIO2RANGE_F_IS_TRACKING)
2674	{
2675	int rc2 = pgmHandlerPhysicalExDeregister(pVM, pCurMmio2->pPhysHandlerR3);
2676	AssertLogRelMsgRC(rc2, ("%#RGp-%#RGp %s failed -> %Rrc\n", pCurMmio2->RamRange.GCPhys, pCurMmio2->RamRange.GCPhysLast,
2677	pCurMmio2->RamRange.pszDesc, rc2));
2678	pCurMmio2->fFlags &= ~PGMREGMMIO2RANGE_F_IS_TRACKING;
2679	}
2680	if (pCurMmio2->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2681	return VINF_SUCCESS;
2682	}
2683	AssertFailed();
2684	return VINF_SUCCESS;
2685
2686	}
2687
2688
2689	/**
2690	* Calculates the number of chunks
2691	*
2692	* @returns Number of registration chunk needed.
2693	* @param pVM The cross context VM structure.
2694	* @param cb The size of the MMIO/MMIO2 range.
2695	* @param pcPagesPerChunk Where to return the number of pages tracked by each
2696	* chunk. Optional.
2697	* @param pcbChunk Where to return the guest mapping size for a chunk.
2698	*/
2699	static uint16_t pgmR3PhysMmio2CalcChunkCount(PVM pVM, RTGCPHYS cb, uint32_t pcPagesPerChunk, uint32_t pcbChunk)
2700	{
2701	RT_NOREF_PV(pVM); /* without raw mode */
2702
2703	/*
2704	* This is the same calculation as PGMR3PhysRegisterRam does, except we'll be
2705	* needing a few bytes extra the PGMREGMMIO2RANGE structure.
2706	*
2707	* Note! In additions, we've got a 24 bit sub-page range for MMIO2 ranges, leaving
2708	* us with an absolute maximum of 16777215 pages per chunk (close to 64 GB).
2709	*/
2710	uint32_t const cPagesPerChunk = _4M;
2711	Assert(RT_ALIGN_32(cPagesPerChunk, X86_PD_PAE_SHIFT - X86_PAGE_SHIFT)); /* NEM large page requirement: 1GB pages. */
2712	uint32_t const cbChunk = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPagesPerChunk]);
2713	AssertRelease(cPagesPerChunk < _16M);
2714
2715	if (pcbChunk)
2716	*pcbChunk = cbChunk;
2717	if (pcPagesPerChunk)
2718	*pcPagesPerChunk = cPagesPerChunk;
2719
2720	/* Calc the number of chunks we need. */
2721	RTGCPHYS const cPages = cb >> X86_PAGE_SHIFT;
2722	uint16_t cChunks = (uint16_t)((cPages + cPagesPerChunk - 1) / cPagesPerChunk);
2723	AssertRelease((RTGCPHYS)cChunks * cPagesPerChunk >= cPages);
2724	return cChunks;
2725	}
2726
2727
2728	/**
2729	* Worker for PGMR3PhysMMIO2Register that allocates and the PGMREGMMIO2RANGE
2730	* structures and does basic initialization.
2731	*
2732	* Caller must set type specfic members and initialize the PGMPAGE structures.
2733	*
2734	* This was previously also used by PGMR3PhysMmio2PreRegister, a function for
2735	* pre-registering MMIO that was later (6.1) replaced by a new handle based IOM
2736	* interface. The reference to caller and type above is purely historical.
2737	*
2738	* @returns VBox status code.
2739	* @param pVM The cross context VM structure.
2740	* @param pDevIns The device instance owning the region.
2741	* @param iSubDev The sub-device number (internal PCI config number).
2742	* @param iRegion The region number. If the MMIO2 memory is a PCI
2743	* I/O region this number has to be the number of that
2744	* region. Otherwise it can be any number safe
2745	* UINT8_MAX.
2746	* @param cb The size of the region. Must be page aligned.
2747	* @param fFlags PGMPHYS_MMIO2_FLAGS_XXX.
2748	* @param idMmio2 The MMIO2 ID for the first chunk.
2749	* @param pszDesc The description.
2750	* @param ppHeadRet Where to return the pointer to the first
2751	* registration chunk.
2752	*
2753	* @thread EMT
2754	*/
2755	static int pgmR3PhysMmio2Create(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb, uint32_t fFlags,
2756	uint8_t idMmio2, const char pszDesc, PPGMREGMMIO2RANGE ppHeadRet)
2757	{
2758	/*
2759	* Figure out how many chunks we need and of which size.
2760	*/
2761	uint32_t cPagesPerChunk;
2762	uint16_t cChunks = pgmR3PhysMmio2CalcChunkCount(pVM, cb, &cPagesPerChunk, NULL);
2763	AssertReturn(cChunks, VERR_PGM_PHYS_MMIO_EX_IPE);
2764
2765	/*
2766	* Allocate the chunks.
2767	*/
2768	PPGMREGMMIO2RANGE *ppNext = ppHeadRet;
2769	*ppNext = NULL;
2770
2771	int rc = VINF_SUCCESS;
2772	uint32_t cPagesLeft = cb >> X86_PAGE_SHIFT;
2773	for (uint16_t iChunk = 0; iChunk < cChunks && RT_SUCCESS(rc); iChunk++, idMmio2++)
2774	{
2775	/*
2776	* We currently do a single RAM range for the whole thing. This will
2777	* probably have to change once someone needs really large MMIO regions,
2778	* as we will be running into SUPR3PageAllocEx limitations and such.
2779	*/
2780	const uint32_t cPagesTrackedByChunk = RT_MIN(cPagesLeft, cPagesPerChunk);
2781	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPagesTrackedByChunk]);
2782	PPGMREGMMIO2RANGE pNew = NULL;
2783	if ( iChunk + 1 < cChunks
2784	\|\| cbRange >= _1M)
2785	{
2786	/*
2787	* Allocate memory for the registration structure.
2788	*/
2789	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2790	size_t const cbChunk = (1 + cChunkPages + 1) << PAGE_SHIFT;
2791	AssertLogRelBreakStmt(cbChunk == (uint32_t)cbChunk, rc = VERR_OUT_OF_RANGE);
2792	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
2793	AssertBreakStmt(paChunkPages, rc = VERR_NO_TMP_MEMORY);
2794	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
2795	void *pvChunk = NULL;
2796	rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk, &R0PtrChunk, paChunkPages);
2797	AssertLogRelMsgRCBreakStmt(rc, ("rc=%Rrc, cChunkPages=%#zx\n", rc, cChunkPages), RTMemTmpFree(paChunkPages));
2798
2799	Assert(R0PtrChunk != NIL_RTR0PTR \|\| PGM_IS_IN_NEM_MODE(pVM));
2800	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
2801
2802	pNew = (PPGMREGMMIO2RANGE)pvChunk;
2803	pNew->RamRange.fFlags = PGM_RAM_RANGE_FLAGS_FLOATING;
2804	pNew->RamRange.pSelfR0 = R0PtrChunk + RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange);
2805
2806	RTMemTmpFree(paChunkPages);
2807	}
2808	/*
2809	* Not so big, do a one time hyper allocation.
2810	*/
2811	else
2812	{
2813	rc = MMR3HyperAllocOnceNoRel(pVM, cbRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
2814	AssertLogRelMsgRCBreak(rc, ("cbRange=%zu\n", cbRange));
2815
2816	/*
2817	* Initialize allocation specific items.
2818	*/
2819	//pNew->RamRange.fFlags = 0;
2820	pNew->RamRange.pSelfR0 = MMHyperCCToR0(pVM, &pNew->RamRange);
2821	}
2822
2823	/*
2824	* Initialize the registration structure (caller does specific bits).
2825	*/
2826	pNew->pDevInsR3 = pDevIns;
2827	//pNew->pvR3 = NULL;
2828	//pNew->pNext = NULL;
2829	if (iChunk == 0)
2830	pNew->fFlags \|= PGMREGMMIO2RANGE_F_FIRST_CHUNK;
2831	if (iChunk + 1 == cChunks)
2832	pNew->fFlags \|= PGMREGMMIO2RANGE_F_LAST_CHUNK;
2833	if (fFlags & PGMPHYS_MMIO2_FLAGS_TRACK_DIRTY_PAGES)
2834	pNew->fFlags \|= PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES;
2835	pNew->iSubDev = iSubDev;
2836	pNew->iRegion = iRegion;
2837	pNew->idSavedState = UINT8_MAX;
2838	pNew->idMmio2 = idMmio2;
2839	//pNew->pPhysHandlerR3 = NULL;
2840	//pNew->paLSPages = NULL;
2841	pNew->RamRange.GCPhys = NIL_RTGCPHYS;
2842	pNew->RamRange.GCPhysLast = NIL_RTGCPHYS;
2843	pNew->RamRange.pszDesc = pszDesc;
2844	pNew->RamRange.cb = pNew->cbReal = (RTGCPHYS)cPagesTrackedByChunk << X86_PAGE_SHIFT;
2845	pNew->RamRange.fFlags \|= PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO_EX;
2846	pNew->RamRange.uNemRange = UINT32_MAX;
2847	//pNew->RamRange.pvR3 = NULL;
2848	//pNew->RamRange.paLSPages = NULL;
2849
2850	*ppNext = pNew;
2851	ASMCompilerBarrier();
2852	cPagesLeft -= cPagesTrackedByChunk;
2853	ppNext = &pNew->pNextR3;
2854
2855	/*
2856	* Pre-allocate a handler if we're tracking dirty pages, unless NEM takes care of this.
2857	*/
2858	if ( (fFlags & PGMPHYS_MMIO2_FLAGS_TRACK_DIRTY_PAGES)
2859	#ifdef VBOX_WITH_PGM_NEM_MODE
2860	&& (!VM_IS_NEM_ENABLED(pVM) \|\| !NEMR3IsMmio2DirtyPageTrackingSupported(pVM))
2861	#endif
2862	)
2863
2864	{
2865	rc = pgmHandlerPhysicalExCreate(pVM, pVM->pgm.s.hMmio2DirtyPhysHandlerType,
2866	(RTR3PTR)(uintptr_t)idMmio2, idMmio2, idMmio2, pszDesc, &pNew->pPhysHandlerR3);
2867	AssertLogRelMsgRCBreak(rc, ("idMmio2=%zu\n", idMmio2));
2868	}
2869	}
2870	Assert(cPagesLeft == 0);
2871
2872	if (RT_SUCCESS(rc))
2873	{
2874	Assert((*ppHeadRet)->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
2875	return VINF_SUCCESS;
2876	}
2877
2878	/*
2879	* Free floating ranges.
2880	*/
2881	while (*ppHeadRet)
2882	{
2883	PPGMREGMMIO2RANGE pFree = *ppHeadRet;
2884	*ppHeadRet = pFree->pNextR3;
2885
2886	if (pFree->pPhysHandlerR3)
2887	{
2888	pgmHandlerPhysicalExDestroy(pVM, pFree->pPhysHandlerR3);
2889	pFree->pPhysHandlerR3 = NULL;
2890	}
2891
2892	if (pFree->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
2893	{
2894	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[pFree->RamRange.cb >> X86_PAGE_SHIFT]);
2895	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2896	SUPR3PageFreeEx(pFree, cChunkPages);
2897	}
2898	}
2899
2900	return rc;
2901	}
2902
2903
2904	/**
2905	* Common worker PGMR3PhysMmio2PreRegister & PGMR3PhysMMIO2Register that links a
2906	* complete registration entry into the lists and lookup tables.
2907	*
2908	* @param pVM The cross context VM structure.
2909	* @param pNew The new MMIO / MMIO2 registration to link.
2910	*/
2911	static void pgmR3PhysMmio2Link(PVM pVM, PPGMREGMMIO2RANGE pNew)
2912	{
2913	Assert(pNew->idMmio2 != UINT8_MAX);
2914
2915	/*
2916	* Link it into the list (order doesn't matter, so insert it at the head).
2917	*
2918	* Note! The range we're linking may consist of multiple chunks, so we
2919	* have to find the last one.
2920	*/
2921	PPGMREGMMIO2RANGE pLast = pNew;
2922	for (pLast = pNew; ; pLast = pLast->pNextR3)
2923	{
2924	if (pLast->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2925	break;
2926	Assert(pLast->pNextR3);
2927	Assert(pLast->pNextR3->pDevInsR3 == pNew->pDevInsR3);
2928	Assert(pLast->pNextR3->iSubDev == pNew->iSubDev);
2929	Assert(pLast->pNextR3->iRegion == pNew->iRegion);
2930	Assert(pLast->pNextR3->idMmio2 == pLast->idMmio2 + 1);
2931	}
2932
2933	PGM_LOCK_VOID(pVM);
2934
2935	/* Link in the chain of ranges at the head of the list. */
2936	pLast->pNextR3 = pVM->pgm.s.pRegMmioRangesR3;
2937	pVM->pgm.s.pRegMmioRangesR3 = pNew;
2938
2939	/* Insert the MMIO2 range/page IDs. */
2940	uint8_t idMmio2 = pNew->idMmio2;
2941	for (;;)
2942	{
2943	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == NULL);
2944	Assert(pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] == NIL_RTR0PTR);
2945	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = pNew;
2946	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = pNew->RamRange.pSelfR0 - RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange);
2947	if (pNew->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2948	break;
2949	pNew = pNew->pNextR3;
2950	idMmio2++;
2951	}
2952
2953	pgmPhysInvalidatePageMapTLB(pVM);
2954	PGM_UNLOCK(pVM);
2955	}
2956
2957
2958	/**
2959	* Allocate and register an MMIO2 region.
2960	*
2961	* As mentioned elsewhere, MMIO2 is just RAM spelled differently. It's RAM
2962	* associated with a device. It is also non-shared memory with a permanent
2963	* ring-3 mapping and page backing (presently).
2964	*
2965	* A MMIO2 range may overlap with base memory if a lot of RAM is configured for
2966	* the VM, in which case we'll drop the base memory pages. Presently we will
2967	* make no attempt to preserve anything that happens to be present in the base
2968	* memory that is replaced, this is of course incorrect but it's too much
2969	* effort.
2970	*
2971	* @returns VBox status code.
2972	* @retval VINF_SUCCESS on success, *ppv pointing to the R3 mapping of the
2973	* memory.
2974	* @retval VERR_ALREADY_EXISTS if the region already exists.
2975	*
2976	* @param pVM The cross context VM structure.
2977	* @param pDevIns The device instance owning the region.
2978	* @param iSubDev The sub-device number.
2979	* @param iRegion The region number. If the MMIO2 memory is a PCI
2980	* I/O region this number has to be the number of that
2981	* region. Otherwise it can be any number save
2982	* UINT8_MAX.
2983	* @param cb The size of the region. Must be page aligned.
2984	* @param fFlags Reserved for future use, must be zero.
2985	* @param pszDesc The description.
2986	* @param ppv Where to store the pointer to the ring-3 mapping of
2987	* the memory.
2988	* @param phRegion Where to return the MMIO2 region handle. Optional.
2989	* @thread EMT
2990	*/
2991	VMMR3_INT_DECL(int) PGMR3PhysMmio2Register(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb,
2992	uint32_t fFlags, const char pszDesc, void ppv, PGMMMIO2HANDLE phRegion)
2993	{
2994	/*
2995	* Validate input.
2996	*/
2997	AssertPtrReturn(ppv, VERR_INVALID_POINTER);
2998	*ppv = NULL;
2999	if (phRegion)
3000	{
3001	AssertPtrReturn(phRegion, VERR_INVALID_POINTER);
3002	*phRegion = NIL_PGMMMIO2HANDLE;
3003	}
3004	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3005	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3006	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3007	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3008	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
3009	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
3010	AssertReturn(pgmR3PhysMmio2Find(pVM, pDevIns, iSubDev, iRegion, NIL_PGMMMIO2HANDLE) == NULL, VERR_ALREADY_EXISTS);
3011	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3012	AssertReturn(cb, VERR_INVALID_PARAMETER);
3013	AssertReturn(!(fFlags & ~PGMPHYS_MMIO2_FLAGS_VALID_MASK), VERR_INVALID_FLAGS);
3014
3015	const uint32_t cPages = cb >> PAGE_SHIFT;
3016	AssertLogRelReturn(((RTGCPHYS)cPages << PAGE_SHIFT) == cb, VERR_INVALID_PARAMETER);
3017	AssertLogRelReturn(cPages <= (MM_MMIO_64_MAX >> X86_PAGE_SHIFT), VERR_OUT_OF_RANGE);
3018	AssertLogRelReturn(cPages <= PGM_MMIO2_MAX_PAGE_COUNT, VERR_OUT_OF_RANGE);
3019
3020	/*
3021	* For the 2nd+ instance, mangle the description string so it's unique.
3022	*/
3023	if (pDevIns->iInstance > 0) /** @todo Move to PDMDevHlp.cpp and use a real string cache. */
3024	{
3025	pszDesc = MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s [%u]", pszDesc, pDevIns->iInstance);
3026	if (!pszDesc)
3027	return VERR_NO_MEMORY;
3028	}
3029
3030	/*
3031	* Allocate an MMIO2 range ID (not freed on failure).
3032	*
3033	* The zero ID is not used as it could be confused with NIL_GMM_PAGEID, so
3034	* the IDs goes from 1 thru PGM_MMIO2_MAX_RANGES.
3035	*/
3036	unsigned cChunks = pgmR3PhysMmio2CalcChunkCount(pVM, cb, NULL, NULL);
3037
3038	PGM_LOCK_VOID(pVM);
3039	AssertCompile(PGM_MMIO2_MAX_RANGES < 255);
3040	uint8_t const idMmio2 = pVM->pgm.s.cMmio2Regions + 1;
3041	unsigned const cNewMmio2Regions = pVM->pgm.s.cMmio2Regions + cChunks;
3042	if (cNewMmio2Regions > PGM_MMIO2_MAX_RANGES)
3043	{
3044	PGM_UNLOCK(pVM);
3045	AssertLogRelFailedReturn(VERR_PGM_TOO_MANY_MMIO2_RANGES);
3046	}
3047	pVM->pgm.s.cMmio2Regions = cNewMmio2Regions;
3048	PGM_UNLOCK(pVM);
3049
3050	/*
3051	* Try reserve and allocate the backing memory first as this is what is
3052	* most likely to fail.
3053	*/
3054	int rc = MMR3AdjustFixedReservation(pVM, cPages, pszDesc);
3055	if (RT_SUCCESS(rc))
3056	{
3057	PSUPPAGE paPages = (PSUPPAGE)RTMemTmpAlloc(cPages * sizeof(SUPPAGE));
3058	if (RT_SUCCESS(rc))
3059	{
3060	void *pvPages = NULL;
3061	#ifndef VBOX_WITH_LINEAR_HOST_PHYS_MEM
3062	RTR0PTR pvPagesR0 = NIL_RTR0PTR;
3063	#endif
3064	#ifdef VBOX_WITH_PGM_NEM_MODE
3065	if (PGM_IS_IN_NEM_MODE(pVM))
3066	{
3067	for (uint32_t i = 0; i < cPages; i++)
3068	{
3069	paPages[i].Phys = UINT64_C(0x0000fffffffff000);
3070	paPages[i].uReserved = 0;
3071	}
3072	rc = SUPR3PageAlloc(cPages, pVM->pgm.s.fUseLargePages ? SUP_PAGE_ALLOC_F_LARGE_PAGES : 0, &pvPages);
3073	}
3074	else
3075	#endif
3076	{
3077	#ifndef VBOX_WITH_LINEAR_HOST_PHYS_MEM
3078	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, &pvPagesR0, paPages);
3079	#else
3080	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, NULL /pR0Ptr/, paPages);
3081	#endif
3082	}
3083	if (RT_SUCCESS(rc))
3084	{
3085	memset(pvPages, 0, cPages * PAGE_SIZE);
3086
3087	/*
3088	* Create the registered MMIO range record for it.
3089	*/
3090	PPGMREGMMIO2RANGE pNew;
3091	rc = pgmR3PhysMmio2Create(pVM, pDevIns, iSubDev, iRegion, cb, fFlags, idMmio2, pszDesc, &pNew);
3092	if (RT_SUCCESS(rc))
3093	{
3094	if (phRegion)
3095	phRegion = idMmio2; / The ID of the first chunk. */
3096
3097	uint32_t iSrcPage = 0;
3098	uint8_t pbCurPages = (uint8_t )pvPages;
3099	for (PPGMREGMMIO2RANGE pCur = pNew; pCur; pCur = pCur->pNextR3)
3100	{
3101	pCur->pvR3 = pbCurPages;
3102	#ifndef VBOX_WITH_LINEAR_HOST_PHYS_MEM
3103	pCur->pvR0 = pvPagesR0 + (iSrcPage << PAGE_SHIFT);
3104	#endif
3105	pCur->RamRange.pvR3 = pbCurPages;
3106
3107	uint32_t iDstPage = pCur->RamRange.cb >> X86_PAGE_SHIFT;
3108	while (iDstPage-- > 0)
3109	{
3110	PGM_PAGE_INIT(&pNew->RamRange.aPages[iDstPage],
3111	paPages[iDstPage + iSrcPage].Phys,
3112	PGM_MMIO2_PAGEID_MAKE(idMmio2, iDstPage),
3113	PGMPAGETYPE_MMIO2, PGM_PAGE_STATE_ALLOCATED);
3114	}
3115
3116	/* advance. */
3117	iSrcPage += pCur->RamRange.cb >> X86_PAGE_SHIFT;
3118	pbCurPages += pCur->RamRange.cb;
3119	}
3120
3121	RTMemTmpFree(paPages);
3122
3123	/*
3124	* Update the page count stats, link the registration and we're done.
3125	*/
3126	pVM->pgm.s.cAllPages += cPages;
3127	pVM->pgm.s.cPrivatePages += cPages;
3128
3129	pgmR3PhysMmio2Link(pVM, pNew);
3130
3131	*ppv = pvPages;
3132	return VINF_SUCCESS;
3133	}
3134
3135	SUPR3PageFreeEx(pvPages, cPages);
3136	}
3137	}
3138	RTMemTmpFree(paPages);
3139	MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pszDesc);
3140	}
3141	if (pDevIns->iInstance > 0)
3142	MMR3HeapFree((void *)pszDesc);
3143	return rc;
3144	}
3145
3146
3147	/**
3148	* Deregisters and frees an MMIO2 region.
3149	*
3150	* Any physical access handlers registered for the region must be deregistered
3151	* before calling this function.
3152	*
3153	* @returns VBox status code.
3154	* @param pVM The cross context VM structure.
3155	* @param pDevIns The device instance owning the region.
3156	* @param hMmio2 The MMIO2 handle to deregister, or NIL if all
3157	* regions for the given device is to be deregistered.
3158	*/
3159	VMMR3_INT_DECL(int) PGMR3PhysMmio2Deregister(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3160	{
3161	/*
3162	* Validate input.
3163	*/
3164	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3165	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3166
3167	/*
3168	* The loop here scanning all registrations will make sure that multi-chunk ranges
3169	* get properly deregistered, though it's original purpose was the wildcard iRegion.
3170	*/
3171	PGM_LOCK_VOID(pVM);
3172	int rc = VINF_SUCCESS;
3173	unsigned cFound = 0;
3174	PPGMREGMMIO2RANGE pPrev = NULL;
3175	PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3;
3176	while (pCur)
3177	{
3178	uint32_t const fFlags = pCur->fFlags;
3179	if ( pCur->pDevInsR3 == pDevIns
3180	&& ( hMmio2 == NIL_PGMMMIO2HANDLE
3181	\|\| pCur->idMmio2 == hMmio2))
3182	{
3183	cFound++;
3184
3185	/*
3186	* Unmap it if it's mapped.
3187	*/
3188	if (fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3189	{
3190	int rc2 = PGMR3PhysMmio2Unmap(pVM, pCur->pDevInsR3, pCur->idMmio2, pCur->RamRange.GCPhys);
3191	AssertRC(rc2);
3192	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3193	rc = rc2;
3194	}
3195
3196	/*
3197	* Unlink it
3198	*/
3199	PPGMREGMMIO2RANGE pNext = pCur->pNextR3;
3200	if (pPrev)
3201	pPrev->pNextR3 = pNext;
3202	else
3203	pVM->pgm.s.pRegMmioRangesR3 = pNext;
3204	pCur->pNextR3 = NULL;
3205
3206	uint8_t idMmio2 = pCur->idMmio2;
3207	Assert(idMmio2 <= RT_ELEMENTS(pVM->pgm.s.apMmio2RangesR3));
3208	if (idMmio2 <= RT_ELEMENTS(pVM->pgm.s.apMmio2RangesR3))
3209	{
3210	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == pCur);
3211	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = NULL;
3212	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = NIL_RTR0PTR;
3213	}
3214
3215	/*
3216	* Free the memory.
3217	*/
3218	uint32_t const cPages = pCur->cbReal >> PAGE_SHIFT;
3219	#ifdef VBOX_WITH_PGM_NEM_MODE
3220	if (!pVM->pgm.s.fNemMode)
3221	#endif
3222	{
3223	int rc2 = SUPR3PageFreeEx(pCur->pvR3, cPages);
3224	AssertRC(rc2);
3225	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3226	rc = rc2;
3227
3228	rc2 = MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pCur->RamRange.pszDesc);
3229	AssertRC(rc2);
3230	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3231	rc = rc2;
3232	}
3233	#ifdef VBOX_WITH_PGM_NEM_MODE
3234	else
3235	{
3236	int rc2 = SUPR3PageFree(pCur->pvR3, cPages);
3237	AssertRC(rc2);
3238	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3239	rc = rc2;
3240	}
3241	#endif
3242
3243	if (pCur->pPhysHandlerR3)
3244	{
3245	pgmHandlerPhysicalExDestroy(pVM, pCur->pPhysHandlerR3);
3246	pCur->pPhysHandlerR3 = NULL;
3247	}
3248
3249	/* we're leaking hyper memory here if done at runtime. */
3250	#ifdef VBOX_STRICT
3251	VMSTATE const enmState = VMR3GetState(pVM);
3252	AssertMsg( enmState == VMSTATE_POWERING_OFF
3253	\|\| enmState == VMSTATE_POWERING_OFF_LS
3254	\|\| enmState == VMSTATE_OFF
3255	\|\| enmState == VMSTATE_OFF_LS
3256	\|\| enmState == VMSTATE_DESTROYING
3257	\|\| enmState == VMSTATE_TERMINATED
3258	\|\| enmState == VMSTATE_CREATING
3259	, ("%s\n", VMR3GetStateName(enmState)));
3260	#endif
3261
3262	if (pCur->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
3263	{
3264	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPages]);
3265	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
3266	SUPR3PageFreeEx(pCur, cChunkPages);
3267	}
3268	/*else
3269	{
3270	rc = MMHyperFree(pVM, pCur); - does not work, see the alloc call.
3271	AssertRCReturn(rc, rc);
3272	} */
3273
3274
3275	/* update page count stats */
3276	pVM->pgm.s.cAllPages -= cPages;
3277	pVM->pgm.s.cPrivatePages -= cPages;
3278
3279	/* next */
3280	pCur = pNext;
3281	if (hMmio2 != NIL_PGMMMIO2HANDLE)
3282	{
3283	if (fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3284	break;
3285	hMmio2++;
3286	Assert(pCur->idMmio2 == hMmio2);
3287	Assert(pCur->pDevInsR3 == pDevIns);
3288	Assert(!(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK));
3289	}
3290	}
3291	else
3292	{
3293	pPrev = pCur;
3294	pCur = pCur->pNextR3;
3295	}
3296	}
3297	pgmPhysInvalidatePageMapTLB(pVM);
3298	PGM_UNLOCK(pVM);
3299	return !cFound && hMmio2 != NIL_PGMMMIO2HANDLE ? VERR_NOT_FOUND : rc;
3300	}
3301
3302
3303	/**
3304	* Maps a MMIO2 region.
3305	*
3306	* This is typically done when a guest / the bios / state loading changes the
3307	* PCI config. The replacing of base memory has the same restrictions as during
3308	* registration, of course.
3309	*
3310	* @returns VBox status code.
3311	*
3312	* @param pVM The cross context VM structure.
3313	* @param pDevIns The device instance owning the region.
3314	* @param hMmio2 The handle of the region to map.
3315	* @param GCPhys The guest-physical address to be remapped.
3316	*/
3317	VMMR3_INT_DECL(int) PGMR3PhysMmio2Map(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS GCPhys)
3318	{
3319	/*
3320	* Validate input.
3321	*
3322	* Note! It's safe to walk the MMIO/MMIO2 list since registrations only
3323	* happens during VM construction.
3324	*/
3325	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3326	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3327	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
3328	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3329	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3330	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3331
3332	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3333	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3334	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3335
3336	PPGMREGMMIO2RANGE pLastMmio = pFirstMmio;
3337	RTGCPHYS cbRange = 0;
3338	for (;;)
3339	{
3340	AssertReturn(!(pLastMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED), VERR_WRONG_ORDER);
3341	Assert(pLastMmio->RamRange.GCPhys == NIL_RTGCPHYS);
3342	Assert(pLastMmio->RamRange.GCPhysLast == NIL_RTGCPHYS);
3343	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3344	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3345	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3346	cbRange += pLastMmio->RamRange.cb;
3347	if (pLastMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3348	break;
3349	pLastMmio = pLastMmio->pNextR3;
3350	}
3351
3352	RTGCPHYS GCPhysLast = GCPhys + cbRange - 1;
3353	AssertLogRelReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3354
3355	/*
3356	* Find our location in the ram range list, checking for restriction
3357	* we don't bother implementing yet (partially overlapping, multiple
3358	* ram ranges).
3359	*/
3360	PGM_LOCK_VOID(pVM);
3361
3362	AssertReturnStmt(!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED), PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3363
3364	bool fRamExists = false;
3365	PPGMRAMRANGE pRamPrev = NULL;
3366	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3367	while (pRam && GCPhysLast >= pRam->GCPhys)
3368	{
3369	if ( GCPhys <= pRam->GCPhysLast
3370	&& GCPhysLast >= pRam->GCPhys)
3371	{
3372	/* Completely within? */
3373	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
3374	&& GCPhysLast <= pRam->GCPhysLast,
3375	("%RGp-%RGp (MMIOEx/%s) falls partly outside %RGp-%RGp (%s)\n",
3376	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc,
3377	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3378	PGM_UNLOCK(pVM),
3379	VERR_PGM_RAM_CONFLICT);
3380
3381	/* Check that all the pages are RAM pages. */
3382	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3383	uint32_t cPagesLeft = cbRange >> PAGE_SHIFT;
3384	while (cPagesLeft-- > 0)
3385	{
3386	AssertLogRelMsgReturnStmt(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
3387	("%RGp isn't a RAM page (%d) - mapping %RGp-%RGp (MMIO2/%s).\n",
3388	GCPhys, PGM_PAGE_GET_TYPE(pPage), GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc),
3389	PGM_UNLOCK(pVM),
3390	VERR_PGM_RAM_CONFLICT);
3391	pPage++;
3392	}
3393
3394	/* There can only be one MMIO/MMIO2 chunk matching here! */
3395	AssertLogRelMsgReturnStmt(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK,
3396	("%RGp-%RGp (MMIOEx/%s, flags %#X) consists of multiple chunks whereas the RAM somehow doesn't!\n",
3397	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3398	PGM_UNLOCK(pVM),
3399	VERR_PGM_PHYS_MMIO_EX_IPE);
3400
3401	fRamExists = true;
3402	break;
3403	}
3404
3405	/* next */
3406	pRamPrev = pRam;
3407	pRam = pRam->pNextR3;
3408	}
3409	Log(("PGMR3PhysMmio2Map: %RGp-%RGp fRamExists=%RTbool %s\n", GCPhys, GCPhysLast, fRamExists, pFirstMmio->RamRange.pszDesc));
3410
3411
3412	/*
3413	* Make the changes.
3414	*/
3415	RTGCPHYS GCPhysCur = GCPhys;
3416	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3417	{
3418	pCurMmio->RamRange.GCPhys = GCPhysCur;
3419	pCurMmio->RamRange.GCPhysLast = GCPhysCur + pCurMmio->RamRange.cb - 1;
3420	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3421	{
3422	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3423	break;
3424	}
3425	GCPhysCur += pCurMmio->RamRange.cb;
3426	}
3427
3428	if (fRamExists)
3429	{
3430	/*
3431	* Make all the pages in the range MMIO/ZERO pages, freeing any
3432	* RAM pages currently mapped here. This might not be 100% correct
3433	* for PCI memory, but we're doing the same thing for MMIO2 pages.
3434	*
3435	* We replace these MMIO/ZERO pages with real pages in the MMIO2 case.
3436	*/
3437	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK); /* Only one chunk */
3438	Assert(pFirstMmio->pvR3 == pFirstMmio->RamRange.pvR3);
3439	Assert(pFirstMmio->RamRange.pvR3 != NULL);
3440
3441	#ifdef VBOX_WITH_PGM_NEM_MODE
3442	/* We cannot mix MMIO2 into a RAM range in simplified memory mode because pRam->pvR3 can't point
3443	both at the RAM and MMIO2, so we won't ever write & read from the actual MMIO2 memory if we try. */
3444	AssertLogRelMsgReturn(!pVM->pgm.s.fNemMode, ("%s at %RGp-%RGp\n", pFirstMmio->RamRange.pszDesc, GCPhys, GCPhysLast),
3445	VERR_PGM_NOT_SUPPORTED_FOR_NEM_MODE);
3446	#endif
3447
3448	int rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, pFirstMmio->RamRange.pvR3);
3449	AssertRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
3450
3451	/* Replace the pages, freeing all present RAM pages. */
3452	PPGMPAGE pPageSrc = &pFirstMmio->RamRange.aPages[0];
3453	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3454	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3455	while (cPagesLeft-- > 0)
3456	{
3457	Assert(PGM_PAGE_IS_MMIO(pPageDst));
3458
3459	RTHCPHYS const HCPhys = PGM_PAGE_GET_HCPHYS(pPageSrc);
3460	uint32_t const idPage = PGM_PAGE_GET_PAGEID(pPageSrc);
3461	PGM_PAGE_SET_PAGEID(pVM, pPageDst, idPage);
3462	PGM_PAGE_SET_HCPHYS(pVM, pPageDst, HCPhys);
3463	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO2);
3464	PGM_PAGE_SET_STATE(pVM, pPageDst, PGM_PAGE_STATE_ALLOCATED);
3465	PGM_PAGE_SET_PDE_TYPE(pVM, pPageDst, PGM_PAGE_PDE_TYPE_DONTCARE);
3466	PGM_PAGE_SET_PTE_INDEX(pVM, pPageDst, 0);
3467	PGM_PAGE_SET_TRACKING(pVM, pPageDst, 0);
3468	/* NEM state is set by pgmR3PhysFreePageRange. */
3469
3470	pVM->pgm.s.cZeroPages--;
3471	GCPhys += PAGE_SIZE;
3472	pPageSrc++;
3473	pPageDst++;
3474	}
3475
3476	/* Flush physical page map TLB. */
3477	pgmPhysInvalidatePageMapTLB(pVM);
3478
3479	/* Force a PGM pool flush as guest ram references have been changed. */
3480	/** @todo not entirely SMP safe; assuming for now the guest takes care of
3481	* this internally (not touch mapped mmio while changing the mapping). */
3482	PVMCPU pVCpu = VMMGetCpu(pVM);
3483	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3484	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3485	}
3486	else
3487	{
3488	/*
3489	* No RAM range, insert the ones prepared during registration.
3490	*/
3491	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3492	{
3493	#ifdef VBOX_WITH_NATIVE_NEM
3494	/* Tell NEM and get the new NEM state for the pages. */
3495	uint8_t u2NemState = 0;
3496	if (VM_IS_NEM_ENABLED(pVM))
3497	{
3498	int rc = NEMR3NotifyPhysMmioExMapEarly(pVM, pCurMmio->RamRange.GCPhys,
3499	pCurMmio->RamRange.GCPhysLast - pCurMmio->RamRange.GCPhys + 1,
3500	NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2
3501	\| (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES
3502	? NEM_NOTIFY_PHYS_MMIO_EX_F_TRACK_DIRTY_PAGES : 0),
3503	NULL /pvRam/, pCurMmio->RamRange.pvR3,
3504	&u2NemState, &pCurMmio->RamRange.uNemRange);
3505	AssertLogRelRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
3506	}
3507	#endif
3508
3509	/* Clear the tracking data of pages we're going to reactivate. */
3510	PPGMPAGE pPageSrc = &pCurMmio->RamRange.aPages[0];
3511	uint32_t cPagesLeft = pCurMmio->RamRange.cb >> PAGE_SHIFT;
3512	while (cPagesLeft-- > 0)
3513	{
3514	PGM_PAGE_SET_TRACKING(pVM, pPageSrc, 0);
3515	PGM_PAGE_SET_PTE_INDEX(pVM, pPageSrc, 0);
3516	#ifdef VBOX_WITH_NATIVE_NEM
3517	PGM_PAGE_SET_NEM_STATE(pPageSrc, u2NemState);
3518	#endif
3519	pPageSrc++;
3520	}
3521
3522	/* link in the ram range */
3523	pgmR3PhysLinkRamRange(pVM, &pCurMmio->RamRange, pRamPrev);
3524
3525	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3526	{
3527	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3528	break;
3529	}
3530	pRamPrev = &pCurMmio->RamRange;
3531	}
3532	}
3533
3534	/*
3535	* If the range have dirty page monitoring enabled, enable that.
3536	*
3537	* We ignore failures here for now because if we fail, the whole mapping
3538	* will have to be reversed and we'll end up with nothing at all on the
3539	* screen and a grumpy guest, whereas if we just go on, we'll only have
3540	* visual distortions to gripe about. There will be something in the
3541	* release log.
3542	*/
3543	if ( pFirstMmio->pPhysHandlerR3
3544	&& (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
3545	pgmR3PhysMmio2EnableDirtyPageTracing(pVM, pFirstMmio);
3546
3547	/*
3548	* We're good, set the flags and invalid the mapping TLB.
3549	*/
3550	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3551	{
3552	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_MAPPED;
3553	if (fRamExists)
3554	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_OVERLAPPING;
3555	else
3556	pCurMmio->fFlags &= ~PGMREGMMIO2RANGE_F_OVERLAPPING;
3557	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3558	break;
3559	}
3560	pgmPhysInvalidatePageMapTLB(pVM);
3561
3562	#ifdef VBOX_WITH_NATIVE_NEM
3563	/*
3564	* Late NEM notification.
3565	*/
3566	if (VM_IS_NEM_ENABLED(pVM))
3567	{
3568	int rc;
3569	uint32_t fNemFlags = NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2;
3570	if (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES)
3571	fNemFlags \|= NEM_NOTIFY_PHYS_MMIO_EX_F_TRACK_DIRTY_PAGES;
3572	if (fRamExists)
3573	rc = NEMR3NotifyPhysMmioExMapLate(pVM, GCPhys, GCPhysLast - GCPhys + 1, fNemFlags \| NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE,
3574	pRam->pvR3 ? (uint8_t *)pRam->pvR3 + GCPhys - pRam->GCPhys : NULL, pFirstMmio->pvR3,
3575	NULL /puNemRange/);
3576	else
3577	{
3578	rc = VINF_SUCCESS;
3579	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3580	{
3581	rc = NEMR3NotifyPhysMmioExMapLate(pVM, pCurMmio->RamRange.GCPhys, pCurMmio->RamRange.cb, fNemFlags,
3582	NULL, pCurMmio->RamRange.pvR3, &pCurMmio->RamRange.uNemRange);
3583	if ((pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK) \|\| RT_FAILURE(rc))
3584	break;
3585	}
3586	}
3587	AssertLogRelRCReturnStmt(rc, PGMR3PhysMmio2Unmap(pVM, pDevIns, hMmio2, GCPhys); PGM_UNLOCK(pVM), rc);
3588	}
3589	#endif
3590
3591	PGM_UNLOCK(pVM);
3592
3593	return VINF_SUCCESS;
3594	}
3595
3596
3597	/**
3598	* Unmaps an MMIO2 region.
3599	*
3600	* This is typically done when a guest / the bios / state loading changes the
3601	* PCI config. The replacing of base memory has the same restrictions as during
3602	* registration, of course.
3603	*/
3604	VMMR3_INT_DECL(int) PGMR3PhysMmio2Unmap(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS GCPhys)
3605	{
3606	/*
3607	* Validate input
3608	*/
3609	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3610	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3611	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3612	if (GCPhys != NIL_RTGCPHYS)
3613	{
3614	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3615	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3616	}
3617
3618	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3619	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3620	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3621
3622	int rc = PGM_LOCK(pVM);
3623	AssertRCReturn(rc, rc);
3624
3625	PPGMREGMMIO2RANGE pLastMmio = pFirstMmio;
3626	RTGCPHYS cbRange = 0;
3627	for (;;)
3628	{
3629	AssertReturnStmt(pLastMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED, PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3630	AssertReturnStmt(pLastMmio->RamRange.GCPhys == GCPhys + cbRange \|\| GCPhys == NIL_RTGCPHYS, PGM_UNLOCK(pVM), VERR_INVALID_PARAMETER);
3631	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3632	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3633	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3634	cbRange += pLastMmio->RamRange.cb;
3635	if (pLastMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3636	break;
3637	pLastMmio = pLastMmio->pNextR3;
3638	}
3639
3640	Log(("PGMR3PhysMmio2Unmap: %RGp-%RGp %s\n",
3641	pFirstMmio->RamRange.GCPhys, pLastMmio->RamRange.GCPhysLast, pFirstMmio->RamRange.pszDesc));
3642
3643	uint16_t const fOldFlags = pFirstMmio->fFlags;
3644	AssertReturnStmt(fOldFlags & PGMREGMMIO2RANGE_F_MAPPED, PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3645
3646	/*
3647	* If monitoring dirty pages, we must deregister the handlers first.
3648	*/
3649	if ( pFirstMmio->pPhysHandlerR3
3650	&& (fOldFlags & PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
3651	pgmR3PhysMmio2DisableDirtyPageTracing(pVM, pFirstMmio);
3652
3653	/*
3654	* Unmap it.
3655	*/
3656	int rcRet = VINF_SUCCESS;
3657	#ifdef VBOX_WITH_NATIVE_NEM
3658	uint32_t const fNemFlags = NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2
3659	\| (fOldFlags & PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES
3660	? NEM_NOTIFY_PHYS_MMIO_EX_F_TRACK_DIRTY_PAGES : 0);
3661	#endif
3662	if (fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING)
3663	{
3664	/*
3665	* We've replaced RAM, replace with zero pages.
3666	*
3667	* Note! This is where we might differ a little from a real system, because
3668	* it's likely to just show the RAM pages as they were before the
3669	* MMIO/MMIO2 region was mapped here.
3670	*/
3671	/* Only one chunk allowed when overlapping! */
3672	Assert(fOldFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK);
3673
3674	/* Restore the RAM pages we've replaced. */
3675	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3676	while (pRam->GCPhys > pFirstMmio->RamRange.GCPhysLast)
3677	pRam = pRam->pNextR3;
3678
3679	PPGMPAGE pPageDst = &pRam->aPages[(pFirstMmio->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3680	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3681	pVM->pgm.s.cZeroPages += cPagesLeft; /** @todo not correct for NEM mode */
3682
3683	#ifdef VBOX_WITH_NATIVE_NEM
3684	if (VM_IS_NEM_ENABLED(pVM)) /* Notify NEM. Note! we cannot be here in simple memory mode, see mapping function. */
3685	{
3686	uint8_t u2State = UINT8_MAX;
3687	rc = NEMR3NotifyPhysMmioExUnmap(pVM, pFirstMmio->RamRange.GCPhys, pFirstMmio->RamRange.cb,
3688	fNemFlags \| NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE,
3689	pRam->pvR3
3690	? (uint8_t *)pRam->pvR3 + pFirstMmio->RamRange.GCPhys - pRam->GCPhys : NULL,
3691	pFirstMmio->pvR3, &u2State, &pRam->uNemRange);
3692	AssertRCStmt(rc, rcRet = rc);
3693	if (u2State != UINT8_MAX)
3694	pgmPhysSetNemStateForPages(pPageDst, cPagesLeft, u2State);
3695	}
3696	#endif
3697
3698	while (cPagesLeft-- > 0)
3699	{
3700	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
3701	pPageDst++;
3702	}
3703
3704	/* Flush physical page map TLB. */
3705	pgmPhysInvalidatePageMapTLB(pVM);
3706
3707	/* Update range state. */
3708	pFirstMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3709	pFirstMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3710	pFirstMmio->fFlags &= ~(PGMREGMMIO2RANGE_F_OVERLAPPING \| PGMREGMMIO2RANGE_F_MAPPED);
3711	}
3712	else
3713	{
3714	/*
3715	* Unlink the chunks related to the MMIO/MMIO2 region.
3716	*/
3717	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3718	{
3719	#ifdef VBOX_WITH_NATIVE_NEM
3720	if (VM_IS_NEM_ENABLED(pVM)) /* Notify NEM. */
3721	{
3722	uint8_t u2State = UINT8_MAX;
3723	rc = NEMR3NotifyPhysMmioExUnmap(pVM, pCurMmio->RamRange.GCPhys, pCurMmio->RamRange.cb, fNemFlags,
3724	NULL, pCurMmio->pvR3, &u2State, &pCurMmio->RamRange.uNemRange);
3725	AssertRCStmt(rc, rcRet = rc);
3726	if (u2State != UINT8_MAX)
3727	pgmPhysSetNemStateForPages(pCurMmio->RamRange.aPages, pCurMmio->RamRange.cb >> PAGE_SHIFT, u2State);
3728	}
3729	#endif
3730	pgmR3PhysUnlinkRamRange(pVM, &pCurMmio->RamRange);
3731	pCurMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3732	pCurMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3733	pCurMmio->fFlags &= ~(PGMREGMMIO2RANGE_F_OVERLAPPING \| PGMREGMMIO2RANGE_F_MAPPED);
3734	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3735	break;
3736	}
3737	}
3738
3739	/* Force a PGM pool flush as guest ram references have been changed. */
3740	/** @todo not entirely SMP safe; assuming for now the guest takes care
3741	* of this internally (not touch mapped mmio while changing the
3742	* mapping). */
3743	PVMCPU pVCpu = VMMGetCpu(pVM);
3744	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3745	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3746
3747	pgmPhysInvalidatePageMapTLB(pVM);
3748	pgmPhysInvalidRamRangeTlbs(pVM);
3749
3750	PGM_UNLOCK(pVM);
3751	return rcRet;
3752	}
3753
3754
3755	/**
3756	* Reduces the mapping size of a MMIO2 region.
3757	*
3758	* This is mainly for dealing with old saved states after changing the default
3759	* size of a mapping region. See PGMDevHlpMMIOExReduce and
3760	* PDMPCIDEV::pfnRegionLoadChangeHookR3.
3761	*
3762	* The region must not currently be mapped when making this call. The VM state
3763	* must be state restore or VM construction.
3764	*
3765	* @returns VBox status code.
3766	* @param pVM The cross context VM structure.
3767	* @param pDevIns The device instance owning the region.
3768	* @param hMmio2 The handle of the region to reduce.
3769	* @param cbRegion The new mapping size.
3770	*/
3771	VMMR3_INT_DECL(int) PGMR3PhysMmio2Reduce(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS cbRegion)
3772	{
3773	/*
3774	* Validate input
3775	*/
3776	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3777	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3778	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3779	AssertReturn(cbRegion >= X86_PAGE_SIZE, VERR_INVALID_PARAMETER);
3780	AssertReturn(!(cbRegion & X86_PAGE_OFFSET_MASK), VERR_UNSUPPORTED_ALIGNMENT);
3781	VMSTATE enmVmState = VMR3GetState(pVM);
3782	AssertLogRelMsgReturn( enmVmState == VMSTATE_CREATING
3783	\|\| enmVmState == VMSTATE_LOADING,
3784	("enmVmState=%d (%s)\n", enmVmState, VMR3GetStateName(enmVmState)),
3785	VERR_VM_INVALID_VM_STATE);
3786
3787	int rc = PGM_LOCK(pVM);
3788	AssertRCReturn(rc, rc);
3789
3790	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3791	if (pFirstMmio)
3792	{
3793	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3794	if (!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED))
3795	{
3796	/*
3797	* NOTE! Current implementation does not support multiple ranges.
3798	* Implement when there is a real world need and thus a testcase.
3799	*/
3800	AssertLogRelMsgStmt(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK,
3801	("%s: %#x\n", pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3802	rc = VERR_NOT_SUPPORTED);
3803	if (RT_SUCCESS(rc))
3804	{
3805	/*
3806	* Make the change.
3807	*/
3808	Log(("PGMR3PhysMmio2Reduce: %s changes from %RGp bytes (%RGp) to %RGp bytes.\n",
3809	pFirstMmio->RamRange.pszDesc, pFirstMmio->RamRange.cb, pFirstMmio->cbReal, cbRegion));
3810
3811	AssertLogRelMsgStmt(cbRegion <= pFirstMmio->cbReal,
3812	("%s: cbRegion=%#RGp cbReal=%#RGp\n", pFirstMmio->RamRange.pszDesc, cbRegion, pFirstMmio->cbReal),
3813	rc = VERR_OUT_OF_RANGE);
3814	if (RT_SUCCESS(rc))
3815	{
3816	pFirstMmio->RamRange.cb = cbRegion;
3817	}
3818	}
3819	}
3820	else
3821	rc = VERR_WRONG_ORDER;
3822	}
3823	else
3824	rc = VERR_NOT_FOUND;
3825
3826	PGM_UNLOCK(pVM);
3827	return rc;
3828	}
3829
3830
3831	/**
3832	* Validates @a hMmio2, making sure it belongs to @a pDevIns.
3833	*
3834	* @returns VBox status code.
3835	* @param pVM The cross context VM structure.
3836	* @param pDevIns The device which allegedly owns @a hMmio2.
3837	* @param hMmio2 The handle to validate.
3838	*/
3839	VMMR3_INT_DECL(int) PGMR3PhysMmio2ValidateHandle(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3840	{
3841	/*
3842	* Validate input
3843	*/
3844	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3845	AssertPtrReturn(pDevIns, VERR_INVALID_POINTER);
3846
3847	/*
3848	* Just do this the simple way. No need for locking as this is only taken at
3849	*/
3850	PGM_LOCK_VOID(pVM);
3851	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3852	PGM_UNLOCK(pVM);
3853	AssertReturn(pFirstMmio, VERR_INVALID_HANDLE);
3854	AssertReturn(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, VERR_INVALID_HANDLE);
3855	return VINF_SUCCESS;
3856	}
3857
3858
3859	/**
3860	* Gets the mapping address of an MMIO2 region.
3861	*
3862	* @returns Mapping address, NIL_RTGCPHYS if not mapped or invalid handle.
3863	*
3864	* @param pVM The cross context VM structure.
3865	* @param pDevIns The device owning the MMIO2 handle.
3866	* @param hMmio2 The region handle.
3867	*/
3868	VMMR3_INT_DECL(RTGCPHYS) PGMR3PhysMmio2GetMappingAddress(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3869	{
3870	AssertPtrReturn(pDevIns, NIL_RTGCPHYS);
3871
3872	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3873	AssertReturn(pFirstRegMmio, NIL_RTGCPHYS);
3874
3875	if (pFirstRegMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3876	return pFirstRegMmio->RamRange.GCPhys;
3877	return NIL_RTGCPHYS;
3878	}
3879
3880
3881	/**
3882	* Worker for PGMR3PhysMmio2QueryAndResetDirtyBitmap.
3883	*
3884	* Called holding the PGM lock.
3885	*/
3886	static int pgmR3PhysMmio2QueryAndResetDirtyBitmapLocked(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2,
3887	void *pvBitmap, size_t cbBitmap)
3888	{
3889	/*
3890	* Continue validation.
3891	*/
3892	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3893	AssertReturn(pFirstRegMmio, VERR_INVALID_HANDLE);
3894	AssertReturn( (pFirstRegMmio->fFlags & (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK))
3895	== (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK),
3896	VERR_INVALID_FUNCTION);
3897	AssertReturn(pDevIns == pFirstRegMmio->pDevInsR3, VERR_NOT_OWNER);
3898
3899	RTGCPHYS cbTotal = 0;
3900	uint16_t fTotalDirty = 0;
3901	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio;;)
3902	{
3903	cbTotal += pCur->RamRange.cb; /* Not using cbReal here, because NEM is not in on the creating, only the mapping. */
3904	fTotalDirty \|= pCur->fFlags;
3905	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3906	break;
3907	pCur = pCur->pNextR3;
3908	AssertPtrReturn(pCur, VERR_INTERNAL_ERROR_5);
3909	AssertReturn( (pCur->fFlags & (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK))
3910	== PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES,
3911	VERR_INTERNAL_ERROR_4);
3912	}
3913	size_t const cbTotalBitmap = RT_ALIGN_T(cbTotal, PAGE_SIZE * 64, RTGCPHYS) / PAGE_SIZE / 8;
3914
3915	if (cbBitmap)
3916	{
3917	AssertPtrReturn(pvBitmap, VERR_INVALID_POINTER);
3918	AssertReturn(RT_ALIGN_P(pvBitmap, sizeof(uint64_t)) == pvBitmap, VERR_INVALID_POINTER);
3919	AssertReturn(cbBitmap == cbTotalBitmap, VERR_INVALID_PARAMETER);
3920	}
3921
3922	/*
3923	* Do the work.
3924	*/
3925	int rc = VINF_SUCCESS;
3926	if (pvBitmap)
3927	{
3928	#ifdef VBOX_WITH_PGM_NEM_MODE
3929	if (pFirstRegMmio->pPhysHandlerR3 == NULL)
3930	{
3931	AssertReturn(VM_IS_NEM_ENABLED(pVM), VERR_INTERNAL_ERROR_4);
3932	uint8_t pbBitmap = (uint8_t )pvBitmap;
3933	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
3934	{
3935	size_t const cbBitmapChunk = pCur->RamRange.cb / PAGE_SIZE / 8;
3936	Assert((RTGCPHYS)cbBitmapChunk * PAGE_SIZE * 8 == pCur->RamRange.cb);
3937	int rc2 = NEMR3PhysMmio2QueryAndResetDirtyBitmap(pVM, pCur->RamRange.GCPhys, pCur->RamRange.cb,
3938	pCur->RamRange.uNemRange, pbBitmap, cbBitmapChunk);
3939	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3940	rc = rc2;
3941	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3942	break;
3943	pbBitmap += pCur->RamRange.cb / PAGE_SIZE / 8;
3944	}
3945	}
3946	else
3947	#endif
3948	if (fTotalDirty & PGMREGMMIO2RANGE_F_IS_DIRTY)
3949	{
3950	if ( (pFirstRegMmio->fFlags & (PGMREGMMIO2RANGE_F_MAPPED \| PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
3951	== (PGMREGMMIO2RANGE_F_MAPPED \| PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
3952	{
3953	/*
3954	* Reset each chunk, gathering dirty bits.
3955	*/
3956	RT_BZERO(pvBitmap, cbBitmap); /* simpler for now. */
3957	uint32_t iPageNo = 0;
3958	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
3959	{
3960	if (pCur->fFlags & PGMREGMMIO2RANGE_F_IS_DIRTY)
3961	{
3962	int rc2 = pgmHandlerPhysicalResetMmio2WithBitmap(pVM, pCur->RamRange.GCPhys, pvBitmap, iPageNo);
3963	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3964	rc = rc2;
3965	pCur->fFlags &= ~PGMREGMMIO2RANGE_F_IS_DIRTY;
3966	}
3967	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3968	break;
3969	iPageNo += pCur->RamRange.cb >> PAGE_SHIFT;
3970	}
3971	}
3972	else
3973	{
3974	/*
3975	* If not mapped or tracking is disabled, we return the
3976	* PGMREGMMIO2RANGE_F_IS_DIRTY status for all pages. We cannot
3977	* get more accurate data than that after unmapping or disabling.
3978	*/
3979	RT_BZERO(pvBitmap, cbBitmap);
3980	uint32_t iPageNo = 0;
3981	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
3982	{
3983	if (pCur->fFlags & PGMREGMMIO2RANGE_F_IS_DIRTY)
3984	{
3985	ASMBitSetRange(pvBitmap, iPageNo, iPageNo + (pCur->RamRange.cb >> PAGE_SHIFT));
3986	pCur->fFlags &= ~PGMREGMMIO2RANGE_F_IS_DIRTY;
3987	}
3988	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3989	break;
3990	iPageNo += pCur->RamRange.cb >> PAGE_SHIFT;
3991	}
3992	}
3993	}
3994	/*
3995	* No dirty chunks.
3996	*/
3997	else
3998	RT_BZERO(pvBitmap, cbBitmap);
3999	}
4000	/*
4001	* No bitmap. Reset the region if tracking is currently enabled.
4002	*/
4003	else if ( (pFirstRegMmio->fFlags & (PGMREGMMIO2RANGE_F_MAPPED \| PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
4004	== (PGMREGMMIO2RANGE_F_MAPPED \| PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
4005	{
4006	#ifdef VBOX_WITH_PGM_NEM_MODE
4007	if (pFirstRegMmio->pPhysHandlerR3 == NULL)
4008	{
4009	AssertReturn(VM_IS_NEM_ENABLED(pVM), VERR_INTERNAL_ERROR_4);
4010	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
4011	{
4012	int rc2 = NEMR3PhysMmio2QueryAndResetDirtyBitmap(pVM, pCur->RamRange.GCPhys, pCur->RamRange.cb,
4013	pCur->RamRange.uNemRange, NULL, 0);
4014	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
4015	rc = rc2;
4016	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
4017	break;
4018	}
4019	}
4020	else
4021	#endif
4022	{
4023	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
4024	{
4025	pCur->fFlags &= ~PGMREGMMIO2RANGE_F_IS_DIRTY;
4026	int rc2 = PGMHandlerPhysicalReset(pVM, pCur->RamRange.GCPhys);
4027	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
4028	rc = rc2;
4029	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
4030	break;
4031	}
4032	}
4033	}
4034
4035	return rc;
4036	}
4037
4038
4039	/**
4040	* Queries the dirty page bitmap and resets the monitoring.
4041	*
4042	* The PGMPHYS_MMIO2_FLAGS_TRACK_DIRTY_PAGES flag must be specified when
4043	* creating the range for this to work.
4044	*
4045	* @returns VBox status code.
4046	* @retval VERR_INVALID_FUNCTION if not created using
4047	* PGMPHYS_MMIO2_FLAGS_TRACK_DIRTY_PAGES.
4048	* @param pVM The cross context VM structure.
4049	* @param pDevIns The device owning the MMIO2 handle.
4050	* @param hMmio2 The region handle.
4051	* @param pvBitmap The output bitmap. Must be 8-byte aligned. Ignored
4052	* when @a cbBitmap is zero.
4053	* @param cbBitmap The size of the bitmap. Must be the size of the whole
4054	* MMIO2 range, rounded up to the nearest 8 bytes.
4055	* When zero only a reset is done.
4056	*/
4057	VMMR3_INT_DECL(int) PGMR3PhysMmio2QueryAndResetDirtyBitmap(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2,
4058	void *pvBitmap, size_t cbBitmap)
4059	{
4060	/*
4061	* Do some basic validation before grapping the PGM lock and continuing.
4062	*/
4063	AssertPtrReturn(pDevIns, VERR_INVALID_POINTER);
4064	AssertReturn(RT_ALIGN_Z(cbBitmap, sizeof(uint64_t)) == cbBitmap, VERR_INVALID_PARAMETER);
4065	int rc = PGM_LOCK(pVM);
4066	if (RT_SUCCESS(rc))
4067	{
4068	STAM_PROFILE_START(&pVM->pgm.s.StatMmio2QueryAndResetDirtyBitmap, a);
4069	rc = pgmR3PhysMmio2QueryAndResetDirtyBitmapLocked(pVM, pDevIns, hMmio2, pvBitmap, cbBitmap);
4070	STAM_PROFILE_STOP(&pVM->pgm.s.StatMmio2QueryAndResetDirtyBitmap, a);
4071	PGM_UNLOCK(pVM);
4072	}
4073	return rc;
4074	}
4075
4076	/**
4077	* Worker for PGMR3PhysMmio2ControlDirtyPageTracking
4078	*
4079	* Called owning the PGM lock.
4080	*/
4081	static int pgmR3PhysMmio2ControlDirtyPageTrackingLocked(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, bool fEnabled)
4082	{
4083	/*
4084	* Continue validation.
4085	*/
4086	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
4087	AssertReturn(pFirstRegMmio, VERR_INVALID_HANDLE);
4088	AssertReturn( (pFirstRegMmio->fFlags & (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK))
4089	== (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK)
4090	, VERR_INVALID_FUNCTION);
4091	AssertReturn(pDevIns == pFirstRegMmio->pDevInsR3, VERR_NOT_OWNER);
4092
4093	#ifdef VBOX_WITH_PGM_NEM_MODE
4094	/*
4095	* This is a nop if NEM is responsible for doing the tracking, we simply
4096	* leave the tracking on all the time there.
4097	*/
4098	if (pFirstRegMmio->pPhysHandlerR3 == NULL)
4099	{
4100	AssertReturn(VM_IS_NEM_ENABLED(pVM), VERR_INTERNAL_ERROR_4);
4101	return VINF_SUCCESS;
4102	}
4103	#endif
4104
4105	/*
4106	* Anyting needing doing?
4107	*/
4108	if (fEnabled != RT_BOOL(pFirstRegMmio->fFlags & PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
4109	{
4110	LogFlowFunc(("fEnabled=%RTbool %s\n", fEnabled, pFirstRegMmio->RamRange.pszDesc));
4111
4112	/*
4113	* Update the PGMREGMMIO2RANGE_F_TRACKING_ENABLED flag.
4114	*/
4115	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio;;)
4116	{
4117	if (fEnabled)
4118	pCur->fFlags \|= PGMREGMMIO2RANGE_F_TRACKING_ENABLED;
4119	else
4120	pCur->fFlags &= ~PGMREGMMIO2RANGE_F_TRACKING_ENABLED;
4121	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
4122	break;
4123	pCur = pCur->pNextR3;
4124	AssertPtrReturn(pCur, VERR_INTERNAL_ERROR_5);
4125	AssertReturn( (pCur->fFlags & (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK))
4126	== PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES
4127	, VERR_INTERNAL_ERROR_4);
4128	}
4129
4130	/*
4131	* Enable/disable handlers if currently mapped.
4132	*
4133	* We ignore status codes here as we've already changed the flags and
4134	* returning a failure status now would be confusing. Besides, the two
4135	* functions will continue past failures. As argued in the mapping code,
4136	* it's in the release log.
4137	*/
4138	if (pFirstRegMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED)
4139	{
4140	if (fEnabled)
4141	pgmR3PhysMmio2EnableDirtyPageTracing(pVM, pFirstRegMmio);
4142	else
4143	pgmR3PhysMmio2DisableDirtyPageTracing(pVM, pFirstRegMmio);
4144	}
4145	}
4146	else
4147	LogFlowFunc(("fEnabled=%RTbool %s - no change\n", fEnabled, pFirstRegMmio->RamRange.pszDesc));
4148
4149	return VINF_SUCCESS;
4150	}
4151
4152
4153	/**
4154	* Controls the dirty page tracking for an MMIO2 range.
4155	*
4156	* @returns VBox status code.
4157	* @param pVM The cross context VM structure.
4158	* @param pDevIns The device owning the MMIO2 memory.
4159	* @param hMmio2 The handle of the region.
4160	* @param fEnabled The new tracking state.
4161	*/
4162	VMMR3_INT_DECL(int) PGMR3PhysMmio2ControlDirtyPageTracking(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, bool fEnabled)
4163	{
4164	/*
4165	* Do some basic validation before grapping the PGM lock and continuing.
4166	*/
4167	AssertPtrReturn(pDevIns, VERR_INVALID_POINTER);
4168	int rc = PGM_LOCK(pVM);
4169	if (RT_SUCCESS(rc))
4170	{
4171	rc = pgmR3PhysMmio2ControlDirtyPageTrackingLocked(pVM, pDevIns, hMmio2, fEnabled);
4172	PGM_UNLOCK(pVM);
4173	}
4174	return rc;
4175	}
4176
4177
4178	/**
4179	* Changes the region number of an MMIO2 region.
4180	*
4181	* This is only for dealing with save state issues, nothing else.
4182	*
4183	* @return VBox status code.
4184	*
4185	* @param pVM The cross context VM structure.
4186	* @param pDevIns The device owning the MMIO2 memory.
4187	* @param hMmio2 The handle of the region.
4188	* @param iNewRegion The new region index.
4189	*
4190	* @thread EMT(0)
4191	* @sa @bugref{9359}
4192	*/
4193	VMMR3_INT_DECL(int) PGMR3PhysMmio2ChangeRegionNo(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, uint32_t iNewRegion)
4194	{
4195	/*
4196	* Validate input.
4197	*/
4198	VM_ASSERT_EMT0_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
4199	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_LOADING, VERR_VM_INVALID_VM_STATE);
4200	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
4201	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
4202	AssertReturn(iNewRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
4203
4204	AssertReturn(pVM->enmVMState == VMSTATE_LOADING, VERR_INVALID_STATE);
4205
4206	int rc = PGM_LOCK(pVM);
4207	AssertRCReturn(rc, rc);
4208
4209	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
4210	AssertReturnStmt(pFirstRegMmio, PGM_UNLOCK(pVM), VERR_NOT_FOUND);
4211	AssertReturnStmt(pgmR3PhysMmio2Find(pVM, pDevIns, pFirstRegMmio->iSubDev, iNewRegion, NIL_PGMMMIO2HANDLE) == NULL,
4212	PGM_UNLOCK(pVM), VERR_RESOURCE_IN_USE);
4213
4214	/*
4215	* Make the change.
4216	*/
4217	pFirstRegMmio->iRegion = (uint8_t)iNewRegion;
4218
4219	PGM_UNLOCK(pVM);
4220	return VINF_SUCCESS;
4221	}
4222
4223
4224
4225	/*********************************************************************************************************************************
4226	* ROM *
4227	*********************************************************************************************************************************/
4228
4229	/**
4230	* Worker for PGMR3PhysRomRegister.
4231	*
4232	* This is here to simplify lock management, i.e. the caller does all the
4233	* locking and we can simply return without needing to remember to unlock
4234	* anything first.
4235	*
4236	* @returns VBox status code.
4237	* @param pVM The cross context VM structure.
4238	* @param pDevIns The device instance owning the ROM.
4239	* @param GCPhys First physical address in the range.
4240	* Must be page aligned!
4241	* @param cb The size of the range (in bytes).
4242	* Must be page aligned!
4243	* @param pvBinary Pointer to the binary data backing the ROM image.
4244	* @param cbBinary The size of the binary data pvBinary points to.
4245	* This must be less or equal to @a cb.
4246	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
4247	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
4248	* @param pszDesc Pointer to description string. This must not be freed.
4249	*/
4250	static int pgmR3PhysRomRegisterLocked(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
4251	const void pvBinary, uint32_t cbBinary, uint8_t fFlags, const char pszDesc)
4252	{
4253	/*
4254	* Validate input.
4255	*/
4256	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
4257	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
4258	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
4259	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
4260	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
4261	AssertPtrReturn(pvBinary, VERR_INVALID_PARAMETER);
4262	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
4263	AssertReturn(!(fFlags & ~PGMPHYS_ROM_FLAGS_VALID_MASK), VERR_INVALID_PARAMETER);
4264	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_CREATING, VERR_VM_INVALID_VM_STATE);
4265
4266	const uint32_t cPages = cb >> PAGE_SHIFT;
4267
4268	/*
4269	* Find the ROM location in the ROM list first.
4270	*/
4271	PPGMROMRANGE pRomPrev = NULL;
4272	PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3;
4273	while (pRom && GCPhysLast >= pRom->GCPhys)
4274	{
4275	if ( GCPhys <= pRom->GCPhysLast
4276	&& GCPhysLast >= pRom->GCPhys)
4277	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
4278	GCPhys, GCPhysLast, pszDesc,
4279	pRom->GCPhys, pRom->GCPhysLast, pRom->pszDesc),
4280	VERR_PGM_RAM_CONFLICT);
4281	/* next */
4282	pRomPrev = pRom;
4283	pRom = pRom->pNextR3;
4284	}
4285
4286	/*
4287	* Find the RAM location and check for conflicts.
4288	*
4289	* Conflict detection is a bit different than for RAM registration since a
4290	* ROM can be located within a RAM range. So, what we have to check for is
4291	* other memory types (other than RAM that is) and that we don't span more
4292	* than one RAM range (lazy).
4293	*/
4294	bool fRamExists = false;
4295	PPGMRAMRANGE pRamPrev = NULL;
4296	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
4297	while (pRam && GCPhysLast >= pRam->GCPhys)
4298	{
4299	if ( GCPhys <= pRam->GCPhysLast
4300	&& GCPhysLast >= pRam->GCPhys)
4301	{
4302	/* completely within? */
4303	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
4304	&& GCPhysLast <= pRam->GCPhysLast,
4305	("%RGp-%RGp (%s) falls partly outside %RGp-%RGp (%s)\n",
4306	GCPhys, GCPhysLast, pszDesc,
4307	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
4308	VERR_PGM_RAM_CONFLICT);
4309	fRamExists = true;
4310	break;
4311	}
4312
4313	/* next */
4314	pRamPrev = pRam;
4315	pRam = pRam->pNextR3;
4316	}
4317	if (fRamExists)
4318	{
4319	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
4320	uint32_t cPagesLeft = cPages;
4321	while (cPagesLeft-- > 0)
4322	{
4323	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
4324	("%RGp (%R[pgmpage]) isn't a RAM page - registering %RGp-%RGp (%s).\n",
4325	pRam->GCPhys + ((RTGCPHYS)(uintptr_t)(pPage - &pRam->aPages[0]) << PAGE_SHIFT),
4326	pPage, GCPhys, GCPhysLast, pszDesc), VERR_PGM_RAM_CONFLICT);
4327	Assert(PGM_PAGE_IS_ZERO(pPage) \|\| PGM_IS_IN_NEM_MODE(pVM));
4328	pPage++;
4329	}
4330	}
4331
4332	/*
4333	* Update the base memory reservation if necessary.
4334	*/
4335	uint32_t cExtraBaseCost = fRamExists ? 0 : cPages;
4336	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4337	cExtraBaseCost += cPages;
4338	if (cExtraBaseCost)
4339	{
4340	int rc = MMR3IncreaseBaseReservation(pVM, cExtraBaseCost);
4341	if (RT_FAILURE(rc))
4342	return rc;
4343	}
4344
4345	#ifdef VBOX_WITH_NATIVE_NEM
4346	/*
4347	* Early NEM notification before we've made any changes or anything.
4348	*/
4349	uint32_t const fNemNotify = (fRamExists ? NEM_NOTIFY_PHYS_ROM_F_REPLACE : 0)
4350	\| (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED ? NEM_NOTIFY_PHYS_ROM_F_SHADOW : 0);
4351	uint8_t u2NemState = UINT8_MAX;
4352	uint32_t uNemRange = 0;
4353	if (VM_IS_NEM_ENABLED(pVM))
4354	{
4355	int rc = NEMR3NotifyPhysRomRegisterEarly(pVM, GCPhys, cPages << PAGE_SHIFT,
4356	fRamExists ? PGM_RAMRANGE_CALC_PAGE_R3PTR(pRam, GCPhys) : NULL,
4357	fNemNotify, &u2NemState, fRamExists ? &pRam->uNemRange : &uNemRange);
4358	AssertLogRelRCReturn(rc, rc);
4359	}
4360	#endif
4361
4362	/*
4363	* Allocate memory for the virgin copy of the RAM. In simplified memory mode,
4364	* we allocate memory for any ad-hoc RAM range and for shadow pages.
4365	*/
4366	PGMMALLOCATEPAGESREQ pReq = NULL;
4367	#ifdef VBOX_WITH_PGM_NEM_MODE
4368	void *pvRam = NULL;
4369	void *pvAlt = NULL;
4370	if (pVM->pgm.s.fNemMode)
4371	{
4372	if (!fRamExists)
4373	{
4374	int rc = SUPR3PageAlloc(cPages, 0, &pvRam);
4375	if (RT_FAILURE(rc))
4376	return rc;
4377	}
4378	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4379	{
4380	int rc = SUPR3PageAlloc(cPages, 0, &pvAlt);
4381	if (RT_FAILURE(rc))
4382	{
4383	if (pvRam)
4384	SUPR3PageFree(pvRam, cPages);
4385	return rc;
4386	}
4387	}
4388	}
4389	else
4390	#endif
4391	{
4392	int rc = GMMR3AllocatePagesPrepare(pVM, &pReq, cPages, GMMACCOUNT_BASE);
4393	AssertRCReturn(rc, rc);
4394
4395	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4396	{
4397	pReq->aPages[iPage].HCPhysGCPhys = GCPhys + (iPage << PAGE_SHIFT);
4398	pReq->aPages[iPage].fZeroed = false;
4399	pReq->aPages[iPage].idPage = NIL_GMM_PAGEID;
4400	pReq->aPages[iPage].idSharedPage = NIL_GMM_PAGEID;
4401	}
4402
4403	rc = GMMR3AllocatePagesPerform(pVM, pReq);
4404	if (RT_FAILURE(rc))
4405	{
4406	GMMR3AllocatePagesCleanup(pReq);
4407	return rc;
4408	}
4409	}
4410
4411	/*
4412	* Allocate the new ROM range and RAM range (if necessary).
4413	*/
4414	PPGMROMRANGE pRomNew;
4415	int rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMROMRANGE, aPages[cPages]), 0, MM_TAG_PGM_PHYS, (void **)&pRomNew);
4416	if (RT_SUCCESS(rc))
4417	{
4418	PPGMRAMRANGE pRamNew = NULL;
4419	if (!fRamExists)
4420	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), sizeof(PGMPAGE), MM_TAG_PGM_PHYS, (void **)&pRamNew);
4421	if (RT_SUCCESS(rc))
4422	{
4423	/*
4424	* Initialize and insert the RAM range (if required).
4425	*/
4426	uint32_t const idxFirstRamPage = fRamExists ? (GCPhys - pRam->GCPhys) >> PAGE_SHIFT : 0;
4427	PPGMROMPAGE pRomPage = &pRomNew->aPages[0];
4428	if (!fRamExists)
4429	{
4430	/* New RAM range. */
4431	pRamNew->pSelfR0 = MMHyperCCToR0(pVM, pRamNew);
4432	pRamNew->GCPhys = GCPhys;
4433	pRamNew->GCPhysLast = GCPhysLast;
4434	pRamNew->cb = cb;
4435	pRamNew->pszDesc = pszDesc;
4436	pRamNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_ROM;
4437	pRamNew->pvR3 = NULL;
4438	pRamNew->paLSPages = NULL;
4439	#ifdef VBOX_WITH_NATIVE_NEM
4440	pRamNew->uNemRange = uNemRange;
4441	#endif
4442
4443	PPGMPAGE pRamPage = &pRamNew->aPages[idxFirstRamPage];
4444	#ifdef VBOX_WITH_PGM_NEM_MODE
4445	if (pVM->pgm.s.fNemMode)
4446	{
4447	AssertPtr(pvRam); Assert(pReq == NULL);
4448	pRamNew->pvR3 = pvRam;
4449	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4450	{
4451	PGM_PAGE_INIT(pRamPage, UINT64_C(0x0000fffffffff000), NIL_GMM_PAGEID,
4452	PGMPAGETYPE_ROM, PGM_PAGE_STATE_ALLOCATED);
4453	pRomPage->Virgin = *pRamPage;
4454	}
4455	}
4456	else
4457	#endif
4458	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4459	{
4460	PGM_PAGE_INIT(pRamPage,
4461	pReq->aPages[iPage].HCPhysGCPhys,
4462	pReq->aPages[iPage].idPage,
4463	PGMPAGETYPE_ROM,
4464	PGM_PAGE_STATE_ALLOCATED);
4465
4466	pRomPage->Virgin = *pRamPage;
4467	}
4468
4469	pVM->pgm.s.cAllPages += cPages;
4470	pVM->pgm.s.cPrivatePages += cPages;
4471	pgmR3PhysLinkRamRange(pVM, pRamNew, pRamPrev);
4472	}
4473	else
4474	{
4475	/* Existing RAM range. */
4476	PPGMPAGE pRamPage = &pRam->aPages[idxFirstRamPage];
4477	#ifdef VBOX_WITH_PGM_NEM_MODE
4478	if (pVM->pgm.s.fNemMode)
4479	{
4480	Assert(pvRam == NULL); Assert(pReq == NULL);
4481	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4482	{
4483	Assert(PGM_PAGE_GET_HCPHYS(pRamPage) == UINT64_C(0x0000fffffffff000));
4484	Assert(PGM_PAGE_GET_PAGEID(pRamPage) == NIL_GMM_PAGEID);
4485	Assert(PGM_PAGE_GET_STATE(pRamPage) == PGM_PAGE_STATE_ALLOCATED);
4486	PGM_PAGE_SET_TYPE(pVM, pRamPage, PGMPAGETYPE_ROM);
4487	PGM_PAGE_SET_STATE(pVM, pRamPage, PGM_PAGE_STATE_ALLOCATED);
4488	PGM_PAGE_SET_PDE_TYPE(pVM, pRamPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4489	PGM_PAGE_SET_PTE_INDEX(pVM, pRamPage, 0);
4490	PGM_PAGE_SET_TRACKING(pVM, pRamPage, 0);
4491
4492	pRomPage->Virgin = *pRamPage;
4493	}
4494	}
4495	else
4496	#endif
4497	{
4498	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4499	{
4500	PGM_PAGE_SET_TYPE(pVM, pRamPage, PGMPAGETYPE_ROM);
4501	PGM_PAGE_SET_HCPHYS(pVM, pRamPage, pReq->aPages[iPage].HCPhysGCPhys);
4502	PGM_PAGE_SET_STATE(pVM, pRamPage, PGM_PAGE_STATE_ALLOCATED);
4503	PGM_PAGE_SET_PAGEID(pVM, pRamPage, pReq->aPages[iPage].idPage);
4504	PGM_PAGE_SET_PDE_TYPE(pVM, pRamPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4505	PGM_PAGE_SET_PTE_INDEX(pVM, pRamPage, 0);
4506	PGM_PAGE_SET_TRACKING(pVM, pRamPage, 0);
4507
4508	pRomPage->Virgin = *pRamPage;
4509	}
4510	pVM->pgm.s.cZeroPages -= cPages;
4511	pVM->pgm.s.cPrivatePages += cPages;
4512	}
4513	pRamNew = pRam;
4514	}
4515
4516	#ifdef VBOX_WITH_NATIVE_NEM
4517	/* Set the NEM state of the pages if needed. */
4518	if (u2NemState != UINT8_MAX)
4519	pgmPhysSetNemStateForPages(&pRamNew->aPages[idxFirstRamPage], cPages, u2NemState);
4520	#endif
4521
4522	/* Flush physical page map TLB. */
4523	pgmPhysInvalidatePageMapTLB(pVM);
4524
4525	/*
4526	* Register the ROM access handler.
4527	*/
4528	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, pVM->pgm.s.hRomPhysHandlerType,
4529	pRomNew, MMHyperCCToR0(pVM, pRomNew), NIL_RTRCPTR, pszDesc);
4530	if (RT_SUCCESS(rc))
4531	{
4532	/*
4533	* Copy the image over to the virgin pages.
4534	* This must be done after linking in the RAM range.
4535	*/
4536	size_t cbBinaryLeft = cbBinary;
4537	PPGMPAGE pRamPage = &pRamNew->aPages[idxFirstRamPage];
4538	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
4539	{
4540	void *pvDstPage;
4541	rc = pgmPhysPageMap(pVM, pRamPage, GCPhys + (iPage << PAGE_SHIFT), &pvDstPage);
4542	if (RT_FAILURE(rc))
4543	{
4544	VMSetError(pVM, rc, RT_SRC_POS, "Failed to map virgin ROM page at %RGp", GCPhys);
4545	break;
4546	}
4547	if (cbBinaryLeft >= PAGE_SIZE)
4548	{
4549	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), PAGE_SIZE);
4550	cbBinaryLeft -= PAGE_SIZE;
4551	}
4552	else
4553	{
4554	ASMMemZeroPage(pvDstPage); /* (shouldn't be necessary, but can't hurt either) */
4555	if (cbBinaryLeft > 0)
4556	{
4557	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), cbBinaryLeft);
4558	cbBinaryLeft = 0;
4559	}
4560	}
4561	}
4562	if (RT_SUCCESS(rc))
4563	{
4564	/*
4565	* Initialize the ROM range.
4566	* Note that the Virgin member of the pages has already been initialized above.
4567	*/
4568	pRomNew->GCPhys = GCPhys;
4569	pRomNew->GCPhysLast = GCPhysLast;
4570	pRomNew->cb = cb;
4571	pRomNew->fFlags = fFlags;
4572	pRomNew->idSavedState = UINT8_MAX;
4573	pRomNew->cbOriginal = cbBinary;
4574	pRomNew->pszDesc = pszDesc;
4575	#ifdef VBOX_WITH_PGM_NEM_MODE
4576	pRomNew->pbR3Alternate = (uint8_t *)pvAlt;
4577	#endif
4578	pRomNew->pvOriginal = fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY
4579	? pvBinary : RTMemDup(pvBinary, cbBinary);
4580	if (pRomNew->pvOriginal)
4581	{
4582	for (unsigned iPage = 0; iPage < cPages; iPage++)
4583	{
4584	PPGMROMPAGE pPage = &pRomNew->aPages[iPage];
4585	pPage->enmProt = PGMROMPROT_READ_ROM_WRITE_IGNORE;
4586	#ifdef VBOX_WITH_PGM_NEM_MODE
4587	if (pVM->pgm.s.fNemMode)
4588	PGM_PAGE_INIT(&pPage->Shadow, UINT64_C(0x0000fffffffff000), NIL_GMM_PAGEID,
4589	PGMPAGETYPE_ROM_SHADOW, PGM_PAGE_STATE_ALLOCATED);
4590	else
4591	#endif
4592	PGM_PAGE_INIT_ZERO(&pPage->Shadow, pVM, PGMPAGETYPE_ROM_SHADOW);
4593	}
4594
4595	/* update the page count stats for the shadow pages. */
4596	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4597	{
4598	#ifdef VBOX_WITH_PGM_NEM_MODE
4599	if (pVM->pgm.s.fNemMode)
4600	pVM->pgm.s.cPrivatePages += cPages;
4601	else
4602	#endif
4603	pVM->pgm.s.cZeroPages += cPages;
4604	pVM->pgm.s.cAllPages += cPages;
4605	}
4606
4607	/*
4608	* Insert the ROM range, tell REM and return successfully.
4609	*/
4610	pRomNew->pNextR3 = pRom;
4611	pRomNew->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4612
4613	if (pRomPrev)
4614	{
4615	pRomPrev->pNextR3 = pRomNew;
4616	pRomPrev->pNextR0 = MMHyperCCToR0(pVM, pRomNew);
4617	}
4618	else
4619	{
4620	pVM->pgm.s.pRomRangesR3 = pRomNew;
4621	pVM->pgm.s.pRomRangesR0 = MMHyperCCToR0(pVM, pRomNew);
4622	}
4623
4624	pgmPhysInvalidatePageMapTLB(pVM);
4625	#ifdef VBOX_WITH_PGM_NEM_MODE
4626	if (!pVM->pgm.s.fNemMode)
4627	#endif
4628	GMMR3AllocatePagesCleanup(pReq);
4629
4630	#ifdef VBOX_WITH_NATIVE_NEM
4631	/*
4632	* Notify NEM again.
4633	*/
4634	if (VM_IS_NEM_ENABLED(pVM))
4635	{
4636	u2NemState = UINT8_MAX;
4637	rc = NEMR3NotifyPhysRomRegisterLate(pVM, GCPhys, cb, PGM_RAMRANGE_CALC_PAGE_R3PTR(pRamNew, GCPhys),
4638	fNemNotify, &u2NemState,
4639	fRamExists ? &pRam->uNemRange : &pRamNew->uNemRange);
4640	if (u2NemState != UINT8_MAX)
4641	pgmPhysSetNemStateForPages(&pRamNew->aPages[idxFirstRamPage], cPages, u2NemState);
4642	if (RT_SUCCESS(rc))
4643	return rc;
4644	}
4645	else
4646	#endif
4647	return rc;
4648
4649	/*
4650	* bail out
4651	*/
4652	#ifdef VBOX_WITH_NATIVE_NEM
4653	/* unlink */
4654	if (pRomPrev)
4655	{
4656	pRomPrev->pNextR3 = pRom;
4657	pRomPrev->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4658	}
4659	else
4660	{
4661	pVM->pgm.s.pRomRangesR3 = pRom;
4662	pVM->pgm.s.pRomRangesR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4663	}
4664
4665	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4666	{
4667	# ifdef VBOX_WITH_PGM_NEM_MODE
4668	if (pVM->pgm.s.fNemMode)
4669	pVM->pgm.s.cPrivatePages -= cPages;
4670	else
4671	# endif
4672	pVM->pgm.s.cZeroPages -= cPages;
4673	pVM->pgm.s.cAllPages -= cPages;
4674	}
4675	#endif
4676	}
4677	else
4678	rc = VERR_NO_MEMORY;
4679	}
4680
4681	int rc2 = PGMHandlerPhysicalDeregister(pVM, GCPhys);
4682	AssertRC(rc2);
4683	}
4684
4685	if (!fRamExists)
4686	{
4687	pgmR3PhysUnlinkRamRange2(pVM, pRamNew, pRamPrev);
4688	MMHyperFree(pVM, pRamNew);
4689	}
4690	else
4691	{
4692	PPGMPAGE pRamPage = &pRam->aPages[idxFirstRamPage];
4693	#ifdef VBOX_WITH_PGM_NEM_MODE
4694	if (pVM->pgm.s.fNemMode)
4695	{
4696	Assert(pvRam == NULL); Assert(pReq == NULL);
4697	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4698	{
4699	Assert(PGM_PAGE_GET_HCPHYS(pRamPage) == UINT64_C(0x0000fffffffff000));
4700	Assert(PGM_PAGE_GET_PAGEID(pRamPage) == NIL_GMM_PAGEID);
4701	Assert(PGM_PAGE_GET_STATE(pRamPage) == PGM_PAGE_STATE_ALLOCATED);
4702	PGM_PAGE_SET_TYPE(pVM, pRamPage, PGMPAGETYPE_RAM);
4703	PGM_PAGE_SET_STATE(pVM, pRamPage, PGM_PAGE_STATE_ALLOCATED);
4704	}
4705	}
4706	else
4707	#endif
4708	{
4709	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
4710	PGM_PAGE_INIT_ZERO(pRamPage, pVM, PGMPAGETYPE_RAM);
4711	pVM->pgm.s.cZeroPages += cPages;
4712	pVM->pgm.s.cPrivatePages -= cPages;
4713	}
4714	}
4715	}
4716	MMHyperFree(pVM, pRomNew);
4717	}
4718
4719	/** @todo Purge the mapping cache or something... */
4720	#ifdef VBOX_WITH_PGM_NEM_MODE
4721	if (pVM->pgm.s.fNemMode)
4722	{
4723	Assert(!pReq);
4724	if (pvRam)
4725	SUPR3PageFree(pvRam, cPages);
4726	if (pvAlt)
4727	SUPR3PageFree(pvAlt, cPages);
4728	}
4729	else
4730	#endif
4731	{
4732	GMMR3FreeAllocatedPages(pVM, pReq);
4733	GMMR3AllocatePagesCleanup(pReq);
4734	}
4735	return rc;
4736	}
4737
4738
4739	/**
4740	* Registers a ROM image.
4741	*
4742	* Shadowed ROM images requires double the amount of backing memory, so,
4743	* don't use that unless you have to. Shadowing of ROM images is process
4744	* where we can select where the reads go and where the writes go. On real
4745	* hardware the chipset provides means to configure this. We provide
4746	* PGMR3PhysProtectROM() for this purpose.
4747	*
4748	* A read-only copy of the ROM image will always be kept around while we
4749	* will allocate RAM pages for the changes on demand (unless all memory
4750	* is configured to be preallocated).
4751	*
4752	* @returns VBox status code.
4753	* @param pVM The cross context VM structure.
4754	* @param pDevIns The device instance owning the ROM.
4755	* @param GCPhys First physical address in the range.
4756	* Must be page aligned!
4757	* @param cb The size of the range (in bytes).
4758	* Must be page aligned!
4759	* @param pvBinary Pointer to the binary data backing the ROM image.
4760	* @param cbBinary The size of the binary data pvBinary points to.
4761	* This must be less or equal to @a cb.
4762	* @param fFlags Mask of flags, PGMPHYS_ROM_FLAGS_XXX.
4763	* @param pszDesc Pointer to description string. This must not be freed.
4764	*
4765	* @remark There is no way to remove the rom, automatically on device cleanup or
4766	* manually from the device yet. This isn't difficult in any way, it's
4767	* just not something we expect to be necessary for a while.
4768	*/
4769	VMMR3DECL(int) PGMR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
4770	const void pvBinary, uint32_t cbBinary, uint8_t fFlags, const char pszDesc)
4771	{
4772	Log(("PGMR3PhysRomRegister: pDevIns=%p GCPhys=%RGp(-%RGp) cb=%RGp pvBinary=%p cbBinary=%#x fFlags=%#x pszDesc=%s\n",
4773	pDevIns, GCPhys, GCPhys + cb, cb, pvBinary, cbBinary, fFlags, pszDesc));
4774	PGM_LOCK_VOID(pVM);
4775	int rc = pgmR3PhysRomRegisterLocked(pVM, pDevIns, GCPhys, cb, pvBinary, cbBinary, fFlags, pszDesc);
4776	PGM_UNLOCK(pVM);
4777	return rc;
4778	}
4779
4780
4781	/**
4782	* Called by PGMR3MemSetup to reset the shadow, switch to the virgin, and verify
4783	* that the virgin part is untouched.
4784	*
4785	* This is done after the normal memory has been cleared.
4786	*
4787	* ASSUMES that the caller owns the PGM lock.
4788	*
4789	* @param pVM The cross context VM structure.
4790	*/
4791	int pgmR3PhysRomReset(PVM pVM)
4792	{
4793	PGM_LOCK_ASSERT_OWNER(pVM);
4794	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4795	{
4796	const uint32_t cPages = pRom->cb >> PAGE_SHIFT;
4797
4798	if (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4799	{
4800	/*
4801	* Reset the physical handler.
4802	*/
4803	int rc = PGMR3PhysRomProtect(pVM, pRom->GCPhys, pRom->cb, PGMROMPROT_READ_ROM_WRITE_IGNORE);
4804	AssertRCReturn(rc, rc);
4805
4806	/*
4807	* What we do with the shadow pages depends on the memory
4808	* preallocation option. If not enabled, we'll just throw
4809	* out all the dirty pages and replace them by the zero page.
4810	*/
4811	#ifdef VBOX_WITH_PGM_NEM_MODE
4812	if (pVM->pgm.s.fNemMode)
4813	{
4814	/* Clear all the shadow pages (currently using alternate backing). */
4815	RT_BZERO(pRom->pbR3Alternate, pRom->cb);
4816	}
4817	else
4818	#endif
4819	if (!pVM->pgm.s.fRamPreAlloc)
4820	{
4821	/* Free the dirty pages. */
4822	uint32_t cPendingPages = 0;
4823	PGMMFREEPAGESREQ pReq;
4824	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
4825	AssertRCReturn(rc, rc);
4826
4827	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4828	if ( !PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow)
4829	&& !PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow))
4830	{
4831	Assert(PGM_PAGE_GET_STATE(&pRom->aPages[iPage].Shadow) == PGM_PAGE_STATE_ALLOCATED);
4832	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, &pRom->aPages[iPage].Shadow,
4833	pRom->GCPhys + (iPage << PAGE_SHIFT),
4834	(PGMPAGETYPE)PGM_PAGE_GET_TYPE(&pRom->aPages[iPage].Shadow));
4835	AssertLogRelRCReturn(rc, rc);
4836	}
4837
4838	if (cPendingPages)
4839	{
4840	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
4841	AssertLogRelRCReturn(rc, rc);
4842	}
4843	GMMR3FreePagesCleanup(pReq);
4844	}
4845	else
4846	{
4847	/* clear all the shadow pages. */
4848	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4849	{
4850	if (PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow))
4851	continue;
4852	Assert(!PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow));
4853	void *pvDstPage;
4854	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4855	rc = pgmPhysPageMakeWritableAndMap(pVM, &pRom->aPages[iPage].Shadow, GCPhys, &pvDstPage);
4856	if (RT_FAILURE(rc))
4857	break;
4858	ASMMemZeroPage(pvDstPage);
4859	}
4860	AssertRCReturn(rc, rc);
4861	}
4862	}
4863
4864	/*
4865	* Restore the original ROM pages after a saved state load.
4866	* Also, in strict builds check that ROM pages remain unmodified.
4867	*/
4868	#ifndef VBOX_STRICT
4869	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4870	#endif
4871	{
4872	size_t cbSrcLeft = pRom->cbOriginal;
4873	uint8_t const pbSrcPage = (uint8_t const )pRom->pvOriginal;
4874	uint32_t cRestored = 0;
4875	for (uint32_t iPage = 0; iPage < cPages && cbSrcLeft > 0; iPage++, pbSrcPage += PAGE_SIZE)
4876	{
4877	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4878	PPGMPAGE const pPage = pgmPhysGetPage(pVM, GCPhys);
4879	void const *pvDstPage = NULL;
4880	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhys, &pvDstPage);
4881	if (RT_FAILURE(rc))
4882	break;
4883
4884	if (memcmp(pvDstPage, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE)))
4885	{
4886	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4887	{
4888	void *pvDstPageW = NULL;
4889	rc = pgmPhysPageMap(pVM, pPage, GCPhys, &pvDstPageW);
4890	AssertLogRelRCReturn(rc, rc);
4891	memcpy(pvDstPageW, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE));
4892	cRestored++;
4893	}
4894	else
4895	LogRel(("pgmR3PhysRomReset: %RGp: ROM page changed (%s)\n", GCPhys, pRom->pszDesc));
4896	}
4897	cbSrcLeft -= RT_MIN(cbSrcLeft, PAGE_SIZE);
4898	}
4899	if (cRestored > 0)
4900	LogRel(("PGM: ROM \"%s\": Reloaded %u of %u pages.\n", pRom->pszDesc, cRestored, cPages));
4901	}
4902	}
4903
4904	/* Clear the ROM restore flag now as we only need to do this once after
4905	loading saved state. */
4906	pVM->pgm.s.fRestoreRomPagesOnReset = false;
4907
4908	return VINF_SUCCESS;
4909	}
4910
4911
4912	/**
4913	* Called by PGMR3Term to free resources.
4914	*
4915	* ASSUMES that the caller owns the PGM lock.
4916	*
4917	* @param pVM The cross context VM structure.
4918	*/
4919	void pgmR3PhysRomTerm(PVM pVM)
4920	{
4921	/*
4922	* Free the heap copy of the original bits.
4923	*/
4924	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4925	{
4926	if ( pRom->pvOriginal
4927	&& !(pRom->fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY))
4928	{
4929	RTMemFree((void *)pRom->pvOriginal);
4930	pRom->pvOriginal = NULL;
4931	}
4932	}
4933	}
4934
4935
4936	/**
4937	* Change the shadowing of a range of ROM pages.
4938	*
4939	* This is intended for implementing chipset specific memory registers
4940	* and will not be very strict about the input. It will silently ignore
4941	* any pages that are not the part of a shadowed ROM.
4942	*
4943	* @returns VBox status code.
4944	* @retval VINF_PGM_SYNC_CR3
4945	*
4946	* @param pVM The cross context VM structure.
4947	* @param GCPhys Where to start. Page aligned.
4948	* @param cb How much to change. Page aligned.
4949	* @param enmProt The new ROM protection.
4950	*/
4951	VMMR3DECL(int) PGMR3PhysRomProtect(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMROMPROT enmProt)
4952	{
4953	/*
4954	* Check input
4955	*/
4956	if (!cb)
4957	return VINF_SUCCESS;
4958	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4959	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4960	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
4961	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
4962	AssertReturn(enmProt >= PGMROMPROT_INVALID && enmProt <= PGMROMPROT_END, VERR_INVALID_PARAMETER);
4963
4964	/*
4965	* Process the request.
4966	*/
4967	PGM_LOCK_VOID(pVM);
4968	int rc = VINF_SUCCESS;
4969	bool fFlushTLB = false;
4970	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4971	{
4972	if ( GCPhys <= pRom->GCPhysLast
4973	&& GCPhysLast >= pRom->GCPhys
4974	&& (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED))
4975	{
4976	/*
4977	* Iterate the relevant pages and make necessary the changes.
4978	*/
4979	#ifdef VBOX_WITH_NATIVE_NEM
4980	PPGMRAMRANGE const pRam = pgmPhysGetRange(pVM, GCPhys);
4981	AssertPtrReturn(pRam, VERR_INTERNAL_ERROR_3);
4982	#endif
4983	bool fChanges = false;
4984	uint32_t const cPages = pRom->GCPhysLast <= GCPhysLast
4985	? pRom->cb >> PAGE_SHIFT
4986	: (GCPhysLast - pRom->GCPhys + 1) >> PAGE_SHIFT;
4987	for (uint32_t iPage = (GCPhys - pRom->GCPhys) >> PAGE_SHIFT;
4988	iPage < cPages;
4989	iPage++)
4990	{
4991	PPGMROMPAGE pRomPage = &pRom->aPages[iPage];
4992	if (PGMROMPROT_IS_ROM(pRomPage->enmProt) != PGMROMPROT_IS_ROM(enmProt))
4993	{
4994	fChanges = true;
4995
4996	/* flush references to the page. */
4997	RTGCPHYS const GCPhysPage = pRom->GCPhys + (iPage << PAGE_SHIFT);
4998	PPGMPAGE pRamPage = pgmPhysGetPage(pVM, GCPhysPage);
4999	int rc2 = pgmPoolTrackUpdateGCPhys(pVM, GCPhysPage, pRamPage, true /fFlushPTEs/, &fFlushTLB);
5000	if (rc2 != VINF_SUCCESS && (rc == VINF_SUCCESS \|\| RT_FAILURE(rc2)))
5001	rc = rc2;
5002	#ifdef VBOX_WITH_NATIVE_NEM
5003	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pRamPage);
5004	#endif
5005
5006	PPGMPAGE pOld = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Virgin : &pRomPage->Shadow;
5007	PPGMPAGE pNew = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Shadow : &pRomPage->Virgin;
5008
5009	pOld = pRamPage;
5010	pRamPage = pNew;
5011	/** @todo preserve the volatile flags (handlers) when these have been moved out of HCPhys! */
5012
5013	#ifdef VBOX_WITH_NATIVE_NEM
5014	# ifdef VBOX_WITH_PGM_NEM_MODE
5015	/* In simplified mode we have to switch the page data around too. */
5016	if (pVM->pgm.s.fNemMode)
5017	{
5018	uint8_t abPage[PAGE_SIZE];
5019	uint8_t * const pbRamPage = PGM_RAMRANGE_CALC_PAGE_R3PTR(pRam, GCPhysPage);
5020	memcpy(abPage, &pRom->pbR3Alternate[(size_t)iPage << PAGE_SHIFT], sizeof(abPage));
5021	memcpy(&pRom->pbR3Alternate[(size_t)iPage << PAGE_SHIFT], pbRamPage, sizeof(abPage));
5022	memcpy(pbRamPage, abPage, sizeof(abPage));
5023	}
5024	# endif
5025	/* Tell NEM about the backing and protection change. */
5026	if (VM_IS_NEM_ENABLED(pVM))
5027	{
5028	PGMPAGETYPE enmType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pNew);
5029	NEMHCNotifyPhysPageChanged(pVM, GCPhys, PGM_PAGE_GET_HCPHYS(pOld), PGM_PAGE_GET_HCPHYS(pNew),
5030	PGM_RAMRANGE_CALC_PAGE_R3PTR(pRam, GCPhysPage),
5031	pgmPhysPageCalcNemProtection(pRamPage, enmType), enmType, &u2State);
5032	PGM_PAGE_SET_NEM_STATE(pRamPage, u2State);
5033	}
5034	#endif
5035	}
5036	pRomPage->enmProt = enmProt;
5037	}
5038
5039	/*
5040	* Reset the access handler if we made changes, no need
5041	* to optimize this.
5042	*/
5043	if (fChanges)
5044	{
5045	int rc2 = PGMHandlerPhysicalReset(pVM, pRom->GCPhys);
5046	if (RT_FAILURE(rc2))
5047	{
5048	PGM_UNLOCK(pVM);
5049	AssertRC(rc);
5050	return rc2;
5051	}
5052	}
5053
5054	/* Advance - cb isn't updated. */
5055	GCPhys = pRom->GCPhys + (cPages << PAGE_SHIFT);
5056	}
5057	}
5058	PGM_UNLOCK(pVM);
5059	if (fFlushTLB)
5060	PGM_INVL_ALL_VCPU_TLBS(pVM);
5061
5062	return rc;
5063	}
5064
5065
5066
5067	/*********************************************************************************************************************************
5068	* Ballooning *
5069	*********************************************************************************************************************************/
5070
5071	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
5072
5073	/**
5074	* Rendezvous callback used by PGMR3ChangeMemBalloon that changes the memory balloon size
5075	*
5076	* This is only called on one of the EMTs while the other ones are waiting for
5077	* it to complete this function.
5078	*
5079	* @returns VINF_SUCCESS (VBox strict status code).
5080	* @param pVM The cross context VM structure.
5081	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
5082	* @param pvUser User parameter
5083	*/
5084	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysChangeMemBalloonRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
5085	{
5086	uintptr_t paUser = (uintptr_t )pvUser;
5087	bool fInflate = !!paUser[0];
5088	unsigned cPages = paUser[1];
5089	RTGCPHYS paPhysPage = (RTGCPHYS )paUser[2];
5090	uint32_t cPendingPages = 0;
5091	PGMMFREEPAGESREQ pReq;
5092	int rc;
5093
5094	Log(("pgmR3PhysChangeMemBalloonRendezvous: %s %x pages\n", (fInflate) ? "inflate" : "deflate", cPages));
5095	PGM_LOCK_VOID(pVM);
5096
5097	if (fInflate)
5098	{
5099	/* Flush the PGM pool cache as we might have stale references to pages that we just freed. */
5100	pgmR3PoolClearAllRendezvous(pVM, pVCpu, NULL);
5101
5102	/* Replace pages with ZERO pages. */
5103	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
5104	if (RT_FAILURE(rc))
5105	{
5106	PGM_UNLOCK(pVM);
5107	AssertLogRelRC(rc);
5108	return rc;
5109	}
5110
5111	/* Iterate the pages. */
5112	for (unsigned i = 0; i < cPages; i++)
5113	{
5114	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
5115	if ( pPage == NULL
5116	\|\| PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM)
5117	{
5118	Log(("pgmR3PhysChangeMemBalloonRendezvous: invalid physical page %RGp pPage->u3Type=%d\n", paPhysPage[i], pPage ? PGM_PAGE_GET_TYPE(pPage) : 0));
5119	break;
5120	}
5121
5122	LogFlow(("balloon page: %RGp\n", paPhysPage[i]));
5123
5124	/* Flush the shadow PT if this page was previously used as a guest page table. */
5125	pgmPoolFlushPageByGCPhys(pVM, paPhysPage[i]);
5126
5127	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, paPhysPage[i], (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage));
5128	if (RT_FAILURE(rc))
5129	{
5130	PGM_UNLOCK(pVM);
5131	AssertLogRelRC(rc);
5132	return rc;
5133	}
5134	Assert(PGM_PAGE_IS_ZERO(pPage));
5135	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_BALLOONED);
5136	}
5137
5138	if (cPendingPages)
5139	{
5140	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
5141	if (RT_FAILURE(rc))
5142	{
5143	PGM_UNLOCK(pVM);
5144	AssertLogRelRC(rc);
5145	return rc;
5146	}
5147	}
5148	GMMR3FreePagesCleanup(pReq);
5149	}
5150	else
5151	{
5152	/* Iterate the pages. */
5153	for (unsigned i = 0; i < cPages; i++)
5154	{
5155	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
5156	AssertBreak(pPage && PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM);
5157
5158	LogFlow(("Free ballooned page: %RGp\n", paPhysPage[i]));
5159
5160	Assert(PGM_PAGE_IS_BALLOONED(pPage));
5161
5162	/* Change back to zero page. (NEM does not need to be informed.) */
5163	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
5164	}
5165
5166	/* Note that we currently do not map any ballooned pages in our shadow page tables, so no need to flush the pgm pool. */
5167	}
5168
5169	/* Notify GMM about the balloon change. */
5170	rc = GMMR3BalloonedPages(pVM, (fInflate) ? GMMBALLOONACTION_INFLATE : GMMBALLOONACTION_DEFLATE, cPages);
5171	if (RT_SUCCESS(rc))
5172	{
5173	if (!fInflate)
5174	{
5175	Assert(pVM->pgm.s.cBalloonedPages >= cPages);
5176	pVM->pgm.s.cBalloonedPages -= cPages;
5177	}
5178	else
5179	pVM->pgm.s.cBalloonedPages += cPages;
5180	}
5181
5182	PGM_UNLOCK(pVM);
5183
5184	/* Flush the recompiler's TLB as well. */
5185	for (VMCPUID i = 0; i < pVM->cCpus; i++)
5186	CPUMSetChangedFlags(pVM->apCpusR3[i], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
5187
5188	AssertLogRelRC(rc);
5189	return rc;
5190	}
5191
5192
5193	/**
5194	* Frees a range of ram pages, replacing them with ZERO pages; helper for PGMR3PhysFreeRamPages
5195	*
5196	* @returns VBox status code.
5197	* @param pVM The cross context VM structure.
5198	* @param fInflate Inflate or deflate memory balloon
5199	* @param cPages Number of pages to free
5200	* @param paPhysPage Array of guest physical addresses
5201	*/
5202	static DECLCALLBACK(void) pgmR3PhysChangeMemBalloonHelper(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
5203	{
5204	uintptr_t paUser[3];
5205
5206	paUser[0] = fInflate;
5207	paUser[1] = cPages;
5208	paUser[2] = (uintptr_t)paPhysPage;
5209	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
5210	AssertRC(rc);
5211
5212	/* Made a copy in PGMR3PhysFreeRamPages; free it here. */
5213	RTMemFree(paPhysPage);
5214	}
5215
5216	#endif /* 64-bit host && (Windows \|\| Solaris \|\| Linux \|\| FreeBSD) */
5217
5218	/**
5219	* Inflate or deflate a memory balloon
5220	*
5221	* @returns VBox status code.
5222	* @param pVM The cross context VM structure.
5223	* @param fInflate Inflate or deflate memory balloon
5224	* @param cPages Number of pages to free
5225	* @param paPhysPage Array of guest physical addresses
5226	*/
5227	VMMR3DECL(int) PGMR3PhysChangeMemBalloon(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
5228	{
5229	/* This must match GMMR0Init; currently we only support memory ballooning on all 64-bit hosts except Mac OS X */
5230	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
5231	int rc;
5232
5233	/* Older additions (ancient non-functioning balloon code) pass wrong physical addresses. */
5234	AssertReturn(!(paPhysPage[0] & 0xfff), VERR_INVALID_PARAMETER);
5235
5236	/* We own the IOM lock here and could cause a deadlock by waiting for another VCPU that is blocking on the IOM lock.
5237	* In the SMP case we post a request packet to postpone the job.
5238	*/
5239	if (pVM->cCpus > 1)
5240	{
5241	unsigned cbPhysPage = cPages * sizeof(paPhysPage[0]);
5242	RTGCPHYS paPhysPageCopy = (RTGCPHYS )RTMemAlloc(cbPhysPage);
5243	AssertReturn(paPhysPageCopy, VERR_NO_MEMORY);
5244
5245	memcpy(paPhysPageCopy, paPhysPage, cbPhysPage);
5246
5247	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysChangeMemBalloonHelper, 4, pVM, fInflate, cPages, paPhysPageCopy);
5248	AssertRC(rc);
5249	}
5250	else
5251	{
5252	uintptr_t paUser[3];
5253
5254	paUser[0] = fInflate;
5255	paUser[1] = cPages;
5256	paUser[2] = (uintptr_t)paPhysPage;
5257	rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
5258	AssertRC(rc);
5259	}
5260	return rc;
5261
5262	#else
5263	NOREF(pVM); NOREF(fInflate); NOREF(cPages); NOREF(paPhysPage);
5264	return VERR_NOT_IMPLEMENTED;
5265	#endif
5266	}
5267
5268
5269	/*********************************************************************************************************************************
5270	* Write Monitoring *
5271	*********************************************************************************************************************************/
5272
5273	/**
5274	* Rendezvous callback used by PGMR3WriteProtectRAM that write protects all
5275	* physical RAM.
5276	*
5277	* This is only called on one of the EMTs while the other ones are waiting for
5278	* it to complete this function.
5279	*
5280	* @returns VINF_SUCCESS (VBox strict status code).
5281	* @param pVM The cross context VM structure.
5282	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
5283	* @param pvUser User parameter, unused.
5284	*/
5285	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysWriteProtectRAMRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
5286	{
5287	int rc = VINF_SUCCESS;
5288	NOREF(pvUser); NOREF(pVCpu);
5289
5290	PGM_LOCK_VOID(pVM);
5291	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
5292	pgmPoolResetDirtyPages(pVM);
5293	#endif
5294
5295	/** @todo pointless to write protect the physical page pointed to by RSP. */
5296
5297	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
5298	pRam;
5299	pRam = pRam->CTX_SUFF(pNext))
5300	{
5301	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
5302	for (uint32_t iPage = 0; iPage < cPages; iPage++)
5303	{
5304	PPGMPAGE pPage = &pRam->aPages[iPage];
5305	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
5306
5307	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
5308	\|\| enmPageType == PGMPAGETYPE_MMIO2)
5309	{
5310	/*
5311	* A RAM page.
5312	*/
5313	switch (PGM_PAGE_GET_STATE(pPage))
5314	{
5315	case PGM_PAGE_STATE_ALLOCATED:
5316	/** @todo Optimize this: Don't always re-enable write
5317	* monitoring if the page is known to be very busy. */
5318	if (PGM_PAGE_IS_WRITTEN_TO(pPage))
5319	PGM_PAGE_CLEAR_WRITTEN_TO(pVM, pPage);
5320
5321	pgmPhysPageWriteMonitor(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
5322	break;
5323
5324	case PGM_PAGE_STATE_SHARED:
5325	AssertFailed();
5326	break;
5327
5328	case PGM_PAGE_STATE_WRITE_MONITORED: /* nothing to change. */
5329	default:
5330	break;
5331	}
5332	}
5333	}
5334	}
5335	pgmR3PoolWriteProtectPages(pVM);
5336	PGM_INVL_ALL_VCPU_TLBS(pVM);
5337	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
5338	CPUMSetChangedFlags(pVM->apCpusR3[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
5339
5340	PGM_UNLOCK(pVM);
5341	return rc;
5342	}
5343
5344	/**
5345	* Protect all physical RAM to monitor writes
5346	*
5347	* @returns VBox status code.
5348	* @param pVM The cross context VM structure.
5349	*/
5350	VMMR3DECL(int) PGMR3PhysWriteProtectRAM(PVM pVM)
5351	{
5352	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
5353
5354	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysWriteProtectRAMRendezvous, NULL);
5355	AssertRC(rc);
5356	return rc;
5357	}
5358
5359
5360	/*********************************************************************************************************************************
5361	* Stats. *
5362	*********************************************************************************************************************************/
5363
5364	/**
5365	* Query the amount of free memory inside VMMR0
5366	*
5367	* @returns VBox status code.
5368	* @param pUVM The user mode VM handle.
5369	* @param pcbAllocMem Where to return the amount of memory allocated
5370	* by VMs.
5371	* @param pcbFreeMem Where to return the amount of memory that is
5372	* allocated from the host but not currently used
5373	* by any VMs.
5374	* @param pcbBallonedMem Where to return the sum of memory that is
5375	* currently ballooned by the VMs.
5376	* @param pcbSharedMem Where to return the amount of memory that is
5377	* currently shared.
5378	*/
5379	VMMR3DECL(int) PGMR3QueryGlobalMemoryStats(PUVM pUVM, uint64_t pcbAllocMem, uint64_t pcbFreeMem,
5380	uint64_t pcbBallonedMem, uint64_t pcbSharedMem)
5381	{
5382	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
5383	VM_ASSERT_VALID_EXT_RETURN(pUVM->pVM, VERR_INVALID_VM_HANDLE);
5384
5385	uint64_t cAllocPages = 0;
5386	uint64_t cFreePages = 0;
5387	uint64_t cBalloonPages = 0;
5388	uint64_t cSharedPages = 0;
5389	if (!SUPR3IsDriverless())
5390	{
5391	int rc = GMMR3QueryHypervisorMemoryStats(pUVM->pVM, &cAllocPages, &cFreePages, &cBalloonPages, &cSharedPages);
5392	AssertRCReturn(rc, rc);
5393	}
5394
5395	if (pcbAllocMem)
5396	pcbAllocMem = cAllocPages _4K;
5397
5398	if (pcbFreeMem)
5399	pcbFreeMem = cFreePages _4K;
5400
5401	if (pcbBallonedMem)
5402	pcbBallonedMem = cBalloonPages _4K;
5403
5404	if (pcbSharedMem)
5405	pcbSharedMem = cSharedPages _4K;
5406
5407	Log(("PGMR3QueryVMMMemoryStats: all=%llx free=%llx ballooned=%llx shared=%llx\n",
5408	cAllocPages, cFreePages, cBalloonPages, cSharedPages));
5409	return VINF_SUCCESS;
5410	}
5411
5412
5413	/**
5414	* Query memory stats for the VM.
5415	*
5416	* @returns VBox status code.
5417	* @param pUVM The user mode VM handle.
5418	* @param pcbTotalMem Where to return total amount memory the VM may
5419	* possibly use.
5420	* @param pcbPrivateMem Where to return the amount of private memory
5421	* currently allocated.
5422	* @param pcbSharedMem Where to return the amount of actually shared
5423	* memory currently used by the VM.
5424	* @param pcbZeroMem Where to return the amount of memory backed by
5425	* zero pages.
5426	*
5427	* @remarks The total mem is normally larger than the sum of the three
5428	* components. There are two reasons for this, first the amount of
5429	* shared memory is what we're sure is shared instead of what could
5430	* possibly be shared with someone. Secondly, because the total may
5431	* include some pure MMIO pages that doesn't go into any of the three
5432	* sub-counts.
5433	*
5434	* @todo Why do we return reused shared pages instead of anything that could
5435	* potentially be shared? Doesn't this mean the first VM gets a much
5436	* lower number of shared pages?
5437	*/
5438	VMMR3DECL(int) PGMR3QueryMemoryStats(PUVM pUVM, uint64_t pcbTotalMem, uint64_t pcbPrivateMem,
5439	uint64_t pcbSharedMem, uint64_t pcbZeroMem)
5440	{
5441	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
5442	PVM pVM = pUVM->pVM;
5443	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
5444
5445	if (pcbTotalMem)
5446	pcbTotalMem = (uint64_t)pVM->pgm.s.cAllPages PAGE_SIZE;
5447
5448	if (pcbPrivateMem)
5449	pcbPrivateMem = (uint64_t)pVM->pgm.s.cPrivatePages PAGE_SIZE;
5450
5451	if (pcbSharedMem)
5452	pcbSharedMem = (uint64_t)pVM->pgm.s.cReusedSharedPages PAGE_SIZE;
5453
5454	if (pcbZeroMem)
5455	pcbZeroMem = (uint64_t)pVM->pgm.s.cZeroPages PAGE_SIZE;
5456
5457	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));
5458	return VINF_SUCCESS;
5459	}
5460
5461
5462
5463	/*********************************************************************************************************************************
5464	* Chunk Mappings and Page Allocation *
5465	*********************************************************************************************************************************/
5466
5467	/**
5468	* Tree enumeration callback for dealing with age rollover.
5469	* It will perform a simple compression of the current age.
5470	*/
5471	static DECLCALLBACK(int) pgmR3PhysChunkAgeingRolloverCallback(PAVLU32NODECORE pNode, void *pvUser)
5472	{
5473	/* Age compression - ASSUMES iNow == 4. */
5474	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
5475	if (pChunk->iLastUsed >= UINT32_C(0xffffff00))
5476	pChunk->iLastUsed = 3;
5477	else if (pChunk->iLastUsed >= UINT32_C(0xfffff000))
5478	pChunk->iLastUsed = 2;
5479	else if (pChunk->iLastUsed)
5480	pChunk->iLastUsed = 1;
5481	else /* iLastUsed = 0 */
5482	pChunk->iLastUsed = 4;
5483
5484	NOREF(pvUser);
5485	return 0;
5486	}
5487
5488
5489	/**
5490	* The structure passed in the pvUser argument of pgmR3PhysChunkUnmapCandidateCallback().
5491	*/
5492	typedef struct PGMR3PHYSCHUNKUNMAPCB
5493	{
5494	PVM pVM; /*< Pointer to the VM. /
5495	PPGMCHUNKR3MAP pChunk; /*< The chunk to unmap. /
5496	} PGMR3PHYSCHUNKUNMAPCB, *PPGMR3PHYSCHUNKUNMAPCB;
5497
5498
5499	/**
5500	* Callback used to find the mapping that's been unused for
5501	* the longest time.
5502	*/
5503	static DECLCALLBACK(int) pgmR3PhysChunkUnmapCandidateCallback(PAVLU32NODECORE pNode, void *pvUser)
5504	{
5505	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
5506	PPGMR3PHYSCHUNKUNMAPCB pArg = (PPGMR3PHYSCHUNKUNMAPCB)pvUser;
5507
5508	/*
5509	* Check for locks and compare when last used.
5510	*/
5511	if (pChunk->cRefs)
5512	return 0;
5513	if (pChunk->cPermRefs)
5514	return 0;
5515	if ( pArg->pChunk
5516	&& pChunk->iLastUsed >= pArg->pChunk->iLastUsed)
5517	return 0;
5518
5519	/*
5520	* Check that it's not in any of the TLBs.
5521	*/
5522	PVM pVM = pArg->pVM;
5523	if ( pVM->pgm.s.ChunkR3Map.Tlb.aEntries[PGM_CHUNKR3MAPTLB_IDX(pChunk->Core.Key)].idChunk
5524	== pChunk->Core.Key)
5525	{
5526	pChunk = NULL;
5527	return 0;
5528	}
5529	#ifdef VBOX_STRICT
5530	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
5531	{
5532	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk != pChunk);
5533	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk != pChunk->Core.Key);
5534	}
5535	#endif
5536
5537	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.PhysTlbR3.aEntries); i++)
5538	if (pVM->pgm.s.PhysTlbR3.aEntries[i].pMap == pChunk)
5539	return 0;
5540
5541	pArg->pChunk = pChunk;
5542	return 0;
5543	}
5544
5545
5546	/**
5547	* Finds a good candidate for unmapping when the ring-3 mapping cache is full.
5548	*
5549	* The candidate will not be part of any TLBs, so no need to flush
5550	* anything afterwards.
5551	*
5552	* @returns Chunk id.
5553	* @param pVM The cross context VM structure.
5554	*/
5555	static int32_t pgmR3PhysChunkFindUnmapCandidate(PVM pVM)
5556	{
5557	PGM_LOCK_ASSERT_OWNER(pVM);
5558
5559	/*
5560	* Enumerate the age tree starting with the left most node.
5561	*/
5562	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
5563	PGMR3PHYSCHUNKUNMAPCB Args;
5564	Args.pVM = pVM;
5565	Args.pChunk = NULL;
5566	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkUnmapCandidateCallback, &Args);
5567	Assert(Args.pChunk);
5568	if (Args.pChunk)
5569	{
5570	Assert(Args.pChunk->cRefs == 0);
5571	Assert(Args.pChunk->cPermRefs == 0);
5572	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
5573	return Args.pChunk->Core.Key;
5574	}
5575
5576	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
5577	return INT32_MAX;
5578	}
5579
5580
5581	/**
5582	* Rendezvous callback used by pgmR3PhysUnmapChunk that unmaps a chunk
5583	*
5584	* This is only called on one of the EMTs while the other ones are waiting for
5585	* it to complete this function.
5586	*
5587	* @returns VINF_SUCCESS (VBox strict status code).
5588	* @param pVM The cross context VM structure.
5589	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
5590	* @param pvUser User pointer. Unused
5591	*
5592	*/
5593	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysUnmapChunkRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
5594	{
5595	int rc = VINF_SUCCESS;
5596	PGM_LOCK_VOID(pVM);
5597	NOREF(pVCpu); NOREF(pvUser);
5598
5599	if (pVM->pgm.s.ChunkR3Map.c >= pVM->pgm.s.ChunkR3Map.cMax)
5600	{
5601	/* Flush the pgm pool cache; call the internal rendezvous handler as we're already in a rendezvous handler here. */
5602	/** @todo also not really efficient to unmap a chunk that contains PD
5603	* or PT pages. */
5604	pgmR3PoolClearAllRendezvous(pVM, pVM->apCpusR3[0], NULL /* no need to flush the REM TLB as we already did that above */);
5605
5606	/*
5607	* Request the ring-0 part to unmap a chunk to make space in the mapping cache.
5608	*/
5609	GMMMAPUNMAPCHUNKREQ Req;
5610	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
5611	Req.Hdr.cbReq = sizeof(Req);
5612	Req.pvR3 = NULL;
5613	Req.idChunkMap = NIL_GMM_CHUNKID;
5614	Req.idChunkUnmap = pgmR3PhysChunkFindUnmapCandidate(pVM);
5615	if (Req.idChunkUnmap != INT32_MAX)
5616	{
5617	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkUnmap, a);
5618	rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
5619	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkUnmap, a);
5620	if (RT_SUCCESS(rc))
5621	{
5622	/*
5623	* Remove the unmapped one.
5624	*/
5625	PPGMCHUNKR3MAP pUnmappedChunk = (PPGMCHUNKR3MAP)RTAvlU32Remove(&pVM->pgm.s.ChunkR3Map.pTree, Req.idChunkUnmap);
5626	AssertRelease(pUnmappedChunk);
5627	AssertRelease(!pUnmappedChunk->cRefs);
5628	AssertRelease(!pUnmappedChunk->cPermRefs);
5629	pUnmappedChunk->pv = NULL;
5630	pUnmappedChunk->Core.Key = UINT32_MAX;
5631	MMR3HeapFree(pUnmappedChunk);
5632	pVM->pgm.s.ChunkR3Map.c--;
5633	pVM->pgm.s.cUnmappedChunks++;
5634
5635	/*
5636	* Flush dangling PGM pointers (R3 & R0 ptrs to GC physical addresses).
5637	*/
5638	/** @todo We should not flush chunks which include cr3 mappings. */
5639	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
5640	{
5641	PPGMCPU pPGM = &pVM->apCpusR3[idCpu]->pgm.s;
5642
5643	pPGM->pGst32BitPdR3 = NULL;
5644	pPGM->pGstPaePdptR3 = NULL;
5645	pPGM->pGstAmd64Pml4R3 = NULL;
5646	pPGM->pGstEptPml4R3 = NULL;
5647	pPGM->pGst32BitPdR0 = NIL_RTR0PTR;
5648	pPGM->pGstPaePdptR0 = NIL_RTR0PTR;
5649	pPGM->pGstAmd64Pml4R0 = NIL_RTR0PTR;
5650	pPGM->pGstEptPml4R0 = NIL_RTR0PTR;
5651	for (unsigned i = 0; i < RT_ELEMENTS(pPGM->apGstPaePDsR3); i++)
5652	{
5653	pPGM->apGstPaePDsR3[i] = NULL;
5654	pPGM->apGstPaePDsR0[i] = NIL_RTR0PTR;
5655	}
5656
5657	/* Flush REM TLBs. */
5658	CPUMSetChangedFlags(pVM->apCpusR3[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
5659	}
5660	}
5661	}
5662	}
5663	PGM_UNLOCK(pVM);
5664	return rc;
5665	}
5666
5667	/**
5668	* Unmap a chunk to free up virtual address space (request packet handler for pgmR3PhysChunkMap)
5669	*
5670	* @returns VBox status code.
5671	* @param pVM The cross context VM structure.
5672	*/
5673	static DECLCALLBACK(void) pgmR3PhysUnmapChunk(PVM pVM)
5674	{
5675	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysUnmapChunkRendezvous, NULL);
5676	AssertRC(rc);
5677	}
5678
5679
5680	/**
5681	* Maps the given chunk into the ring-3 mapping cache.
5682	*
5683	* This will call ring-0.
5684	*
5685	* @returns VBox status code.
5686	* @param pVM The cross context VM structure.
5687	* @param idChunk The chunk in question.
5688	* @param ppChunk Where to store the chunk tracking structure.
5689	*
5690	* @remarks Called from within the PGM critical section.
5691	* @remarks Can be called from any thread!
5692	*/
5693	int pgmR3PhysChunkMap(PVM pVM, uint32_t idChunk, PPPGMCHUNKR3MAP ppChunk)
5694	{
5695	int rc;
5696
5697	PGM_LOCK_ASSERT_OWNER(pVM);
5698
5699	/*
5700	* Move the chunk time forward.
5701	*/
5702	pVM->pgm.s.ChunkR3Map.iNow++;
5703	if (pVM->pgm.s.ChunkR3Map.iNow == 0)
5704	{
5705	pVM->pgm.s.ChunkR3Map.iNow = 4;
5706	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkAgeingRolloverCallback, NULL);
5707	}
5708
5709	/*
5710	* Allocate a new tracking structure first.
5711	*/
5712	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3HeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk));
5713	AssertReturn(pChunk, VERR_NO_MEMORY);
5714	pChunk->Core.Key = idChunk;
5715	pChunk->iLastUsed = pVM->pgm.s.ChunkR3Map.iNow;
5716
5717	/*
5718	* Request the ring-0 part to map the chunk in question.
5719	*/
5720	GMMMAPUNMAPCHUNKREQ Req;
5721	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
5722	Req.Hdr.cbReq = sizeof(Req);
5723	Req.pvR3 = NULL;
5724	Req.idChunkMap = idChunk;
5725	Req.idChunkUnmap = NIL_GMM_CHUNKID;
5726
5727	/* Must be callable from any thread, so can't use VMMR3CallR0. */
5728	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkMap, a);
5729	rc = SUPR3CallVMMR0Ex(VMCC_GET_VMR0_FOR_CALL(pVM), NIL_VMCPUID, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
5730	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkMap, a);
5731	if (RT_SUCCESS(rc))
5732	{
5733	pChunk->pv = Req.pvR3;
5734
5735	/*
5736	* If we're running out of virtual address space, then we should
5737	* unmap another chunk.
5738	*
5739	* Currently, an unmap operation requires that all other virtual CPUs
5740	* are idling and not by chance making use of the memory we're
5741	* unmapping. So, we create an async unmap operation here.
5742	*
5743	* Now, when creating or restoring a saved state this wont work very
5744	* well since we may want to restore all guest RAM + a little something.
5745	* So, we have to do the unmap synchronously. Fortunately for us
5746	* though, during these operations the other virtual CPUs are inactive
5747	* and it should be safe to do this.
5748	*/
5749	/** @todo Eventually we should lock all memory when used and do
5750	* map+unmap as one kernel call without any rendezvous or
5751	* other precautions. */
5752	if (pVM->pgm.s.ChunkR3Map.c + 1 >= pVM->pgm.s.ChunkR3Map.cMax)
5753	{
5754	switch (VMR3GetState(pVM))
5755	{
5756	case VMSTATE_LOADING:
5757	case VMSTATE_SAVING:
5758	{
5759	PVMCPU pVCpu = VMMGetCpu(pVM);
5760	if ( pVCpu
5761	&& pVM->pgm.s.cDeprecatedPageLocks == 0)
5762	{
5763	pgmR3PhysUnmapChunkRendezvous(pVM, pVCpu, NULL);
5764	break;
5765	}
5766	}
5767	RT_FALL_THRU();
5768	default:
5769	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysUnmapChunk, 1, pVM);
5770	AssertRC(rc);
5771	break;
5772	}
5773	}
5774
5775	/*
5776	* Update the tree. We must do this after any unmapping to make sure
5777	* the chunk we're going to return isn't unmapped by accident.
5778	*/
5779	AssertPtr(Req.pvR3);
5780	bool fRc = RTAvlU32Insert(&pVM->pgm.s.ChunkR3Map.pTree, &pChunk->Core);
5781	AssertRelease(fRc);
5782	pVM->pgm.s.ChunkR3Map.c++;
5783	pVM->pgm.s.cMappedChunks++;
5784	}
5785	else
5786	{
5787	/** @todo this may fail because of /proc/sys/vm/max_map_count, so we
5788	* should probably restrict ourselves on linux. */
5789	AssertRC(rc);
5790	MMR3HeapFree(pChunk);
5791	pChunk = NULL;
5792	}
5793
5794	*ppChunk = pChunk;
5795	return rc;
5796	}
5797
5798
5799	/**
5800	* Invalidates the TLB for the ring-3 mapping cache.
5801	*
5802	* @param pVM The cross context VM structure.
5803	*/
5804	VMMR3DECL(void) PGMR3PhysChunkInvalidateTLB(PVM pVM)
5805	{
5806	PGM_LOCK_VOID(pVM);
5807	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
5808	{
5809	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk = NIL_GMM_CHUNKID;
5810	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk = NULL;
5811	}
5812	/* The page map TLB references chunks, so invalidate that one too. */
5813	pgmPhysInvalidatePageMapTLB(pVM);
5814	PGM_UNLOCK(pVM);
5815	}
5816
5817
5818	/**
5819	* Response to VM_FF_PGM_NEED_HANDY_PAGES and helper for pgmPhysEnsureHandyPage.
5820	*
5821	* This function will also work the VM_FF_PGM_NO_MEMORY force action flag, to
5822	* signal and clear the out of memory condition. When called, this API is used
5823	* to try clear the condition when the user wants to resume.
5824	*
5825	* @returns The following VBox status codes.
5826	* @retval VINF_SUCCESS on success. FFs cleared.
5827	* @retval VINF_EM_NO_MEMORY if we're out of memory. The FF is not cleared in
5828	* this case and it gets accompanied by VM_FF_PGM_NO_MEMORY.
5829	*
5830	* @param pVM The cross context VM structure.
5831	*
5832	* @remarks The VINF_EM_NO_MEMORY status is for the benefit of the FF processing
5833	* in EM.cpp and shouldn't be propagated outside TRPM, HM, EM and
5834	* pgmPhysEnsureHandyPage. There is one exception to this in the \#PF
5835	* handler.
5836	*/
5837	VMMR3DECL(int) PGMR3PhysAllocateHandyPages(PVM pVM)
5838	{
5839	PGM_LOCK_VOID(pVM);
5840
5841	/*
5842	* Allocate more pages, noting down the index of the first new page.
5843	*/
5844	uint32_t iClear = pVM->pgm.s.cHandyPages;
5845	AssertMsgReturn(iClear <= RT_ELEMENTS(pVM->pgm.s.aHandyPages), ("%d", iClear), VERR_PGM_HANDY_PAGE_IPE);
5846	Log(("PGMR3PhysAllocateHandyPages: %d -> %d\n", iClear, RT_ELEMENTS(pVM->pgm.s.aHandyPages)));
5847	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
5848	/** @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) */
5849	if ( rc == VERR_GMM_HIT_VM_ACCOUNT_LIMIT
5850	&& pVM->pgm.s.cHandyPages > 0)
5851	{
5852	/* Still handy pages left, so don't panic. */
5853	rc = VINF_SUCCESS;
5854	}
5855
5856	if (RT_SUCCESS(rc))
5857	{
5858	AssertMsg(rc == VINF_SUCCESS, ("%Rrc\n", rc));
5859	Assert(pVM->pgm.s.cHandyPages > 0);
5860	#ifdef VBOX_STRICT
5861	uint32_t i;
5862	for (i = iClear; i < pVM->pgm.s.cHandyPages; i++)
5863	if ( pVM->pgm.s.aHandyPages[i].idPage == NIL_GMM_PAGEID
5864	\|\| pVM->pgm.s.aHandyPages[i].idSharedPage != NIL_GMM_PAGEID
5865	\|\| (pVM->pgm.s.aHandyPages[i].HCPhysGCPhys & PAGE_OFFSET_MASK))
5866	break;
5867	if (i != pVM->pgm.s.cHandyPages)
5868	{
5869	RTAssertMsg1Weak(NULL, __LINE__, __FILE__, __FUNCTION__);
5870	RTAssertMsg2Weak("i=%d iClear=%d cHandyPages=%d\n", i, iClear, pVM->pgm.s.cHandyPages);
5871	for (uint32_t j = iClear; j < pVM->pgm.s.cHandyPages; j++)
5872	RTAssertMsg2Add("%03d: idPage=%d HCPhysGCPhys=%RHp idSharedPage=%d%s\n", j,
5873	pVM->pgm.s.aHandyPages[j].idPage,
5874	pVM->pgm.s.aHandyPages[j].HCPhysGCPhys,
5875	pVM->pgm.s.aHandyPages[j].idSharedPage,
5876	j == i ? " <---" : "");
5877	RTAssertPanic();
5878	}
5879	#endif
5880	}
5881	else
5882	{
5883	/*
5884	* We should never get here unless there is a genuine shortage of
5885	* memory (or some internal error). Flag the error so the VM can be
5886	* suspended ASAP and the user informed. If we're totally out of
5887	* handy pages we will return failure.
5888	*/
5889	/* Report the failure. */
5890	LogRel(("PGM: Failed to procure handy pages; rc=%Rrc cHandyPages=%#x\n"
5891	" cAllPages=%#x cPrivatePages=%#x cSharedPages=%#x cZeroPages=%#x\n",
5892	rc, pVM->pgm.s.cHandyPages,
5893	pVM->pgm.s.cAllPages, pVM->pgm.s.cPrivatePages, pVM->pgm.s.cSharedPages, pVM->pgm.s.cZeroPages));
5894
5895	if ( rc != VERR_NO_MEMORY
5896	&& rc != VERR_NO_PHYS_MEMORY
5897	&& rc != VERR_LOCK_FAILED)
5898	for (uint32_t i = 0; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
5899	{
5900	LogRel(("PGM: aHandyPages[#%#04x] = {.HCPhysGCPhys=%RHp, .idPage=%#08x, .idSharedPage=%#08x}\n",
5901	i, pVM->pgm.s.aHandyPages[i].HCPhysGCPhys, pVM->pgm.s.aHandyPages[i].idPage,
5902	pVM->pgm.s.aHandyPages[i].idSharedPage));
5903	uint32_t const idPage = pVM->pgm.s.aHandyPages[i].idPage;
5904	if (idPage != NIL_GMM_PAGEID)
5905	{
5906	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
5907	pRam;
5908	pRam = pRam->pNextR3)
5909	{
5910	uint32_t const cPages = pRam->cb >> PAGE_SHIFT;
5911	for (uint32_t iPage = 0; iPage < cPages; iPage++)
5912	if (PGM_PAGE_GET_PAGEID(&pRam->aPages[iPage]) == idPage)
5913	LogRel(("PGM: Used by %RGp %R[pgmpage] (%s)\n",
5914	pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pRam->aPages[iPage], pRam->pszDesc));
5915	}
5916	}
5917	}
5918
5919	if (rc == VERR_NO_MEMORY)
5920	{
5921	uint64_t cbHostRamAvail = 0;
5922	int rc2 = RTSystemQueryAvailableRam(&cbHostRamAvail);
5923	if (RT_SUCCESS(rc2))
5924	LogRel(("Host RAM: %RU64MB available\n", cbHostRamAvail / _1M));
5925	else
5926	LogRel(("Cannot determine the amount of available host memory\n"));
5927	}
5928
5929	/* Set the FFs and adjust rc. */
5930	VM_FF_SET(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
5931	VM_FF_SET(pVM, VM_FF_PGM_NO_MEMORY);
5932	if ( rc == VERR_NO_MEMORY
5933	\|\| rc == VERR_NO_PHYS_MEMORY
5934	\|\| rc == VERR_LOCK_FAILED)
5935	rc = VINF_EM_NO_MEMORY;
5936	}
5937
5938	PGM_UNLOCK(pVM);
5939	return rc;
5940	}
5941
5942
5943	/*********************************************************************************************************************************
5944	* Other Stuff *
5945	*********************************************************************************************************************************/
5946
5947	/**
5948	* Sets the Address Gate 20 state.
5949	*
5950	* @param pVCpu The cross context virtual CPU structure.
5951	* @param fEnable True if the gate should be enabled.
5952	* False if the gate should be disabled.
5953	*/
5954	VMMDECL(void) PGMR3PhysSetA20(PVMCPU pVCpu, bool fEnable)
5955	{
5956	LogFlow(("PGMR3PhysSetA20 %d (was %d)\n", fEnable, pVCpu->pgm.s.fA20Enabled));
5957	if (pVCpu->pgm.s.fA20Enabled != fEnable)
5958	{
5959	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX
5960	PCCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
5961	if ( CPUMIsGuestInVmxRootMode(pCtx)
5962	&& !fEnable)
5963	{
5964	Log(("Cannot enter A20M mode while in VMX root mode\n"));
5965	return;
5966	}
5967	#endif
5968	pVCpu->pgm.s.fA20Enabled = fEnable;
5969	pVCpu->pgm.s.GCPhysA20Mask = ~((RTGCPHYS)!fEnable << 20);
5970	if (VM_IS_NEM_ENABLED(pVCpu->CTX_SUFF(pVM)))
5971	NEMR3NotifySetA20(pVCpu, fEnable);
5972	#ifdef PGM_WITH_A20
5973	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
5974	pgmR3RefreshShadowModeAfterA20Change(pVCpu);
5975	HMFlushTlb(pVCpu);
5976	#endif
5977	IEMTlbInvalidateAllPhysical(pVCpu);
5978	STAM_REL_COUNTER_INC(&pVCpu->pgm.s.cA20Changes);
5979	}
5980	}
5981

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PGMPhys.cpp@ 93444

Download in other formats: