PGMPhys.cpp@ 92441

Last change on this file since 92441 was 92441, checked in by vboxsync, 3 years ago
VMM/PGM: Increased the size of ram ranges to cover 16 GiB each. bugref:10122
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 222.6 KB

Line
1	/* $Id: PGMPhys.cpp 92441 2021-11-15 20:50:49Z vboxsync $ */
2	/** @file
3	* PGM - Page Manager and Monitor, Physical Memory Addressing.
4	*/
5
6	/*
7	* Copyright (C) 2006-2020 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*********************************************************************************************************************************
20	* Header Files *
21	*********************************************************************************************************************************/
22	#define LOG_GROUP LOG_GROUP_PGM_PHYS
23	#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].idPage = NIL_GMM_PAGEID;
1445	break;
1446	}
1447	if (pVM->pgm.s.aHandyPages[i].idSharedPage == idPage)
1448	{
1449	pVM->pgm.s.aHandyPages[i].idSharedPage = NIL_GMM_PAGEID;
1450	break;
1451	}
1452	}
1453
1454	/*
1455	* Push it onto the page array.
1456	*/
1457	uint32_t iPage = *pcPendingPages;
1458	Assert(iPage < PGMPHYS_FREE_PAGE_BATCH_SIZE);
1459	*pcPendingPages += 1;
1460
1461	pReq->aPages[iPage].idPage = idPage;
1462
1463	if (iPage + 1 < PGMPHYS_FREE_PAGE_BATCH_SIZE)
1464	return VINF_SUCCESS;
1465
1466	/*
1467	* Flush the pages.
1468	*/
1469	int rc = GMMR3FreePagesPerform(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE);
1470	if (RT_SUCCESS(rc))
1471	{
1472	GMMR3FreePagesRePrep(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1473	*pcPendingPages = 0;
1474	}
1475	return rc;
1476	}
1477
1478
1479	/**
1480	* Frees a range of pages, replacing them with ZERO pages of the specified type.
1481	*
1482	* @returns VBox status code.
1483	* @param pVM The cross context VM structure.
1484	* @param pRam The RAM range in which the pages resides.
1485	* @param GCPhys The address of the first page.
1486	* @param GCPhysLast The address of the last page.
1487	* @param pvMmio2 Pointer to the ring-3 mapping of any MMIO2 memory that
1488	* will replace the pages we're freeing up.
1489	*/
1490	static int pgmR3PhysFreePageRange(PVM pVM, PPGMRAMRANGE pRam, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast, void *pvMmio2)
1491	{
1492	PGM_LOCK_ASSERT_OWNER(pVM);
1493
1494	#ifdef VBOX_WITH_PGM_NEM_MODE
1495	/*
1496	* In simplified memory mode we don't actually free the memory,
1497	* we just unmap it and let NEM do any unlocking of it.
1498	*/
1499	if (pVM->pgm.s.fNemMode)
1500	{
1501	Assert(VM_IS_NEM_ENABLED(pVM));
1502	uint32_t const fNemNotify = (pvMmio2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0) \| NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE;
1503	uint8_t u2State = 0; /* (We don't support UINT8_MAX here.) */
1504	int rc = NEMR3NotifyPhysMmioExMapEarly(pVM, GCPhys, GCPhysLast - GCPhys + 1, fNemNotify,
1505	pRam->pvR3 ? (uint8_t *)pRam->pvR3 + GCPhys - pRam->GCPhys : NULL,
1506	pvMmio2, &u2State, NULL /puNemRange/);
1507	AssertLogRelRCReturn(rc, rc);
1508
1509	/* Iterate the pages. */
1510	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
1511	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
1512	while (cPagesLeft-- > 0)
1513	{
1514	rc = pgmPhysFreePage(pVM, NULL, NULL, pPageDst, GCPhys, PGMPAGETYPE_MMIO);
1515	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
1516
1517	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO);
1518	PGM_PAGE_SET_NEM_STATE(pPageDst, u2State);
1519
1520	GCPhys += PAGE_SIZE;
1521	pPageDst++;
1522	}
1523	return rc;
1524	}
1525	#else /* !VBOX_WITH_PGM_NEM_MODE */
1526	RT_NOREF(pvMmio2);
1527	#endif /* !VBOX_WITH_PGM_NEM_MODE */
1528
1529	/*
1530	* Regular mode.
1531	*/
1532	/* Prepare. */
1533	uint32_t cPendingPages = 0;
1534	PGMMFREEPAGESREQ pReq;
1535	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1536	AssertLogRelRCReturn(rc, rc);
1537
1538	#ifdef VBOX_WITH_NATIVE_NEM
1539	/* Tell NEM up-front. */
1540	uint8_t u2State = UINT8_MAX;
1541	if (VM_IS_NEM_ENABLED(pVM))
1542	{
1543	uint32_t const fNemNotify = (pvMmio2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0) \| NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE;
1544	rc = NEMR3NotifyPhysMmioExMapEarly(pVM, GCPhys, GCPhysLast - GCPhys + 1, fNemNotify, NULL, pvMmio2,
1545	&u2State, NULL /puNemRange/);
1546	AssertLogRelRCReturnStmt(rc, GMMR3FreePagesCleanup(pReq), rc);
1547	}
1548	#endif
1549
1550	/* Iterate the pages. */
1551	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
1552	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
1553	while (cPagesLeft-- > 0)
1554	{
1555	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys, PGMPAGETYPE_MMIO);
1556	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
1557
1558	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO);
1559	#ifdef VBOX_WITH_NATIVE_NEM
1560	if (u2State != UINT8_MAX)
1561	PGM_PAGE_SET_NEM_STATE(pPageDst, u2State);
1562	#endif
1563
1564	GCPhys += PAGE_SIZE;
1565	pPageDst++;
1566	}
1567
1568	/* Finish pending and cleanup. */
1569	if (cPendingPages)
1570	{
1571	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1572	AssertLogRelRCReturn(rc, rc);
1573	}
1574	GMMR3FreePagesCleanup(pReq);
1575
1576	return rc;
1577	}
1578
1579
1580	/**
1581	* PGMR3PhysRegisterRam worker that initializes and links a RAM range.
1582	*
1583	* In NEM mode, this will allocate the pages backing the RAM range and this may
1584	* fail. NEM registration may also fail. (In regular HM mode it won't fail.)
1585	*
1586	* @returns VBox status code.
1587	* @param pVM The cross context VM structure.
1588	* @param pNew The new RAM range.
1589	* @param GCPhys The address of the RAM range.
1590	* @param GCPhysLast The last address of the RAM range.
1591	* @param R0PtrNew Ditto for R0.
1592	* @param fFlags PGM_RAM_RANGE_FLAGS_FLOATING or zero.
1593	* @param pszDesc The description.
1594	* @param pPrev The previous RAM range (for linking).
1595	*/
1596	static int pgmR3PhysInitAndLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast,
1597	RTR0PTR R0PtrNew, uint32_t fFlags, const char *pszDesc, PPGMRAMRANGE pPrev)
1598	{
1599	/*
1600	* Initialize the range.
1601	*/
1602	pNew->pSelfR0 = R0PtrNew;
1603	pNew->GCPhys = GCPhys;
1604	pNew->GCPhysLast = GCPhysLast;
1605	pNew->cb = GCPhysLast - GCPhys + 1;
1606	pNew->pszDesc = pszDesc;
1607	pNew->fFlags = fFlags;
1608	pNew->uNemRange = UINT32_MAX;
1609	pNew->pvR3 = NULL;
1610	pNew->paLSPages = NULL;
1611
1612	uint32_t const cPages = pNew->cb >> PAGE_SHIFT;
1613	#ifdef VBOX_WITH_PGM_NEM_MODE
1614	if (!pVM->pgm.s.fNemMode)
1615	#endif
1616	{
1617	RTGCPHYS iPage = cPages;
1618	while (iPage-- > 0)
1619	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_RAM);
1620
1621	/* Update the page count stats. */
1622	pVM->pgm.s.cZeroPages += cPages;
1623	pVM->pgm.s.cAllPages += cPages;
1624	}
1625	#ifdef VBOX_WITH_PGM_NEM_MODE
1626	else
1627	{
1628	int rc = SUPR3PageAlloc(cPages, &pNew->pvR3);
1629	if (RT_FAILURE(rc))
1630	return rc;
1631
1632	RTGCPHYS iPage = cPages;
1633	while (iPage-- > 0)
1634	PGM_PAGE_INIT(&pNew->aPages[iPage], UINT64_C(0x0000fffffffff000), NIL_GMM_PAGEID,
1635	PGMPAGETYPE_RAM, PGM_PAGE_STATE_ALLOCATED);
1636
1637	/* Update the page count stats. */
1638	pVM->pgm.s.cPrivatePages += cPages;
1639	pVM->pgm.s.cAllPages += cPages;
1640	}
1641	#endif
1642
1643	/*
1644	* Link it.
1645	*/
1646	pgmR3PhysLinkRamRange(pVM, pNew, pPrev);
1647
1648	#ifdef VBOX_WITH_NATIVE_NEM
1649	/*
1650	* Notify NEM now that it has been linked.
1651	*/
1652	if (VM_IS_NEM_ENABLED(pVM))
1653	{
1654	uint8_t u2State = UINT8_MAX;
1655	int rc = NEMR3NotifyPhysRamRegister(pVM, GCPhys, pNew->cb, pNew->pvR3, &u2State, &pNew->uNemRange);
1656	if (RT_SUCCESS(rc))
1657	{
1658	if (u2State != UINT8_MAX)
1659	pgmPhysSetNemStateForPages(&pNew->aPages[0], cPages, u2State);
1660	}
1661	else
1662	pgmR3PhysUnlinkRamRange2(pVM, pNew, pPrev);
1663	return rc;
1664	}
1665	#endif
1666	return VINF_SUCCESS;
1667	}
1668
1669
1670	/**
1671	* PGMR3PhysRegisterRam worker that registers a high chunk.
1672	*
1673	* @returns VBox status code.
1674	* @param pVM The cross context VM structure.
1675	* @param GCPhys The address of the RAM.
1676	* @param cRamPages The number of RAM pages to register.
1677	* @param iChunk The chunk number.
1678	* @param pszDesc The RAM range description.
1679	* @param ppPrev Previous RAM range pointer. In/Out.
1680	*/
1681	static int pgmR3PhysRegisterHighRamChunk(PVM pVM, RTGCPHYS GCPhys, uint32_t cRamPages, uint32_t iChunk,
1682	const char pszDesc, PPGMRAMRANGE ppPrev)
1683	{
1684	const char *pszDescChunk = iChunk == 0
1685	? pszDesc
1686	: MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s (#%u)", pszDesc, iChunk + 1);
1687	AssertReturn(pszDescChunk, VERR_NO_MEMORY);
1688
1689	/*
1690	* Allocate memory for the new chunk.
1691	*/
1692	size_t const cChunkPages = RT_ALIGN_Z(RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cRamPages]), PAGE_SIZE) >> PAGE_SHIFT;
1693	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
1694	AssertReturn(paChunkPages, VERR_NO_TMP_MEMORY);
1695	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
1696	void *pvChunk = NULL;
1697	int rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk, &R0PtrChunk, paChunkPages);
1698	if (RT_SUCCESS(rc))
1699	{
1700	Assert(R0PtrChunk != NIL_RTR0PTR);
1701	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
1702
1703	PPGMRAMRANGE pNew = (PPGMRAMRANGE)pvChunk;
1704
1705	/*
1706	* Ok, init and link the range.
1707	*/
1708	rc = pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhys + ((RTGCPHYS)cRamPages << PAGE_SHIFT) - 1,
1709	R0PtrChunk, PGM_RAM_RANGE_FLAGS_FLOATING, pszDescChunk, *ppPrev);
1710	if (RT_SUCCESS(rc))
1711	*ppPrev = pNew;
1712
1713	if (RT_FAILURE(rc))
1714	SUPR3PageFreeEx(pvChunk, cChunkPages);
1715	}
1716
1717	RTMemTmpFree(paChunkPages);
1718	return rc;
1719	}
1720
1721
1722	/**
1723	* Sets up a range RAM.
1724	*
1725	* This will check for conflicting registrations, make a resource
1726	* reservation for the memory (with GMM), and setup the per-page
1727	* tracking structures (PGMPAGE).
1728	*
1729	* @returns VBox status code.
1730	* @param pVM The cross context VM structure.
1731	* @param GCPhys The physical address of the RAM.
1732	* @param cb The size of the RAM.
1733	* @param pszDesc The description - not copied, so, don't free or change it.
1734	*/
1735	VMMR3DECL(int) PGMR3PhysRegisterRam(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, const char *pszDesc)
1736	{
1737	/*
1738	* Validate input.
1739	*/
1740	Log(("PGMR3PhysRegisterRam: GCPhys=%RGp cb=%RGp pszDesc=%s\n", GCPhys, cb, pszDesc));
1741	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
1742	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
1743	AssertReturn(cb > 0, VERR_INVALID_PARAMETER);
1744	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
1745	AssertMsgReturn(GCPhysLast > GCPhys, ("The range wraps! GCPhys=%RGp cb=%RGp\n", GCPhys, cb), VERR_INVALID_PARAMETER);
1746	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
1747	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1748
1749	PGM_LOCK_VOID(pVM);
1750
1751	/*
1752	* Find range location and check for conflicts.
1753	*/
1754	PPGMRAMRANGE pPrev = NULL;
1755	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
1756	while (pRam && GCPhysLast >= pRam->GCPhys)
1757	{
1758	AssertLogRelMsgReturnStmt( GCPhysLast < pRam->GCPhys
1759	\|\| GCPhys > pRam->GCPhysLast,
1760	("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
1761	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
1762	PGM_UNLOCK(pVM), VERR_PGM_RAM_CONFLICT);
1763
1764	/* next */
1765	pPrev = pRam;
1766	pRam = pRam->pNextR3;
1767	}
1768
1769	/*
1770	* Register it with GMM (the API bitches).
1771	*/
1772	const RTGCPHYS cPages = cb >> PAGE_SHIFT;
1773	int rc = MMR3IncreaseBaseReservation(pVM, cPages);
1774	if (RT_FAILURE(rc))
1775	{
1776	PGM_UNLOCK(pVM);
1777	return rc;
1778	}
1779
1780	if ( GCPhys >= _4G
1781	&& cPages > 256)
1782	{
1783	/*
1784	* The PGMRAMRANGE structures for the high memory can get very big.
1785	* There used to be some limitations on SUPR3PageAllocEx allocation
1786	* sizes, so traditionally we limited this to 16MB chunks. These days
1787	* we do ~64 MB chunks each covering 16GB of guest RAM, making sure
1788	* each range is a multiple of 1GB to enable eager hosts to use 1GB
1789	* pages in NEM mode.
1790	*
1791	* See also pgmR3PhysMmio2CalcChunkCount.
1792	*/
1793	uint32_t const cPagesPerChunk = _4M;
1794	Assert(RT_ALIGN_32(cPagesPerChunk, X86_PD_PAE_SHIFT - X86_PAGE_SHIFT)); /* NEM large page requirement: 1GB pages. */
1795
1796	RTGCPHYS cPagesLeft = cPages;
1797	RTGCPHYS GCPhysChunk = GCPhys;
1798	uint32_t iChunk = 0;
1799	while (cPagesLeft > 0)
1800	{
1801	uint32_t cPagesInChunk = cPagesLeft;
1802	if (cPagesInChunk > cPagesPerChunk)
1803	cPagesInChunk = cPagesPerChunk;
1804
1805	rc = pgmR3PhysRegisterHighRamChunk(pVM, GCPhysChunk, cPagesInChunk, iChunk, pszDesc, &pPrev);
1806	AssertRCReturn(rc, rc);
1807
1808	/* advance */
1809	GCPhysChunk += (RTGCPHYS)cPagesInChunk << PAGE_SHIFT;
1810	cPagesLeft -= cPagesInChunk;
1811	iChunk++;
1812	}
1813	}
1814	else
1815	{
1816	/*
1817	* Allocate, initialize and link the new RAM range.
1818	*/
1819	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
1820	PPGMRAMRANGE pNew;
1821	rc = MMR3HyperAllocOnceNoRel(pVM, cbRamRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
1822	AssertLogRelMsgRCReturn(rc, ("rc=%Rrc cbRamRange=%zu\n", rc, cbRamRange), rc);
1823
1824	rc = pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhysLast, MMHyperCCToR0(pVM, pNew), 0 /fFlags/, pszDesc, pPrev);
1825	AssertLogRelMsgRCReturn(rc, ("rc=%Rrc cbRamRange=%zu\n", rc, cbRamRange), rc);
1826	}
1827	pgmPhysInvalidatePageMapTLB(pVM);
1828
1829	PGM_UNLOCK(pVM);
1830	return rc;
1831	}
1832
1833
1834	/**
1835	* Worker called by PGMR3InitFinalize if we're configured to pre-allocate RAM.
1836	*
1837	* We do this late in the init process so that all the ROM and MMIO ranges have
1838	* been registered already and we don't go wasting memory on them.
1839	*
1840	* @returns VBox status code.
1841	*
1842	* @param pVM The cross context VM structure.
1843	*/
1844	int pgmR3PhysRamPreAllocate(PVM pVM)
1845	{
1846	Assert(pVM->pgm.s.fRamPreAlloc);
1847	Log(("pgmR3PhysRamPreAllocate: enter\n"));
1848	#ifdef VBOX_WITH_PGM_NEM_MODE
1849	AssertLogRelReturn(!pVM->pgm.s.fNemMode, VERR_PGM_NOT_SUPPORTED_FOR_NEM_MODE);
1850	#endif
1851
1852	/*
1853	* Walk the RAM ranges and allocate all RAM pages, halt at
1854	* the first allocation error.
1855	*/
1856	uint64_t cPages = 0;
1857	uint64_t NanoTS = RTTimeNanoTS();
1858	PGM_LOCK_VOID(pVM);
1859	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1860	{
1861	PPGMPAGE pPage = &pRam->aPages[0];
1862	RTGCPHYS GCPhys = pRam->GCPhys;
1863	uint32_t cLeft = pRam->cb >> PAGE_SHIFT;
1864	while (cLeft-- > 0)
1865	{
1866	if (PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1867	{
1868	switch (PGM_PAGE_GET_STATE(pPage))
1869	{
1870	case PGM_PAGE_STATE_ZERO:
1871	{
1872	int rc = pgmPhysAllocPage(pVM, pPage, GCPhys);
1873	if (RT_FAILURE(rc))
1874	{
1875	LogRel(("PGM: RAM Pre-allocation failed at %RGp (in %s) with rc=%Rrc\n", GCPhys, pRam->pszDesc, rc));
1876	PGM_UNLOCK(pVM);
1877	return rc;
1878	}
1879	cPages++;
1880	break;
1881	}
1882
1883	case PGM_PAGE_STATE_BALLOONED:
1884	case PGM_PAGE_STATE_ALLOCATED:
1885	case PGM_PAGE_STATE_WRITE_MONITORED:
1886	case PGM_PAGE_STATE_SHARED:
1887	/* nothing to do here. */
1888	break;
1889	}
1890	}
1891
1892	/* next */
1893	pPage++;
1894	GCPhys += PAGE_SIZE;
1895	}
1896	}
1897	PGM_UNLOCK(pVM);
1898	NanoTS = RTTimeNanoTS() - NanoTS;
1899
1900	LogRel(("PGM: Pre-allocated %llu pages in %llu ms\n", cPages, NanoTS / 1000000));
1901	Log(("pgmR3PhysRamPreAllocate: returns VINF_SUCCESS\n"));
1902	return VINF_SUCCESS;
1903	}
1904
1905
1906	/**
1907	* Checks shared page checksums.
1908	*
1909	* @param pVM The cross context VM structure.
1910	*/
1911	void pgmR3PhysAssertSharedPageChecksums(PVM pVM)
1912	{
1913	#ifdef VBOX_STRICT
1914	PGM_LOCK_VOID(pVM);
1915
1916	if (pVM->pgm.s.cSharedPages > 0)
1917	{
1918	/*
1919	* Walk the ram ranges.
1920	*/
1921	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1922	{
1923	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
1924	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
1925
1926	while (iPage-- > 0)
1927	{
1928	PPGMPAGE pPage = &pRam->aPages[iPage];
1929	if (PGM_PAGE_IS_SHARED(pPage))
1930	{
1931	uint32_t u32Checksum = pPage->s.u2Unused0/* \| ((uint32_t)pPage->s.u2Unused1 << 8)*/;
1932	if (!u32Checksum)
1933	{
1934	RTGCPHYS GCPhysPage = pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT);
1935	void const *pvPage;
1936	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhysPage, &pvPage);
1937	if (RT_SUCCESS(rc))
1938	{
1939	uint32_t u32Checksum2 = RTCrc32(pvPage, PAGE_SIZE);
1940	# if 0
1941	AssertMsg((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum, ("GCPhysPage=%RGp\n", GCPhysPage));
1942	# else
1943	if ((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum)
1944	LogFlow(("shpg %#x @ %RGp %#x [OK]\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
1945	else
1946	AssertMsgFailed(("shpg %#x @ %RGp %#x\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
1947	# endif
1948	}
1949	else
1950	AssertRC(rc);
1951	}
1952	}
1953
1954	} /* for each page */
1955
1956	} /* for each ram range */
1957	}
1958
1959	PGM_UNLOCK(pVM);
1960	#endif /* VBOX_STRICT */
1961	NOREF(pVM);
1962	}
1963
1964
1965	/**
1966	* Resets the physical memory state.
1967	*
1968	* ASSUMES that the caller owns the PGM lock.
1969	*
1970	* @returns VBox status code.
1971	* @param pVM The cross context VM structure.
1972	*/
1973	int pgmR3PhysRamReset(PVM pVM)
1974	{
1975	PGM_LOCK_ASSERT_OWNER(pVM);
1976
1977	/* Reset the memory balloon. */
1978	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
1979	AssertRC(rc);
1980
1981	#ifdef VBOX_WITH_PAGE_SHARING
1982	/* Clear all registered shared modules. */
1983	pgmR3PhysAssertSharedPageChecksums(pVM);
1984	rc = GMMR3ResetSharedModules(pVM);
1985	AssertRC(rc);
1986	#endif
1987	/* Reset counters. */
1988	pVM->pgm.s.cReusedSharedPages = 0;
1989	pVM->pgm.s.cBalloonedPages = 0;
1990
1991	return VINF_SUCCESS;
1992	}
1993
1994
1995	/**
1996	* Resets (zeros) the RAM after all devices and components have been reset.
1997	*
1998	* ASSUMES that the caller owns the PGM lock.
1999	*
2000	* @returns VBox status code.
2001	* @param pVM The cross context VM structure.
2002	*/
2003	int pgmR3PhysRamZeroAll(PVM pVM)
2004	{
2005	PGM_LOCK_ASSERT_OWNER(pVM);
2006
2007	/*
2008	* We batch up pages that should be freed instead of calling GMM for
2009	* each and every one of them.
2010	*/
2011	uint32_t cPendingPages = 0;
2012	PGMMFREEPAGESREQ pReq;
2013	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2014	AssertLogRelRCReturn(rc, rc);
2015
2016	/*
2017	* Walk the ram ranges.
2018	*/
2019	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2020	{
2021	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2022	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2023
2024	if ( !pVM->pgm.s.fRamPreAlloc
2025	#ifdef VBOX_WITH_PGM_NEM_MODE
2026	&& !pVM->pgm.s.fNemMode
2027	#endif
2028	&& pVM->pgm.s.fZeroRamPagesOnReset)
2029	{
2030	/* Replace all RAM pages by ZERO pages. */
2031	while (iPage-- > 0)
2032	{
2033	PPGMPAGE pPage = &pRam->aPages[iPage];
2034	switch (PGM_PAGE_GET_TYPE(pPage))
2035	{
2036	case PGMPAGETYPE_RAM:
2037	/* Do not replace pages part of a 2 MB continuous range
2038	with zero pages, but zero them instead. */
2039	if ( PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE
2040	\|\| PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE_DISABLED)
2041	{
2042	void *pvPage;
2043	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2044	AssertLogRelRCReturn(rc, rc);
2045	ASMMemZeroPage(pvPage);
2046	}
2047	else if (PGM_PAGE_IS_BALLOONED(pPage))
2048	{
2049	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2050	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2051	}
2052	else if (!PGM_PAGE_IS_ZERO(pPage))
2053	{
2054	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2055	PGMPAGETYPE_RAM);
2056	AssertLogRelRCReturn(rc, rc);
2057	}
2058	break;
2059
2060	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2061	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2062	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2063	pRam, true /fDoAccounting/);
2064	break;
2065
2066	case PGMPAGETYPE_MMIO2:
2067	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2068	case PGMPAGETYPE_ROM:
2069	case PGMPAGETYPE_MMIO:
2070	break;
2071	default:
2072	AssertFailed();
2073	}
2074	} /* for each page */
2075	}
2076	else
2077	{
2078	/* Zero the memory. */
2079	while (iPage-- > 0)
2080	{
2081	PPGMPAGE pPage = &pRam->aPages[iPage];
2082	switch (PGM_PAGE_GET_TYPE(pPage))
2083	{
2084	case PGMPAGETYPE_RAM:
2085	switch (PGM_PAGE_GET_STATE(pPage))
2086	{
2087	case PGM_PAGE_STATE_ZERO:
2088	break;
2089
2090	case PGM_PAGE_STATE_BALLOONED:
2091	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2092	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2093	break;
2094
2095	case PGM_PAGE_STATE_SHARED:
2096	case PGM_PAGE_STATE_WRITE_MONITORED:
2097	rc = pgmPhysPageMakeWritable(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2098	AssertLogRelRCReturn(rc, rc);
2099	RT_FALL_THRU();
2100
2101	case PGM_PAGE_STATE_ALLOCATED:
2102	if (pVM->pgm.s.fZeroRamPagesOnReset)
2103	{
2104	void *pvPage;
2105	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2106	AssertLogRelRCReturn(rc, rc);
2107	ASMMemZeroPage(pvPage);
2108	}
2109	break;
2110	}
2111	break;
2112
2113	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2114	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2115	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2116	pRam, true /fDoAccounting/);
2117	break;
2118
2119	case PGMPAGETYPE_MMIO2:
2120	case PGMPAGETYPE_ROM_SHADOW:
2121	case PGMPAGETYPE_ROM:
2122	case PGMPAGETYPE_MMIO:
2123	break;
2124	default:
2125	AssertFailed();
2126
2127	}
2128	} /* for each page */
2129	}
2130
2131	}
2132
2133	/*
2134	* Finish off any pages pending freeing.
2135	*/
2136	if (cPendingPages)
2137	{
2138	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2139	AssertLogRelRCReturn(rc, rc);
2140	}
2141	GMMR3FreePagesCleanup(pReq);
2142	return VINF_SUCCESS;
2143	}
2144
2145
2146	/**
2147	* Frees all RAM during VM termination
2148	*
2149	* ASSUMES that the caller owns the PGM lock.
2150	*
2151	* @returns VBox status code.
2152	* @param pVM The cross context VM structure.
2153	*/
2154	int pgmR3PhysRamTerm(PVM pVM)
2155	{
2156	PGM_LOCK_ASSERT_OWNER(pVM);
2157
2158	/* Reset the memory balloon. */
2159	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2160	AssertRC(rc);
2161
2162	#ifdef VBOX_WITH_PAGE_SHARING
2163	/*
2164	* Clear all registered shared modules.
2165	*/
2166	pgmR3PhysAssertSharedPageChecksums(pVM);
2167	rc = GMMR3ResetSharedModules(pVM);
2168	AssertRC(rc);
2169
2170	/*
2171	* Flush the handy pages updates to make sure no shared pages are hiding
2172	* in there. (Not unlikely if the VM shuts down, apparently.)
2173	*/
2174	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_FLUSH_HANDY_PAGES, 0, NULL);
2175	#endif
2176
2177	/*
2178	* We batch up pages that should be freed instead of calling GMM for
2179	* each and every one of them.
2180	*/
2181	uint32_t cPendingPages = 0;
2182	PGMMFREEPAGESREQ pReq;
2183	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2184	AssertLogRelRCReturn(rc, rc);
2185
2186	/*
2187	* Walk the ram ranges.
2188	*/
2189	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2190	{
2191	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2192	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2193
2194	while (iPage-- > 0)
2195	{
2196	PPGMPAGE pPage = &pRam->aPages[iPage];
2197	switch (PGM_PAGE_GET_TYPE(pPage))
2198	{
2199	case PGMPAGETYPE_RAM:
2200	/* Free all shared pages. Private pages are automatically freed during GMM VM cleanup. */
2201	/** @todo change this to explicitly free private pages here. */
2202	if (PGM_PAGE_IS_SHARED(pPage))
2203	{
2204	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2205	PGMPAGETYPE_RAM);
2206	AssertLogRelRCReturn(rc, rc);
2207	}
2208	break;
2209
2210	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2211	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO:
2212	case PGMPAGETYPE_MMIO2:
2213	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2214	case PGMPAGETYPE_ROM:
2215	case PGMPAGETYPE_MMIO:
2216	break;
2217	default:
2218	AssertFailed();
2219	}
2220	} /* for each page */
2221	}
2222
2223	/*
2224	* Finish off any pages pending freeing.
2225	*/
2226	if (cPendingPages)
2227	{
2228	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2229	AssertLogRelRCReturn(rc, rc);
2230	}
2231	GMMR3FreePagesCleanup(pReq);
2232	return VINF_SUCCESS;
2233	}
2234
2235
2236
2237	/*********************************************************************************************************************************
2238	* MMIO *
2239	*********************************************************************************************************************************/
2240
2241	/**
2242	* This is the interface IOM is using to register an MMIO region.
2243	*
2244	* It will check for conflicts and ensure that a RAM range structure
2245	* is present before calling the PGMR3HandlerPhysicalRegister API to
2246	* register the callbacks.
2247	*
2248	* @returns VBox status code.
2249	*
2250	* @param pVM The cross context VM structure.
2251	* @param GCPhys The start of the MMIO region.
2252	* @param cb The size of the MMIO region.
2253	* @param hType The physical access handler type registration.
2254	* @param pvUserR3 The user argument for R3.
2255	* @param pvUserR0 The user argument for R0.
2256	* @param pvUserRC The user argument for RC.
2257	* @param pszDesc The description of the MMIO region.
2258	*/
2259	VMMR3DECL(int) PGMR3PhysMMIORegister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMPHYSHANDLERTYPE hType,
2260	RTR3PTR pvUserR3, RTR0PTR pvUserR0, RTRCPTR pvUserRC, const char *pszDesc)
2261	{
2262	/*
2263	* Assert on some assumption.
2264	*/
2265	VM_ASSERT_EMT(pVM);
2266	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2267	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2268	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2269	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2270	Assert(((PPGMPHYSHANDLERTYPEINT)MMHyperHeapOffsetToPtr(pVM, hType))->enmKind == PGMPHYSHANDLERKIND_MMIO);
2271
2272	int rc = PGM_LOCK(pVM);
2273	AssertRCReturn(rc, rc);
2274
2275	/*
2276	* Make sure there's a RAM range structure for the region.
2277	*/
2278	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2279	bool fRamExists = false;
2280	PPGMRAMRANGE pRamPrev = NULL;
2281	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2282	while (pRam && GCPhysLast >= pRam->GCPhys)
2283	{
2284	if ( GCPhysLast >= pRam->GCPhys
2285	&& GCPhys <= pRam->GCPhysLast)
2286	{
2287	/* Simplification: all within the same range. */
2288	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
2289	&& GCPhysLast <= pRam->GCPhysLast,
2290	("%RGp-%RGp (MMIO/%s) falls partly outside %RGp-%RGp (%s)\n",
2291	GCPhys, GCPhysLast, pszDesc,
2292	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2293	PGM_UNLOCK(pVM),
2294	VERR_PGM_RAM_CONFLICT);
2295
2296	/* Check that it's all RAM or MMIO pages. */
2297	PCPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2298	uint32_t cLeft = cb >> PAGE_SHIFT;
2299	while (cLeft-- > 0)
2300	{
2301	AssertLogRelMsgReturnStmt( PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM
2302	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO,
2303	("%RGp-%RGp (MMIO/%s): %RGp is not a RAM or MMIO page - type=%d desc=%s\n",
2304	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, PGM_PAGE_GET_TYPE(pPage), pRam->pszDesc),
2305	PGM_UNLOCK(pVM),
2306	VERR_PGM_RAM_CONFLICT);
2307	pPage++;
2308	}
2309
2310	/* Looks good. */
2311	fRamExists = true;
2312	break;
2313	}
2314
2315	/* next */
2316	pRamPrev = pRam;
2317	pRam = pRam->pNextR3;
2318	}
2319	PPGMRAMRANGE pNew;
2320	if (fRamExists)
2321	{
2322	pNew = NULL;
2323
2324	/*
2325	* Make all the pages in the range MMIO/ZERO pages, freeing any
2326	* RAM pages currently mapped here. This might not be 100% correct
2327	* for PCI memory, but we're doing the same thing for MMIO2 pages.
2328	*/
2329	rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, NULL);
2330	AssertRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
2331
2332	/* Force a PGM pool flush as guest ram references have been changed. */
2333	/** @todo not entirely SMP safe; assuming for now the guest takes
2334	* care of this internally (not touch mapped mmio while changing the
2335	* mapping). */
2336	PVMCPU pVCpu = VMMGetCpu(pVM);
2337	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2338	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2339	}
2340	else
2341	{
2342	/*
2343	* No RAM range, insert an ad hoc one.
2344	*
2345	* Note that we don't have to tell REM about this range because
2346	* PGMHandlerPhysicalRegisterEx will do that for us.
2347	*/
2348	Log(("PGMR3PhysMMIORegister: Adding ad hoc MMIO range for %RGp-%RGp %s\n", GCPhys, GCPhysLast, pszDesc));
2349
2350	/* Alloc. */
2351	const uint32_t cPages = cb >> PAGE_SHIFT;
2352	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
2353	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), 16, MM_TAG_PGM_PHYS, (void **)&pNew);
2354	AssertLogRelMsgRCReturnStmt(rc, ("cbRamRange=%zu\n", cbRamRange), PGM_UNLOCK(pVM), rc);
2355
2356	#ifdef VBOX_WITH_NATIVE_NEM
2357	/* Notify NEM. */
2358	uint8_t u2State = 0; /* (must have valid state as there can't be anything to preserve) */
2359	if (VM_IS_NEM_ENABLED(pVM))
2360	{
2361	rc = NEMR3NotifyPhysMmioExMapEarly(pVM, GCPhys, cPages << PAGE_SHIFT, 0 /fFlags/, NULL, NULL,
2362	&u2State, &pNew->uNemRange);
2363	AssertLogRelRCReturnStmt(rc, MMHyperFree(pVM, pNew), rc);
2364	}
2365	#endif
2366
2367	/* Initialize the range. */
2368	pNew->pSelfR0 = MMHyperCCToR0(pVM, pNew);
2369	pNew->GCPhys = GCPhys;
2370	pNew->GCPhysLast = GCPhysLast;
2371	pNew->cb = cb;
2372	pNew->pszDesc = pszDesc;
2373	pNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO;
2374	pNew->pvR3 = NULL;
2375	pNew->paLSPages = NULL;
2376
2377	uint32_t iPage = cPages;
2378	while (iPage-- > 0)
2379	{
2380	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_MMIO);
2381	#ifdef VBOX_WITH_NATIVE_NEM
2382	PGM_PAGE_SET_NEM_STATE(&pNew->aPages[iPage], u2State);
2383	#endif
2384	}
2385	Assert(PGM_PAGE_GET_TYPE(&pNew->aPages[0]) == PGMPAGETYPE_MMIO);
2386
2387	/* update the page count stats. */
2388	pVM->pgm.s.cPureMmioPages += cPages;
2389	pVM->pgm.s.cAllPages += cPages;
2390
2391	/* link it */
2392	pgmR3PhysLinkRamRange(pVM, pNew, pRamPrev);
2393	}
2394
2395	/*
2396	* Register the access handler.
2397	*/
2398	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, hType, pvUserR3, pvUserR0, pvUserRC, pszDesc);
2399	if (RT_SUCCESS(rc))
2400	{
2401	#ifdef VBOX_WITH_NATIVE_NEM
2402	/* Late NEM notification. */
2403	if (VM_IS_NEM_ENABLED(pVM))
2404	{
2405	uint32_t const fNemNotify = (fRamExists ? NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE : 0);
2406	rc = NEMR3NotifyPhysMmioExMapLate(pVM, GCPhys, GCPhysLast - GCPhys + 1, fNemNotify,
2407	fRamExists ? (uint8_t *)pRam->pvR3 + (uintptr_t)(GCPhys - pRam->GCPhys) : NULL,
2408	NULL, !fRamExists ? &pRam->uNemRange : NULL);
2409	AssertLogRelRCReturn(rc, rc);
2410	}
2411	#endif
2412	}
2413	/** @todo the phys handler failure handling isn't complete, esp. wrt NEM. */
2414	else if (!fRamExists)
2415	{
2416	pVM->pgm.s.cPureMmioPages -= cb >> PAGE_SHIFT;
2417	pVM->pgm.s.cAllPages -= cb >> PAGE_SHIFT;
2418
2419	/* remove the ad hoc range. */
2420	pgmR3PhysUnlinkRamRange2(pVM, pNew, pRamPrev);
2421	pNew->cb = pNew->GCPhys = pNew->GCPhysLast = NIL_RTGCPHYS;
2422	MMHyperFree(pVM, pRam);
2423	}
2424	pgmPhysInvalidatePageMapTLB(pVM);
2425
2426	PGM_UNLOCK(pVM);
2427	return rc;
2428	}
2429
2430
2431	/**
2432	* This is the interface IOM is using to register an MMIO region.
2433	*
2434	* It will take care of calling PGMHandlerPhysicalDeregister and clean up
2435	* any ad hoc PGMRAMRANGE left behind.
2436	*
2437	* @returns VBox status code.
2438	* @param pVM The cross context VM structure.
2439	* @param GCPhys The start of the MMIO region.
2440	* @param cb The size of the MMIO region.
2441	*/
2442	VMMR3DECL(int) PGMR3PhysMMIODeregister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb)
2443	{
2444	VM_ASSERT_EMT(pVM);
2445
2446	int rc = PGM_LOCK(pVM);
2447	AssertRCReturn(rc, rc);
2448
2449	/*
2450	* First deregister the handler, then check if we should remove the ram range.
2451	*/
2452	rc = PGMHandlerPhysicalDeregister(pVM, GCPhys);
2453	if (RT_SUCCESS(rc))
2454	{
2455	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2456	PPGMRAMRANGE pRamPrev = NULL;
2457	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2458	while (pRam && GCPhysLast >= pRam->GCPhys)
2459	{
2460	/** @todo We're being a bit too careful here. rewrite. */
2461	if ( GCPhysLast == pRam->GCPhysLast
2462	&& GCPhys == pRam->GCPhys)
2463	{
2464	Assert(pRam->cb == cb);
2465
2466	/*
2467	* See if all the pages are dead MMIO pages.
2468	*/
2469	uint32_t const cPages = cb >> PAGE_SHIFT;
2470	bool fAllMMIO = true;
2471	uint32_t iPage = 0;
2472	uint32_t cLeft = cPages;
2473	while (cLeft-- > 0)
2474	{
2475	PPGMPAGE pPage = &pRam->aPages[iPage];
2476	if ( !PGM_PAGE_IS_MMIO_OR_ALIAS(pPage)
2477	/\|\| not-out-of-action later /)
2478	{
2479	fAllMMIO = false;
2480	AssertMsgFailed(("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2481	break;
2482	}
2483	Assert( PGM_PAGE_IS_ZERO(pPage)
2484	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2485	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO);
2486	pPage++;
2487	}
2488	if (fAllMMIO)
2489	{
2490	/*
2491	* Ad-hoc range, unlink and free it.
2492	*/
2493	Log(("PGMR3PhysMMIODeregister: Freeing ad hoc MMIO range for %RGp-%RGp %s\n",
2494	GCPhys, GCPhysLast, pRam->pszDesc));
2495	/** @todo check the ad-hoc flags? */
2496
2497	#ifdef VBOX_WITH_NATIVE_NEM
2498	if (VM_IS_NEM_ENABLED(pVM)) /* Notify REM before we unlink the range. */
2499	{
2500	rc = NEMR3NotifyPhysMmioExUnmap(pVM, GCPhys, GCPhysLast - GCPhys + 1, 0 /fFlags/, NULL, NULL, NULL);
2501	AssertLogRelRCReturn(rc, rc);
2502	}
2503	#endif
2504
2505	pVM->pgm.s.cAllPages -= cPages;
2506	pVM->pgm.s.cPureMmioPages -= cPages;
2507
2508	pgmR3PhysUnlinkRamRange2(pVM, pRam, pRamPrev);
2509	pRam->cb = pRam->GCPhys = pRam->GCPhysLast = NIL_RTGCPHYS;
2510	MMHyperFree(pVM, pRam);
2511	break;
2512	}
2513	}
2514
2515	/*
2516	* Range match? It will all be within one range (see PGMAllHandler.cpp).
2517	*/
2518	if ( GCPhysLast >= pRam->GCPhys
2519	&& GCPhys <= pRam->GCPhysLast)
2520	{
2521	Assert(GCPhys >= pRam->GCPhys);
2522	Assert(GCPhysLast <= pRam->GCPhysLast);
2523
2524	/*
2525	* Turn the pages back into RAM pages.
2526	*/
2527	uint32_t iPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2528	uint32_t cLeft = cb >> PAGE_SHIFT;
2529	while (cLeft--)
2530	{
2531	PPGMPAGE pPage = &pRam->aPages[iPage];
2532	AssertMsg( (PGM_PAGE_IS_MMIO(pPage) && PGM_PAGE_IS_ZERO(pPage))
2533	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2534	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO,
2535	("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2536	if (PGM_PAGE_IS_MMIO_OR_ALIAS(pPage))
2537	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_RAM);
2538	iPage++;
2539	}
2540
2541	#ifdef VBOX_WITH_NATIVE_NEM
2542	/* Notify REM (failure will probably leave things in a non-working state). */
2543	if (VM_IS_NEM_ENABLED(pVM))
2544	{
2545	uint8_t u2State = UINT8_MAX;
2546	rc = NEMR3NotifyPhysMmioExUnmap(pVM, GCPhys, GCPhysLast - GCPhys + 1, NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE,
2547	pRam->pvR3 ? (uint8_t *)pRam->pvR3 + GCPhys - pRam->GCPhys : NULL,
2548	NULL, &u2State);
2549	AssertLogRelRCReturn(rc, rc);
2550	if (u2State != UINT8_MAX)
2551	pgmPhysSetNemStateForPages(&pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT],
2552	cb >> PAGE_SHIFT, u2State);
2553	}
2554	#endif
2555	break;
2556	}
2557
2558	/* next */
2559	pRamPrev = pRam;
2560	pRam = pRam->pNextR3;
2561	}
2562	}
2563
2564	/* Force a PGM pool flush as guest ram references have been changed. */
2565	/** @todo Not entirely SMP safe; assuming for now the guest takes care of
2566	* this internally (not touch mapped mmio while changing the mapping). */
2567	PVMCPU pVCpu = VMMGetCpu(pVM);
2568	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2569	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2570
2571	pgmPhysInvalidatePageMapTLB(pVM);
2572	pgmPhysInvalidRamRangeTlbs(pVM);
2573	PGM_UNLOCK(pVM);
2574	return rc;
2575	}
2576
2577
2578
2579	/*********************************************************************************************************************************
2580	* MMIO2 *
2581	*********************************************************************************************************************************/
2582
2583	/**
2584	* Locate a MMIO2 range.
2585	*
2586	* @returns Pointer to the MMIO2 range.
2587	* @param pVM The cross context VM structure.
2588	* @param pDevIns The device instance owning the region.
2589	* @param iSubDev The sub-device number.
2590	* @param iRegion The region.
2591	* @param hMmio2 Handle to look up. If NIL, use the @a iSubDev and
2592	* @a iRegion.
2593	*/
2594	DECLINLINE(PPGMREGMMIO2RANGE) pgmR3PhysMmio2Find(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev,
2595	uint32_t iRegion, PGMMMIO2HANDLE hMmio2)
2596	{
2597	if (hMmio2 != NIL_PGMMMIO2HANDLE)
2598	{
2599	if (hMmio2 <= RT_ELEMENTS(pVM->pgm.s.apMmio2RangesR3) && hMmio2 != 0)
2600	{
2601	PPGMREGMMIO2RANGE pCur = pVM->pgm.s.apMmio2RangesR3[hMmio2 - 1];
2602	if (pCur && pCur->pDevInsR3 == pDevIns)
2603	{
2604	Assert(pCur->idMmio2 == hMmio2);
2605	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, NULL);
2606	return pCur;
2607	}
2608	Assert(!pCur);
2609	}
2610	for (PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2611	if (pCur->idMmio2 == hMmio2)
2612	{
2613	AssertBreak(pCur->pDevInsR3 == pDevIns);
2614	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, NULL);
2615	return pCur;
2616	}
2617	}
2618	else
2619	{
2620	/*
2621	* Search the list. There shouldn't be many entries.
2622	*/
2623	/** @todo Optimize this lookup! There may now be many entries and it'll
2624	* become really slow when doing MMR3HyperMapMMIO2 and similar. */
2625	for (PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2626	if ( pCur->pDevInsR3 == pDevIns
2627	&& pCur->iRegion == iRegion
2628	&& pCur->iSubDev == iSubDev)
2629	return pCur;
2630	}
2631	return NULL;
2632	}
2633
2634
2635	/**
2636	* Worker for PGMR3PhysMmio2ControlDirtyPageTracking and PGMR3PhysMmio2Map.
2637	*/
2638	static int pgmR3PhysMmio2EnableDirtyPageTracing(PVM pVM, PPGMREGMMIO2RANGE pFirstMmio2)
2639	{
2640	int rc = VINF_SUCCESS;
2641	for (PPGMREGMMIO2RANGE pCurMmio2 = pFirstMmio2; pCurMmio2; pCurMmio2 = pCurMmio2->pNextR3)
2642	{
2643	Assert(!(pCurMmio2->fFlags & PGMREGMMIO2RANGE_F_IS_TRACKING));
2644	int rc2 = pgmHandlerPhysicalExRegister(pVM, pCurMmio2->pPhysHandlerR3, pCurMmio2->RamRange.GCPhys,
2645	pCurMmio2->RamRange.GCPhysLast);
2646	AssertLogRelMsgRC(rc2, ("%#RGp-%#RGp %s failed -> %Rrc\n", pCurMmio2->RamRange.GCPhys, pCurMmio2->RamRange.GCPhysLast,
2647	pCurMmio2->RamRange.pszDesc, rc2));
2648	if (RT_SUCCESS(rc2))
2649	pCurMmio2->fFlags \|= PGMREGMMIO2RANGE_F_IS_TRACKING;
2650	else if (RT_SUCCESS(rc))
2651	rc = rc2;
2652	if (pCurMmio2->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2653	return rc;
2654	}
2655	AssertFailed();
2656	return rc;
2657	}
2658
2659
2660	/**
2661	* Worker for PGMR3PhysMmio2ControlDirtyPageTracking and PGMR3PhysMmio2Unmap.
2662	*/
2663	static int pgmR3PhysMmio2DisableDirtyPageTracing(PVM pVM, PPGMREGMMIO2RANGE pFirstMmio2)
2664	{
2665	for (PPGMREGMMIO2RANGE pCurMmio2 = pFirstMmio2; pCurMmio2; pCurMmio2 = pCurMmio2->pNextR3)
2666	{
2667	if (pCurMmio2->fFlags & PGMREGMMIO2RANGE_F_IS_TRACKING)
2668	{
2669	int rc2 = pgmHandlerPhysicalExDeregister(pVM, pCurMmio2->pPhysHandlerR3);
2670	AssertLogRelMsgRC(rc2, ("%#RGp-%#RGp %s failed -> %Rrc\n", pCurMmio2->RamRange.GCPhys, pCurMmio2->RamRange.GCPhysLast,
2671	pCurMmio2->RamRange.pszDesc, rc2));
2672	pCurMmio2->fFlags &= ~PGMREGMMIO2RANGE_F_IS_TRACKING;
2673	}
2674	if (pCurMmio2->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2675	return VINF_SUCCESS;
2676	}
2677	AssertFailed();
2678	return VINF_SUCCESS;
2679
2680	}
2681
2682
2683	/**
2684	* Calculates the number of chunks
2685	*
2686	* @returns Number of registration chunk needed.
2687	* @param pVM The cross context VM structure.
2688	* @param cb The size of the MMIO/MMIO2 range.
2689	* @param pcPagesPerChunk Where to return the number of pages tracked by each
2690	* chunk. Optional.
2691	* @param pcbChunk Where to return the guest mapping size for a chunk.
2692	*/
2693	static uint16_t pgmR3PhysMmio2CalcChunkCount(PVM pVM, RTGCPHYS cb, uint32_t pcPagesPerChunk, uint32_t pcbChunk)
2694	{
2695	RT_NOREF_PV(pVM); /* without raw mode */
2696
2697	/*
2698	* This is the same calculation as PGMR3PhysRegisterRam does, except we'll be
2699	* needing a few bytes extra the PGMREGMMIO2RANGE structure.
2700	*
2701	* Note! In additions, we've got a 24 bit sub-page range for MMIO2 ranges, leaving
2702	* us with an absolute maximum of 16777215 pages per chunk (close to 64 GB).
2703	*/
2704	uint32_t const cPagesPerChunk = _4M;
2705	Assert(RT_ALIGN_32(cPagesPerChunk, X86_PD_PAE_SHIFT - X86_PAGE_SHIFT)); /* NEM large page requirement: 1GB pages. */
2706	uint32_t const cbChunk = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPagesPerChunk]);
2707	AssertRelease(cPagesPerChunk < _16M);
2708
2709	if (pcbChunk)
2710	*pcbChunk = cbChunk;
2711	if (pcPagesPerChunk)
2712	*pcPagesPerChunk = cPagesPerChunk;
2713
2714	/* Calc the number of chunks we need. */
2715	RTGCPHYS const cPages = cb >> X86_PAGE_SHIFT;
2716	uint16_t cChunks = (uint16_t)((cPages + cPagesPerChunk - 1) / cPagesPerChunk);
2717	AssertRelease((RTGCPHYS)cChunks * cPagesPerChunk >= cPages);
2718	return cChunks;
2719	}
2720
2721
2722	/**
2723	* Worker for PGMR3PhysMMIO2Register that allocates and the PGMREGMMIO2RANGE
2724	* structures and does basic initialization.
2725	*
2726	* Caller must set type specfic members and initialize the PGMPAGE structures.
2727	*
2728	* This was previously also used by PGMR3PhysMmio2PreRegister, a function for
2729	* pre-registering MMIO that was later (6.1) replaced by a new handle based IOM
2730	* interface. The reference to caller and type above is purely historical.
2731	*
2732	* @returns VBox status code.
2733	* @param pVM The cross context VM structure.
2734	* @param pDevIns The device instance owning the region.
2735	* @param iSubDev The sub-device number (internal PCI config number).
2736	* @param iRegion The region number. If the MMIO2 memory is a PCI
2737	* I/O region this number has to be the number of that
2738	* region. Otherwise it can be any number safe
2739	* UINT8_MAX.
2740	* @param cb The size of the region. Must be page aligned.
2741	* @param fFlags PGMPHYS_MMIO2_FLAGS_XXX.
2742	* @param idMmio2 The MMIO2 ID for the first chunk.
2743	* @param pszDesc The description.
2744	* @param ppHeadRet Where to return the pointer to the first
2745	* registration chunk.
2746	*
2747	* @thread EMT
2748	*/
2749	static int pgmR3PhysMmio2Create(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb, uint32_t fFlags,
2750	uint8_t idMmio2, const char pszDesc, PPGMREGMMIO2RANGE ppHeadRet)
2751	{
2752	/*
2753	* Figure out how many chunks we need and of which size.
2754	*/
2755	uint32_t cPagesPerChunk;
2756	uint16_t cChunks = pgmR3PhysMmio2CalcChunkCount(pVM, cb, &cPagesPerChunk, NULL);
2757	AssertReturn(cChunks, VERR_PGM_PHYS_MMIO_EX_IPE);
2758
2759	/*
2760	* Allocate the chunks.
2761	*/
2762	PPGMREGMMIO2RANGE *ppNext = ppHeadRet;
2763	*ppNext = NULL;
2764
2765	int rc = VINF_SUCCESS;
2766	uint32_t cPagesLeft = cb >> X86_PAGE_SHIFT;
2767	for (uint16_t iChunk = 0; iChunk < cChunks && RT_SUCCESS(rc); iChunk++, idMmio2++)
2768	{
2769	/*
2770	* We currently do a single RAM range for the whole thing. This will
2771	* probably have to change once someone needs really large MMIO regions,
2772	* as we will be running into SUPR3PageAllocEx limitations and such.
2773	*/
2774	const uint32_t cPagesTrackedByChunk = RT_MIN(cPagesLeft, cPagesPerChunk);
2775	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPagesTrackedByChunk]);
2776	PPGMREGMMIO2RANGE pNew = NULL;
2777	if ( iChunk + 1 < cChunks
2778	\|\| cbRange >= _1M)
2779	{
2780	/*
2781	* Allocate memory for the registration structure.
2782	*/
2783	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2784	size_t const cbChunk = (1 + cChunkPages + 1) << PAGE_SHIFT;
2785	AssertLogRelBreakStmt(cbChunk == (uint32_t)cbChunk, rc = VERR_OUT_OF_RANGE);
2786	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
2787	AssertBreakStmt(paChunkPages, rc = VERR_NO_TMP_MEMORY);
2788	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
2789	void *pvChunk = NULL;
2790	rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk, &R0PtrChunk, paChunkPages);
2791	AssertLogRelMsgRCBreakStmt(rc, ("rc=%Rrc, cChunkPages=%#zx\n", rc, cChunkPages), RTMemTmpFree(paChunkPages));
2792
2793	Assert(R0PtrChunk != NIL_RTR0PTR);
2794	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
2795
2796	pNew = (PPGMREGMMIO2RANGE)pvChunk;
2797	pNew->RamRange.fFlags = PGM_RAM_RANGE_FLAGS_FLOATING;
2798	pNew->RamRange.pSelfR0 = R0PtrChunk + RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange);
2799
2800	RTMemTmpFree(paChunkPages);
2801	}
2802	/*
2803	* Not so big, do a one time hyper allocation.
2804	*/
2805	else
2806	{
2807	rc = MMR3HyperAllocOnceNoRel(pVM, cbRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
2808	AssertLogRelMsgRCBreak(rc, ("cbRange=%zu\n", cbRange));
2809
2810	/*
2811	* Initialize allocation specific items.
2812	*/
2813	//pNew->RamRange.fFlags = 0;
2814	pNew->RamRange.pSelfR0 = MMHyperCCToR0(pVM, &pNew->RamRange);
2815	}
2816
2817	/*
2818	* Initialize the registration structure (caller does specific bits).
2819	*/
2820	pNew->pDevInsR3 = pDevIns;
2821	//pNew->pvR3 = NULL;
2822	//pNew->pNext = NULL;
2823	if (iChunk == 0)
2824	pNew->fFlags \|= PGMREGMMIO2RANGE_F_FIRST_CHUNK;
2825	if (iChunk + 1 == cChunks)
2826	pNew->fFlags \|= PGMREGMMIO2RANGE_F_LAST_CHUNK;
2827	if (fFlags & PGMPHYS_MMIO2_FLAGS_TRACK_DIRTY_PAGES)
2828	pNew->fFlags \|= PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES;
2829	pNew->iSubDev = iSubDev;
2830	pNew->iRegion = iRegion;
2831	pNew->idSavedState = UINT8_MAX;
2832	pNew->idMmio2 = idMmio2;
2833	//pNew->pPhysHandlerR3 = NULL;
2834	//pNew->paLSPages = NULL;
2835	pNew->RamRange.GCPhys = NIL_RTGCPHYS;
2836	pNew->RamRange.GCPhysLast = NIL_RTGCPHYS;
2837	pNew->RamRange.pszDesc = pszDesc;
2838	pNew->RamRange.cb = pNew->cbReal = (RTGCPHYS)cPagesTrackedByChunk << X86_PAGE_SHIFT;
2839	pNew->RamRange.fFlags \|= PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO_EX;
2840	pNew->RamRange.uNemRange = UINT32_MAX;
2841	//pNew->RamRange.pvR3 = NULL;
2842	//pNew->RamRange.paLSPages = NULL;
2843
2844	*ppNext = pNew;
2845	ASMCompilerBarrier();
2846	cPagesLeft -= cPagesTrackedByChunk;
2847	ppNext = &pNew->pNextR3;
2848
2849	/*
2850	* Pre-allocate a handler if we're tracking dirty pages, unless NEM takes care of this.
2851	*/
2852	if ( (fFlags & PGMPHYS_MMIO2_FLAGS_TRACK_DIRTY_PAGES)
2853	#ifdef VBOX_WITH_PGM_NEM_MODE
2854	&& !NEMR3IsMmio2DirtyPageTrackingSupported(pVM)
2855	#endif
2856	)
2857
2858	{
2859	rc = pgmHandlerPhysicalExCreate(pVM, pVM->pgm.s.hMmio2DirtyPhysHandlerType,
2860	(RTR3PTR)(uintptr_t)idMmio2, idMmio2, idMmio2, pszDesc, &pNew->pPhysHandlerR3);
2861	AssertLogRelMsgRCBreak(rc, ("idMmio2=%zu\n", idMmio2));
2862	}
2863	}
2864	Assert(cPagesLeft == 0);
2865
2866	if (RT_SUCCESS(rc))
2867	{
2868	Assert((*ppHeadRet)->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
2869	return VINF_SUCCESS;
2870	}
2871
2872	/*
2873	* Free floating ranges.
2874	*/
2875	while (*ppHeadRet)
2876	{
2877	PPGMREGMMIO2RANGE pFree = *ppHeadRet;
2878	*ppHeadRet = pFree->pNextR3;
2879
2880	if (pFree->pPhysHandlerR3)
2881	{
2882	pgmHandlerPhysicalExDestroy(pVM, pFree->pPhysHandlerR3);
2883	pFree->pPhysHandlerR3 = NULL;
2884	}
2885
2886	if (pFree->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
2887	{
2888	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[pFree->RamRange.cb >> X86_PAGE_SHIFT]);
2889	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2890	SUPR3PageFreeEx(pFree, cChunkPages);
2891	}
2892	}
2893
2894	return rc;
2895	}
2896
2897
2898	/**
2899	* Common worker PGMR3PhysMmio2PreRegister & PGMR3PhysMMIO2Register that links a
2900	* complete registration entry into the lists and lookup tables.
2901	*
2902	* @param pVM The cross context VM structure.
2903	* @param pNew The new MMIO / MMIO2 registration to link.
2904	*/
2905	static void pgmR3PhysMmio2Link(PVM pVM, PPGMREGMMIO2RANGE pNew)
2906	{
2907	Assert(pNew->idMmio2 != UINT8_MAX);
2908
2909	/*
2910	* Link it into the list (order doesn't matter, so insert it at the head).
2911	*
2912	* Note! The range we're linking may consist of multiple chunks, so we
2913	* have to find the last one.
2914	*/
2915	PPGMREGMMIO2RANGE pLast = pNew;
2916	for (pLast = pNew; ; pLast = pLast->pNextR3)
2917	{
2918	if (pLast->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2919	break;
2920	Assert(pLast->pNextR3);
2921	Assert(pLast->pNextR3->pDevInsR3 == pNew->pDevInsR3);
2922	Assert(pLast->pNextR3->iSubDev == pNew->iSubDev);
2923	Assert(pLast->pNextR3->iRegion == pNew->iRegion);
2924	Assert(pLast->pNextR3->idMmio2 == pLast->idMmio2 + 1);
2925	}
2926
2927	PGM_LOCK_VOID(pVM);
2928
2929	/* Link in the chain of ranges at the head of the list. */
2930	pLast->pNextR3 = pVM->pgm.s.pRegMmioRangesR3;
2931	pVM->pgm.s.pRegMmioRangesR3 = pNew;
2932
2933	/* Insert the MMIO2 range/page IDs. */
2934	uint8_t idMmio2 = pNew->idMmio2;
2935	for (;;)
2936	{
2937	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == NULL);
2938	Assert(pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] == NIL_RTR0PTR);
2939	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = pNew;
2940	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = pNew->RamRange.pSelfR0 - RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange);
2941	if (pNew->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2942	break;
2943	pNew = pNew->pNextR3;
2944	idMmio2++;
2945	}
2946
2947	pgmPhysInvalidatePageMapTLB(pVM);
2948	PGM_UNLOCK(pVM);
2949	}
2950
2951
2952	/**
2953	* Allocate and register an MMIO2 region.
2954	*
2955	* As mentioned elsewhere, MMIO2 is just RAM spelled differently. It's RAM
2956	* associated with a device. It is also non-shared memory with a permanent
2957	* ring-3 mapping and page backing (presently).
2958	*
2959	* A MMIO2 range may overlap with base memory if a lot of RAM is configured for
2960	* the VM, in which case we'll drop the base memory pages. Presently we will
2961	* make no attempt to preserve anything that happens to be present in the base
2962	* memory that is replaced, this is of course incorrect but it's too much
2963	* effort.
2964	*
2965	* @returns VBox status code.
2966	* @retval VINF_SUCCESS on success, *ppv pointing to the R3 mapping of the
2967	* memory.
2968	* @retval VERR_ALREADY_EXISTS if the region already exists.
2969	*
2970	* @param pVM The cross context VM structure.
2971	* @param pDevIns The device instance owning the region.
2972	* @param iSubDev The sub-device number.
2973	* @param iRegion The region number. If the MMIO2 memory is a PCI
2974	* I/O region this number has to be the number of that
2975	* region. Otherwise it can be any number save
2976	* UINT8_MAX.
2977	* @param cb The size of the region. Must be page aligned.
2978	* @param fFlags Reserved for future use, must be zero.
2979	* @param pszDesc The description.
2980	* @param ppv Where to store the pointer to the ring-3 mapping of
2981	* the memory.
2982	* @param phRegion Where to return the MMIO2 region handle. Optional.
2983	* @thread EMT
2984	*/
2985	VMMR3_INT_DECL(int) PGMR3PhysMmio2Register(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb,
2986	uint32_t fFlags, const char pszDesc, void ppv, PGMMMIO2HANDLE phRegion)
2987	{
2988	/*
2989	* Validate input.
2990	*/
2991	AssertPtrReturn(ppv, VERR_INVALID_POINTER);
2992	*ppv = NULL;
2993	if (phRegion)
2994	{
2995	AssertPtrReturn(phRegion, VERR_INVALID_POINTER);
2996	*phRegion = NIL_PGMMMIO2HANDLE;
2997	}
2998	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2999	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3000	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3001	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3002	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
3003	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
3004	AssertReturn(pgmR3PhysMmio2Find(pVM, pDevIns, iSubDev, iRegion, NIL_PGMMMIO2HANDLE) == NULL, VERR_ALREADY_EXISTS);
3005	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3006	AssertReturn(cb, VERR_INVALID_PARAMETER);
3007	AssertReturn(!(fFlags & ~PGMPHYS_MMIO2_FLAGS_VALID_MASK), VERR_INVALID_FLAGS);
3008
3009	const uint32_t cPages = cb >> PAGE_SHIFT;
3010	AssertLogRelReturn(((RTGCPHYS)cPages << PAGE_SHIFT) == cb, VERR_INVALID_PARAMETER);
3011	AssertLogRelReturn(cPages <= (MM_MMIO_64_MAX >> X86_PAGE_SHIFT), VERR_OUT_OF_RANGE);
3012	AssertLogRelReturn(cPages <= PGM_MMIO2_MAX_PAGE_COUNT, VERR_OUT_OF_RANGE);
3013
3014	/*
3015	* For the 2nd+ instance, mangle the description string so it's unique.
3016	*/
3017	if (pDevIns->iInstance > 0) /** @todo Move to PDMDevHlp.cpp and use a real string cache. */
3018	{
3019	pszDesc = MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s [%u]", pszDesc, pDevIns->iInstance);
3020	if (!pszDesc)
3021	return VERR_NO_MEMORY;
3022	}
3023
3024	/*
3025	* Allocate an MMIO2 range ID (not freed on failure).
3026	*
3027	* The zero ID is not used as it could be confused with NIL_GMM_PAGEID, so
3028	* the IDs goes from 1 thru PGM_MMIO2_MAX_RANGES.
3029	*/
3030	unsigned cChunks = pgmR3PhysMmio2CalcChunkCount(pVM, cb, NULL, NULL);
3031
3032	PGM_LOCK_VOID(pVM);
3033	AssertCompile(PGM_MMIO2_MAX_RANGES < 255);
3034	uint8_t const idMmio2 = pVM->pgm.s.cMmio2Regions + 1;
3035	unsigned const cNewMmio2Regions = pVM->pgm.s.cMmio2Regions + cChunks;
3036	if (cNewMmio2Regions > PGM_MMIO2_MAX_RANGES)
3037	{
3038	PGM_UNLOCK(pVM);
3039	AssertLogRelFailedReturn(VERR_PGM_TOO_MANY_MMIO2_RANGES);
3040	}
3041	pVM->pgm.s.cMmio2Regions = cNewMmio2Regions;
3042	PGM_UNLOCK(pVM);
3043
3044	/*
3045	* Try reserve and allocate the backing memory first as this is what is
3046	* most likely to fail.
3047	*/
3048	int rc = VINF_SUCCESS;
3049	#ifdef VBOX_WITH_PGM_NEM_MODE
3050	if (!pVM->pgm.s.fNemMode)
3051	#endif
3052	rc = MMR3AdjustFixedReservation(pVM, cPages, pszDesc);
3053	if (RT_SUCCESS(rc))
3054	{
3055	PSUPPAGE paPages = (PSUPPAGE)RTMemTmpAlloc(cPages * sizeof(SUPPAGE));
3056	if (RT_SUCCESS(rc))
3057	{
3058	void *pvPages;
3059	#ifndef VBOX_WITH_LINEAR_HOST_PHYS_MEM
3060	RTR0PTR pvPagesR0 = NIL_RTR0PTR;
3061	#endif
3062
3063	#ifdef VBOX_WITH_PGM_NEM_MODE
3064	if (!pVM->pgm.s.fNemMode)
3065	#endif
3066	{
3067	#ifndef VBOX_WITH_LINEAR_HOST_PHYS_MEM
3068	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, &pvPagesR0, paPages);
3069	#else
3070	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, NULL /pR0Ptr/, paPages);
3071	#endif
3072	}
3073	#ifdef VBOX_WITH_PGM_NEM_MODE
3074	else
3075	{
3076	rc = SUPR3PageAlloc(cPages, &pvPages);
3077	if (RT_SUCCESS(rc))
3078	for (uint32_t i = 0; i < cPages; i++)
3079	paPages[i].Phys = UINT64_C(0x0000fffffffff000);
3080	}
3081	#endif
3082	if (RT_SUCCESS(rc))
3083	{
3084	memset(pvPages, 0, cPages * PAGE_SIZE);
3085
3086	/*
3087	* Create the registered MMIO range record for it.
3088	*/
3089	PPGMREGMMIO2RANGE pNew;
3090	rc = pgmR3PhysMmio2Create(pVM, pDevIns, iSubDev, iRegion, cb, fFlags, idMmio2, pszDesc, &pNew);
3091	if (RT_SUCCESS(rc))
3092	{
3093	if (phRegion)
3094	phRegion = idMmio2; / The ID of the first chunk. */
3095
3096	uint32_t iSrcPage = 0;
3097	uint8_t pbCurPages = (uint8_t )pvPages;
3098	for (PPGMREGMMIO2RANGE pCur = pNew; pCur; pCur = pCur->pNextR3)
3099	{
3100	pCur->pvR3 = pbCurPages;
3101	#ifndef VBOX_WITH_LINEAR_HOST_PHYS_MEM
3102	pCur->pvR0 = pvPagesR0 + (iSrcPage << PAGE_SHIFT);
3103	#endif
3104	pCur->RamRange.pvR3 = pbCurPages;
3105
3106	uint32_t iDstPage = pCur->RamRange.cb >> X86_PAGE_SHIFT;
3107	while (iDstPage-- > 0)
3108	{
3109	PGM_PAGE_INIT(&pNew->RamRange.aPages[iDstPage],
3110	paPages[iDstPage + iSrcPage].Phys,
3111	PGM_MMIO2_PAGEID_MAKE(idMmio2, iDstPage),
3112	PGMPAGETYPE_MMIO2, PGM_PAGE_STATE_ALLOCATED);
3113	}
3114
3115	/* advance. */
3116	iSrcPage += pCur->RamRange.cb >> X86_PAGE_SHIFT;
3117	pbCurPages += pCur->RamRange.cb;
3118	}
3119
3120	RTMemTmpFree(paPages);
3121
3122	/*
3123	* Update the page count stats, link the registration and we're done.
3124	*/
3125	pVM->pgm.s.cAllPages += cPages;
3126	pVM->pgm.s.cPrivatePages += cPages;
3127
3128	pgmR3PhysMmio2Link(pVM, pNew);
3129
3130	*ppv = pvPages;
3131	return VINF_SUCCESS;
3132	}
3133
3134	SUPR3PageFreeEx(pvPages, cPages);
3135	}
3136	}
3137	RTMemTmpFree(paPages);
3138	MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pszDesc);
3139	}
3140	if (pDevIns->iInstance > 0)
3141	MMR3HeapFree((void *)pszDesc);
3142	return rc;
3143	}
3144
3145
3146	/**
3147	* Deregisters and frees an MMIO2 region.
3148	*
3149	* Any physical access handlers registered for the region must be deregistered
3150	* before calling this function.
3151	*
3152	* @returns VBox status code.
3153	* @param pVM The cross context VM structure.
3154	* @param pDevIns The device instance owning the region.
3155	* @param hMmio2 The MMIO2 handle to deregister, or NIL if all
3156	* regions for the given device is to be deregistered.
3157	*/
3158	VMMR3_INT_DECL(int) PGMR3PhysMmio2Deregister(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3159	{
3160	/*
3161	* Validate input.
3162	*/
3163	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3164	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3165
3166	/*
3167	* The loop here scanning all registrations will make sure that multi-chunk ranges
3168	* get properly deregistered, though it's original purpose was the wildcard iRegion.
3169	*/
3170	PGM_LOCK_VOID(pVM);
3171	int rc = VINF_SUCCESS;
3172	unsigned cFound = 0;
3173	PPGMREGMMIO2RANGE pPrev = NULL;
3174	PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3;
3175	while (pCur)
3176	{
3177	uint32_t const fFlags = pCur->fFlags;
3178	if ( pCur->pDevInsR3 == pDevIns
3179	&& ( hMmio2 == NIL_PGMMMIO2HANDLE
3180	\|\| pCur->idMmio2 == hMmio2))
3181	{
3182	cFound++;
3183
3184	/*
3185	* Unmap it if it's mapped.
3186	*/
3187	if (fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3188	{
3189	int rc2 = PGMR3PhysMmio2Unmap(pVM, pCur->pDevInsR3, pCur->idMmio2, pCur->RamRange.GCPhys);
3190	AssertRC(rc2);
3191	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3192	rc = rc2;
3193	}
3194
3195	/*
3196	* Unlink it
3197	*/
3198	PPGMREGMMIO2RANGE pNext = pCur->pNextR3;
3199	if (pPrev)
3200	pPrev->pNextR3 = pNext;
3201	else
3202	pVM->pgm.s.pRegMmioRangesR3 = pNext;
3203	pCur->pNextR3 = NULL;
3204
3205	uint8_t idMmio2 = pCur->idMmio2;
3206	Assert(idMmio2 <= RT_ELEMENTS(pVM->pgm.s.apMmio2RangesR3));
3207	if (idMmio2 <= RT_ELEMENTS(pVM->pgm.s.apMmio2RangesR3))
3208	{
3209	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == pCur);
3210	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = NULL;
3211	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = NIL_RTR0PTR;
3212	}
3213
3214	/*
3215	* Free the memory.
3216	*/
3217	uint32_t const cPages = pCur->cbReal >> PAGE_SHIFT;
3218	#ifdef VBOX_WITH_PGM_NEM_MODE
3219	if (!pVM->pgm.s.fNemMode)
3220	#endif
3221	{
3222	int rc2 = SUPR3PageFreeEx(pCur->pvR3, cPages);
3223	AssertRC(rc2);
3224	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3225	rc = rc2;
3226
3227	rc2 = MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pCur->RamRange.pszDesc);
3228	AssertRC(rc2);
3229	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3230	rc = rc2;
3231	}
3232	#ifdef VBOX_WITH_PGM_NEM_MODE
3233	else
3234	{
3235	int rc2 = SUPR3PageFree(pCur->pvR3, cPages);
3236	AssertRC(rc2);
3237	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3238	rc = rc2;
3239	}
3240	#endif
3241
3242	if (pCur->pPhysHandlerR3)
3243	{
3244	pgmHandlerPhysicalExDestroy(pVM, pCur->pPhysHandlerR3);
3245	pCur->pPhysHandlerR3 = NULL;
3246	}
3247
3248	/* we're leaking hyper memory here if done at runtime. */
3249	#ifdef VBOX_STRICT
3250	VMSTATE const enmState = VMR3GetState(pVM);
3251	AssertMsg( enmState == VMSTATE_POWERING_OFF
3252	\|\| enmState == VMSTATE_POWERING_OFF_LS
3253	\|\| enmState == VMSTATE_OFF
3254	\|\| enmState == VMSTATE_OFF_LS
3255	\|\| enmState == VMSTATE_DESTROYING
3256	\|\| enmState == VMSTATE_TERMINATED
3257	\|\| enmState == VMSTATE_CREATING
3258	, ("%s\n", VMR3GetStateName(enmState)));
3259	#endif
3260
3261	if (pCur->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
3262	{
3263	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPages]);
3264	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
3265	SUPR3PageFreeEx(pCur, cChunkPages);
3266	}
3267	/*else
3268	{
3269	rc = MMHyperFree(pVM, pCur); - does not work, see the alloc call.
3270	AssertRCReturn(rc, rc);
3271	} */
3272
3273
3274	/* update page count stats */
3275	pVM->pgm.s.cAllPages -= cPages;
3276	pVM->pgm.s.cPrivatePages -= cPages;
3277
3278	/* next */
3279	pCur = pNext;
3280	if (hMmio2 != NIL_PGMMMIO2HANDLE)
3281	{
3282	if (fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3283	break;
3284	hMmio2++;
3285	Assert(pCur->idMmio2 == hMmio2);
3286	Assert(pCur->pDevInsR3 == pDevIns);
3287	Assert(!(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK));
3288	}
3289	}
3290	else
3291	{
3292	pPrev = pCur;
3293	pCur = pCur->pNextR3;
3294	}
3295	}
3296	pgmPhysInvalidatePageMapTLB(pVM);
3297	PGM_UNLOCK(pVM);
3298	return !cFound && hMmio2 != NIL_PGMMMIO2HANDLE ? VERR_NOT_FOUND : rc;
3299	}
3300
3301
3302	/**
3303	* Maps a MMIO2 region.
3304	*
3305	* This is typically done when a guest / the bios / state loading changes the
3306	* PCI config. The replacing of base memory has the same restrictions as during
3307	* registration, of course.
3308	*
3309	* @returns VBox status code.
3310	*
3311	* @param pVM The cross context VM structure.
3312	* @param pDevIns The device instance owning the region.
3313	* @param hMmio2 The handle of the region to map.
3314	* @param GCPhys The guest-physical address to be remapped.
3315	*/
3316	VMMR3_INT_DECL(int) PGMR3PhysMmio2Map(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS GCPhys)
3317	{
3318	/*
3319	* Validate input.
3320	*
3321	* Note! It's safe to walk the MMIO/MMIO2 list since registrations only
3322	* happens during VM construction.
3323	*/
3324	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3325	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3326	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
3327	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3328	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3329	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3330
3331	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3332	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3333	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3334
3335	PPGMREGMMIO2RANGE pLastMmio = pFirstMmio;
3336	RTGCPHYS cbRange = 0;
3337	for (;;)
3338	{
3339	AssertReturn(!(pLastMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED), VERR_WRONG_ORDER);
3340	Assert(pLastMmio->RamRange.GCPhys == NIL_RTGCPHYS);
3341	Assert(pLastMmio->RamRange.GCPhysLast == NIL_RTGCPHYS);
3342	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3343	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3344	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3345	cbRange += pLastMmio->RamRange.cb;
3346	if (pLastMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3347	break;
3348	pLastMmio = pLastMmio->pNextR3;
3349	}
3350
3351	RTGCPHYS GCPhysLast = GCPhys + cbRange - 1;
3352	AssertLogRelReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3353
3354	/*
3355	* Find our location in the ram range list, checking for restriction
3356	* we don't bother implementing yet (partially overlapping, multiple
3357	* ram ranges).
3358	*/
3359	PGM_LOCK_VOID(pVM);
3360
3361	AssertReturnStmt(!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED), PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3362
3363	bool fRamExists = false;
3364	PPGMRAMRANGE pRamPrev = NULL;
3365	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3366	while (pRam && GCPhysLast >= pRam->GCPhys)
3367	{
3368	if ( GCPhys <= pRam->GCPhysLast
3369	&& GCPhysLast >= pRam->GCPhys)
3370	{
3371	/* Completely within? */
3372	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
3373	&& GCPhysLast <= pRam->GCPhysLast,
3374	("%RGp-%RGp (MMIOEx/%s) falls partly outside %RGp-%RGp (%s)\n",
3375	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc,
3376	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3377	PGM_UNLOCK(pVM),
3378	VERR_PGM_RAM_CONFLICT);
3379
3380	/* Check that all the pages are RAM pages. */
3381	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3382	uint32_t cPagesLeft = cbRange >> PAGE_SHIFT;
3383	while (cPagesLeft-- > 0)
3384	{
3385	AssertLogRelMsgReturnStmt(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
3386	("%RGp isn't a RAM page (%d) - mapping %RGp-%RGp (MMIO2/%s).\n",
3387	GCPhys, PGM_PAGE_GET_TYPE(pPage), GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc),
3388	PGM_UNLOCK(pVM),
3389	VERR_PGM_RAM_CONFLICT);
3390	pPage++;
3391	}
3392
3393	/* There can only be one MMIO/MMIO2 chunk matching here! */
3394	AssertLogRelMsgReturnStmt(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK,
3395	("%RGp-%RGp (MMIOEx/%s, flags %#X) consists of multiple chunks whereas the RAM somehow doesn't!\n",
3396	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3397	PGM_UNLOCK(pVM),
3398	VERR_PGM_PHYS_MMIO_EX_IPE);
3399
3400	fRamExists = true;
3401	break;
3402	}
3403
3404	/* next */
3405	pRamPrev = pRam;
3406	pRam = pRam->pNextR3;
3407	}
3408	Log(("PGMR3PhysMmio2Map: %RGp-%RGp fRamExists=%RTbool %s\n", GCPhys, GCPhysLast, fRamExists, pFirstMmio->RamRange.pszDesc));
3409
3410
3411	/*
3412	* Make the changes.
3413	*/
3414	RTGCPHYS GCPhysCur = GCPhys;
3415	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3416	{
3417	pCurMmio->RamRange.GCPhys = GCPhysCur;
3418	pCurMmio->RamRange.GCPhysLast = GCPhysCur + pCurMmio->RamRange.cb - 1;
3419	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3420	{
3421	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3422	break;
3423	}
3424	GCPhysCur += pCurMmio->RamRange.cb;
3425	}
3426
3427	if (fRamExists)
3428	{
3429	/*
3430	* Make all the pages in the range MMIO/ZERO pages, freeing any
3431	* RAM pages currently mapped here. This might not be 100% correct
3432	* for PCI memory, but we're doing the same thing for MMIO2 pages.
3433	*
3434	* We replace these MMIO/ZERO pages with real pages in the MMIO2 case.
3435	*/
3436	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK); /* Only one chunk */
3437	Assert(pFirstMmio->pvR3 == pFirstMmio->RamRange.pvR3);
3438	Assert(pFirstMmio->RamRange.pvR3 != NULL);
3439
3440	#ifdef VBOX_WITH_PGM_NEM_MODE
3441	/* We cannot mix MMIO2 into a RAM range in simplified memory mode because pRam->pvR3 can't point
3442	both at the RAM and MMIO2, so we won't ever write & read from the actual MMIO2 memory if we try. */
3443	AssertLogRelMsgReturn(!pVM->pgm.s.fNemMode, ("%s at %RGp-%RGp\n", pFirstMmio->RamRange.pszDesc, GCPhys, GCPhysLast),
3444	VERR_PGM_NOT_SUPPORTED_FOR_NEM_MODE);
3445	#endif
3446
3447	int rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, pFirstMmio->RamRange.pvR3);
3448	AssertRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
3449
3450	/* Replace the pages, freeing all present RAM pages. */
3451	PPGMPAGE pPageSrc = &pFirstMmio->RamRange.aPages[0];
3452	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3453	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3454	while (cPagesLeft-- > 0)
3455	{
3456	Assert(PGM_PAGE_IS_MMIO(pPageDst));
3457
3458	RTHCPHYS const HCPhys = PGM_PAGE_GET_HCPHYS(pPageSrc);
3459	uint32_t const idPage = PGM_PAGE_GET_PAGEID(pPageSrc);
3460	PGM_PAGE_SET_PAGEID(pVM, pPageDst, idPage);
3461	PGM_PAGE_SET_HCPHYS(pVM, pPageDst, HCPhys);
3462	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO2);
3463	PGM_PAGE_SET_STATE(pVM, pPageDst, PGM_PAGE_STATE_ALLOCATED);
3464	PGM_PAGE_SET_PDE_TYPE(pVM, pPageDst, PGM_PAGE_PDE_TYPE_DONTCARE);
3465	PGM_PAGE_SET_PTE_INDEX(pVM, pPageDst, 0);
3466	PGM_PAGE_SET_TRACKING(pVM, pPageDst, 0);
3467	/* NEM state is set by pgmR3PhysFreePageRange. */
3468
3469	pVM->pgm.s.cZeroPages--;
3470	GCPhys += PAGE_SIZE;
3471	pPageSrc++;
3472	pPageDst++;
3473	}
3474
3475	/* Flush physical page map TLB. */
3476	pgmPhysInvalidatePageMapTLB(pVM);
3477
3478	/* Force a PGM pool flush as guest ram references have been changed. */
3479	/** @todo not entirely SMP safe; assuming for now the guest takes care of
3480	* this internally (not touch mapped mmio while changing the mapping). */
3481	PVMCPU pVCpu = VMMGetCpu(pVM);
3482	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3483	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3484	}
3485	else
3486	{
3487	/*
3488	* No RAM range, insert the ones prepared during registration.
3489	*/
3490	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3491	{
3492	#ifdef VBOX_WITH_NATIVE_NEM
3493	/* Tell NEM and get the new NEM state for the pages. */
3494	uint8_t u2NemState = 0;
3495	if (VM_IS_NEM_ENABLED(pVM))
3496	{
3497	int rc = NEMR3NotifyPhysMmioExMapEarly(pVM, pCurMmio->RamRange.GCPhys,
3498	pCurMmio->RamRange.GCPhysLast - pCurMmio->RamRange.GCPhys + 1,
3499	NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2
3500	\| (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES
3501	? NEM_NOTIFY_PHYS_MMIO_EX_F_TRACK_DIRTY_PAGES : 0),
3502	NULL /pvRam/, pCurMmio->RamRange.pvR3,
3503	&u2NemState, &pCurMmio->RamRange.uNemRange);
3504	AssertLogRelRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
3505	}
3506	#endif
3507
3508	/* Clear the tracking data of pages we're going to reactivate. */
3509	PPGMPAGE pPageSrc = &pCurMmio->RamRange.aPages[0];
3510	uint32_t cPagesLeft = pCurMmio->RamRange.cb >> PAGE_SHIFT;
3511	while (cPagesLeft-- > 0)
3512	{
3513	PGM_PAGE_SET_TRACKING(pVM, pPageSrc, 0);
3514	PGM_PAGE_SET_PTE_INDEX(pVM, pPageSrc, 0);
3515	#ifdef VBOX_WITH_NATIVE_NEM
3516	PGM_PAGE_SET_NEM_STATE(pPageSrc, u2NemState);
3517	#endif
3518	pPageSrc++;
3519	}
3520
3521	/* link in the ram range */
3522	pgmR3PhysLinkRamRange(pVM, &pCurMmio->RamRange, pRamPrev);
3523
3524	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3525	{
3526	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3527	break;
3528	}
3529	pRamPrev = &pCurMmio->RamRange;
3530	}
3531	}
3532
3533	/*
3534	* If the range have dirty page monitoring enabled, enable that.
3535	*
3536	* We ignore failures here for now because if we fail, the whole mapping
3537	* will have to be reversed and we'll end up with nothing at all on the
3538	* screen and a grumpy guest, whereas if we just go on, we'll only have
3539	* visual distortions to gripe about. There will be something in the
3540	* release log.
3541	*/
3542	if ( pFirstMmio->pPhysHandlerR3
3543	&& (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
3544	pgmR3PhysMmio2EnableDirtyPageTracing(pVM, pFirstMmio);
3545
3546	/*
3547	* We're good, set the flags and invalid the mapping TLB.
3548	*/
3549	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3550	{
3551	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_MAPPED;
3552	if (fRamExists)
3553	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_OVERLAPPING;
3554	else
3555	pCurMmio->fFlags &= ~PGMREGMMIO2RANGE_F_OVERLAPPING;
3556	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3557	break;
3558	}
3559	pgmPhysInvalidatePageMapTLB(pVM);
3560
3561	#ifdef VBOX_WITH_NATIVE_NEM
3562	/*
3563	* Late NEM notification.
3564	*/
3565	if (VM_IS_NEM_ENABLED(pVM))
3566	{
3567	int rc;
3568	uint32_t fNemFlags = NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2;
3569	if (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES)
3570	fNemFlags \|= NEM_NOTIFY_PHYS_MMIO_EX_F_TRACK_DIRTY_PAGES;
3571	if (fRamExists)
3572	rc = NEMR3NotifyPhysMmioExMapLate(pVM, GCPhys, GCPhysLast - GCPhys + 1, fNemFlags \| NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE,
3573	pRam->pvR3 ? (uint8_t *)pRam->pvR3 + GCPhys - pRam->GCPhys : NULL, pFirstMmio->pvR3,
3574	NULL /puNemRange/);
3575	else
3576	{
3577	rc = VINF_SUCCESS;
3578	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3579	{
3580	rc = NEMR3NotifyPhysMmioExMapLate(pVM, pCurMmio->RamRange.GCPhys, pCurMmio->RamRange.cb, fNemFlags,
3581	NULL, pCurMmio->RamRange.pvR3, &pCurMmio->RamRange.uNemRange);
3582	if ((pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK) \|\| RT_FAILURE(rc))
3583	break;
3584	}
3585	}
3586	AssertLogRelRCReturnStmt(rc, PGMR3PhysMmio2Unmap(pVM, pDevIns, hMmio2, GCPhys); PGM_UNLOCK(pVM), rc);
3587	}
3588	#endif
3589
3590	PGM_UNLOCK(pVM);
3591
3592	return VINF_SUCCESS;
3593	}
3594
3595
3596	/**
3597	* Unmaps an MMIO2 region.
3598	*
3599	* This is typically done when a guest / the bios / state loading changes the
3600	* PCI config. The replacing of base memory has the same restrictions as during
3601	* registration, of course.
3602	*/
3603	VMMR3_INT_DECL(int) PGMR3PhysMmio2Unmap(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS GCPhys)
3604	{
3605	/*
3606	* Validate input
3607	*/
3608	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3609	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3610	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3611	if (GCPhys != NIL_RTGCPHYS)
3612	{
3613	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3614	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3615	}
3616
3617	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3618	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3619	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3620
3621	int rc = PGM_LOCK(pVM);
3622	AssertRCReturn(rc, rc);
3623
3624	PPGMREGMMIO2RANGE pLastMmio = pFirstMmio;
3625	RTGCPHYS cbRange = 0;
3626	for (;;)
3627	{
3628	AssertReturnStmt(pLastMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED, PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3629	AssertReturnStmt(pLastMmio->RamRange.GCPhys == GCPhys + cbRange \|\| GCPhys == NIL_RTGCPHYS, PGM_UNLOCK(pVM), VERR_INVALID_PARAMETER);
3630	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3631	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3632	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3633	cbRange += pLastMmio->RamRange.cb;
3634	if (pLastMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3635	break;
3636	pLastMmio = pLastMmio->pNextR3;
3637	}
3638
3639	Log(("PGMR3PhysMmio2Unmap: %RGp-%RGp %s\n",
3640	pFirstMmio->RamRange.GCPhys, pLastMmio->RamRange.GCPhysLast, pFirstMmio->RamRange.pszDesc));
3641
3642	uint16_t const fOldFlags = pFirstMmio->fFlags;
3643	AssertReturnStmt(fOldFlags & PGMREGMMIO2RANGE_F_MAPPED, PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3644
3645	/*
3646	* If monitoring dirty pages, we must deregister the handlers first.
3647	*/
3648	if ( pFirstMmio->pPhysHandlerR3
3649	&& (fOldFlags & PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
3650	pgmR3PhysMmio2DisableDirtyPageTracing(pVM, pFirstMmio);
3651
3652	/*
3653	* Unmap it.
3654	*/
3655	int rcRet = VINF_SUCCESS;
3656	#ifdef VBOX_WITH_NATIVE_NEM
3657	uint32_t const fNemFlags = NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2
3658	\| (fOldFlags & PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES
3659	? NEM_NOTIFY_PHYS_MMIO_EX_F_TRACK_DIRTY_PAGES : 0);
3660	#endif
3661	if (fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING)
3662	{
3663	/*
3664	* We've replaced RAM, replace with zero pages.
3665	*
3666	* Note! This is where we might differ a little from a real system, because
3667	* it's likely to just show the RAM pages as they were before the
3668	* MMIO/MMIO2 region was mapped here.
3669	*/
3670	/* Only one chunk allowed when overlapping! */
3671	Assert(fOldFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK);
3672
3673	/* Restore the RAM pages we've replaced. */
3674	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3675	while (pRam->GCPhys > pFirstMmio->RamRange.GCPhysLast)
3676	pRam = pRam->pNextR3;
3677
3678	PPGMPAGE pPageDst = &pRam->aPages[(pFirstMmio->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3679	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3680	pVM->pgm.s.cZeroPages += cPagesLeft; /** @todo not correct for NEM mode */
3681
3682	#ifdef VBOX_WITH_NATIVE_NEM
3683	if (VM_IS_NEM_ENABLED(pVM)) /* Notify NEM. Note! we cannot be here in simple memory mode, see mapping function. */
3684	{
3685	uint8_t u2State = UINT8_MAX;
3686	rc = NEMR3NotifyPhysMmioExUnmap(pVM, pFirstMmio->RamRange.GCPhys, pFirstMmio->RamRange.cb,
3687	fNemFlags \| NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE,
3688	pRam->pvR3
3689	? (uint8_t *)pRam->pvR3 + pFirstMmio->RamRange.GCPhys - pRam->GCPhys : NULL,
3690	pFirstMmio->pvR3, &u2State);
3691	AssertRCStmt(rc, rcRet = rc);
3692	if (u2State != UINT8_MAX)
3693	pgmPhysSetNemStateForPages(pPageDst, cPagesLeft, u2State);
3694	}
3695	#endif
3696
3697	while (cPagesLeft-- > 0)
3698	{
3699	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
3700	pPageDst++;
3701	}
3702
3703	/* Flush physical page map TLB. */
3704	pgmPhysInvalidatePageMapTLB(pVM);
3705
3706	/* Update range state. */
3707	pFirstMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3708	pFirstMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3709	pFirstMmio->fFlags &= ~(PGMREGMMIO2RANGE_F_OVERLAPPING \| PGMREGMMIO2RANGE_F_MAPPED);
3710	}
3711	else
3712	{
3713	/*
3714	* Unlink the chunks related to the MMIO/MMIO2 region.
3715	*/
3716	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3717	{
3718	#ifdef VBOX_WITH_NATIVE_NEM
3719	if (VM_IS_NEM_ENABLED(pVM)) /* Notify NEM. */
3720	{
3721	uint8_t u2State = UINT8_MAX;
3722	rc = NEMR3NotifyPhysMmioExUnmap(pVM, pCurMmio->RamRange.GCPhys, pCurMmio->RamRange.cb, fNemFlags,
3723	NULL, pCurMmio->pvR3, &u2State);
3724	AssertRCStmt(rc, rcRet = rc);
3725	if (u2State != UINT8_MAX)
3726	pgmPhysSetNemStateForPages(pCurMmio->RamRange.aPages, pCurMmio->RamRange.cb >> PAGE_SHIFT, u2State);
3727	}
3728	#endif
3729	pgmR3PhysUnlinkRamRange(pVM, &pCurMmio->RamRange);
3730	pCurMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3731	pCurMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3732	pCurMmio->fFlags &= ~(PGMREGMMIO2RANGE_F_OVERLAPPING \| PGMREGMMIO2RANGE_F_MAPPED);
3733	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3734	break;
3735	}
3736	}
3737
3738	/* Force a PGM pool flush as guest ram references have been changed. */
3739	/** @todo not entirely SMP safe; assuming for now the guest takes care
3740	* of this internally (not touch mapped mmio while changing the
3741	* mapping). */
3742	PVMCPU pVCpu = VMMGetCpu(pVM);
3743	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3744	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3745
3746	pgmPhysInvalidatePageMapTLB(pVM);
3747	pgmPhysInvalidRamRangeTlbs(pVM);
3748
3749	PGM_UNLOCK(pVM);
3750	return rcRet;
3751	}
3752
3753
3754	/**
3755	* Reduces the mapping size of a MMIO2 region.
3756	*
3757	* This is mainly for dealing with old saved states after changing the default
3758	* size of a mapping region. See PGMDevHlpMMIOExReduce and
3759	* PDMPCIDEV::pfnRegionLoadChangeHookR3.
3760	*
3761	* The region must not currently be mapped when making this call. The VM state
3762	* must be state restore or VM construction.
3763	*
3764	* @returns VBox status code.
3765	* @param pVM The cross context VM structure.
3766	* @param pDevIns The device instance owning the region.
3767	* @param hMmio2 The handle of the region to reduce.
3768	* @param cbRegion The new mapping size.
3769	*/
3770	VMMR3_INT_DECL(int) PGMR3PhysMmio2Reduce(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS cbRegion)
3771	{
3772	/*
3773	* Validate input
3774	*/
3775	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3776	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3777	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3778	AssertReturn(cbRegion >= X86_PAGE_SIZE, VERR_INVALID_PARAMETER);
3779	AssertReturn(!(cbRegion & X86_PAGE_OFFSET_MASK), VERR_UNSUPPORTED_ALIGNMENT);
3780	VMSTATE enmVmState = VMR3GetState(pVM);
3781	AssertLogRelMsgReturn( enmVmState == VMSTATE_CREATING
3782	\|\| enmVmState == VMSTATE_LOADING,
3783	("enmVmState=%d (%s)\n", enmVmState, VMR3GetStateName(enmVmState)),
3784	VERR_VM_INVALID_VM_STATE);
3785
3786	int rc = PGM_LOCK(pVM);
3787	AssertRCReturn(rc, rc);
3788
3789	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3790	if (pFirstMmio)
3791	{
3792	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3793	if (!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED))
3794	{
3795	/*
3796	* NOTE! Current implementation does not support multiple ranges.
3797	* Implement when there is a real world need and thus a testcase.
3798	*/
3799	AssertLogRelMsgStmt(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK,
3800	("%s: %#x\n", pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3801	rc = VERR_NOT_SUPPORTED);
3802	if (RT_SUCCESS(rc))
3803	{
3804	/*
3805	* Make the change.
3806	*/
3807	Log(("PGMR3PhysMmio2Reduce: %s changes from %RGp bytes (%RGp) to %RGp bytes.\n",
3808	pFirstMmio->RamRange.pszDesc, pFirstMmio->RamRange.cb, pFirstMmio->cbReal, cbRegion));
3809
3810	AssertLogRelMsgStmt(cbRegion <= pFirstMmio->cbReal,
3811	("%s: cbRegion=%#RGp cbReal=%#RGp\n", pFirstMmio->RamRange.pszDesc, cbRegion, pFirstMmio->cbReal),
3812	rc = VERR_OUT_OF_RANGE);
3813	if (RT_SUCCESS(rc))
3814	{
3815	pFirstMmio->RamRange.cb = cbRegion;
3816	}
3817	}
3818	}
3819	else
3820	rc = VERR_WRONG_ORDER;
3821	}
3822	else
3823	rc = VERR_NOT_FOUND;
3824
3825	PGM_UNLOCK(pVM);
3826	return rc;
3827	}
3828
3829
3830	/**
3831	* Validates @a hMmio2, making sure it belongs to @a pDevIns.
3832	*
3833	* @returns VBox status code.
3834	* @param pVM The cross context VM structure.
3835	* @param pDevIns The device which allegedly owns @a hMmio2.
3836	* @param hMmio2 The handle to validate.
3837	*/
3838	VMMR3_INT_DECL(int) PGMR3PhysMmio2ValidateHandle(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3839	{
3840	/*
3841	* Validate input
3842	*/
3843	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3844	AssertPtrReturn(pDevIns, VERR_INVALID_POINTER);
3845
3846	/*
3847	* Just do this the simple way. No need for locking as this is only taken at
3848	*/
3849	PGM_LOCK_VOID(pVM);
3850	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3851	PGM_UNLOCK(pVM);
3852	AssertReturn(pFirstMmio, VERR_INVALID_HANDLE);
3853	AssertReturn(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, VERR_INVALID_HANDLE);
3854	return VINF_SUCCESS;
3855	}
3856
3857
3858	/**
3859	* Gets the mapping address of an MMIO2 region.
3860	*
3861	* @returns Mapping address, NIL_RTGCPHYS if not mapped or invalid handle.
3862	*
3863	* @param pVM The cross context VM structure.
3864	* @param pDevIns The device owning the MMIO2 handle.
3865	* @param hMmio2 The region handle.
3866	*/
3867	VMMR3_INT_DECL(RTGCPHYS) PGMR3PhysMmio2GetMappingAddress(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3868	{
3869	AssertPtrReturn(pDevIns, NIL_RTGCPHYS);
3870
3871	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3872	AssertReturn(pFirstRegMmio, NIL_RTGCPHYS);
3873
3874	if (pFirstRegMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3875	return pFirstRegMmio->RamRange.GCPhys;
3876	return NIL_RTGCPHYS;
3877	}
3878
3879
3880	/**
3881	* Worker for PGMR3PhysMmio2QueryAndResetDirtyBitmap.
3882	*
3883	* Called holding the PGM lock.
3884	*/
3885	static int pgmR3PhysMmio2QueryAndResetDirtyBitmapLocked(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2,
3886	void *pvBitmap, size_t cbBitmap)
3887	{
3888	/*
3889	* Continue validation.
3890	*/
3891	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3892	AssertReturn(pFirstRegMmio, VERR_INVALID_HANDLE);
3893	AssertReturn( (pFirstRegMmio->fFlags & (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK))
3894	== (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK),
3895	VERR_INVALID_FUNCTION);
3896	AssertReturn(pDevIns == pFirstRegMmio->pDevInsR3, VERR_NOT_OWNER);
3897
3898	RTGCPHYS cbTotal = 0;
3899	uint16_t fTotalDirty = 0;
3900	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio;;)
3901	{
3902	cbTotal += pCur->RamRange.cb; /* Not using cbReal here, because NEM is not in on the creating, only the mapping. */
3903	fTotalDirty \|= pCur->fFlags;
3904	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3905	break;
3906	pCur = pCur->pNextR3;
3907	AssertPtrReturn(pCur, VERR_INTERNAL_ERROR_5);
3908	AssertReturn( (pCur->fFlags & (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK))
3909	== PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES,
3910	VERR_INTERNAL_ERROR_4);
3911	}
3912	size_t const cbTotalBitmap = RT_ALIGN_T(cbTotal, PAGE_SIZE * 64, RTGCPHYS) / PAGE_SIZE / 8;
3913
3914	if (cbBitmap)
3915	{
3916	AssertPtrReturn(pvBitmap, VERR_INVALID_POINTER);
3917	AssertReturn(RT_ALIGN_P(pvBitmap, sizeof(uint64_t)) == pvBitmap, VERR_INVALID_POINTER);
3918	AssertReturn(cbBitmap == cbTotalBitmap, VERR_INVALID_PARAMETER);
3919	}
3920
3921	/*
3922	* Do the work.
3923	*/
3924	int rc = VINF_SUCCESS;
3925	if (pvBitmap)
3926	{
3927	#ifdef VBOX_WITH_PGM_NEM_MODE
3928	if (pFirstRegMmio->pPhysHandlerR3 == NULL)
3929	{
3930	AssertReturn(VM_IS_NEM_ENABLED(pVM), VERR_INTERNAL_ERROR_4);
3931	uint8_t pbBitmap = (uint8_t )pvBitmap;
3932	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
3933	{
3934	size_t const cbBitmapChunk = pCur->RamRange.cb / PAGE_SIZE / 8;
3935	Assert((RTGCPHYS)cbBitmapChunk * PAGE_SIZE * 8 == pCur->RamRange.cb);
3936	int rc2 = NEMR3PhysMmio2QueryAndResetDirtyBitmap(pVM, pCur->RamRange.GCPhys, pCur->RamRange.cb,
3937	pCur->RamRange.uNemRange, pbBitmap, cbBitmapChunk);
3938	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3939	rc = rc2;
3940	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3941	break;
3942	pbBitmap += pCur->RamRange.cb / PAGE_SIZE / 8;
3943	}
3944	}
3945	else
3946	#endif
3947	if (fTotalDirty & PGMREGMMIO2RANGE_F_IS_DIRTY)
3948	{
3949	if ( (pFirstRegMmio->fFlags & (PGMREGMMIO2RANGE_F_MAPPED \| PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
3950	== (PGMREGMMIO2RANGE_F_MAPPED \| PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
3951	{
3952	/*
3953	* Reset each chunk, gathering dirty bits.
3954	*/
3955	RT_BZERO(pvBitmap, cbBitmap); /* simpler for now. */
3956	uint32_t iPageNo = 0;
3957	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
3958	{
3959	if (pCur->fFlags & PGMREGMMIO2RANGE_F_IS_DIRTY)
3960	{
3961	int rc2 = pgmHandlerPhysicalResetMmio2WithBitmap(pVM, pCur->RamRange.GCPhys, pvBitmap, iPageNo);
3962	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3963	rc = rc2;
3964	pCur->fFlags &= ~PGMREGMMIO2RANGE_F_IS_DIRTY;
3965	}
3966	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3967	break;
3968	iPageNo += pCur->RamRange.cb >> PAGE_SHIFT;
3969	}
3970	}
3971	else
3972	{
3973	/*
3974	* If not mapped or tracking is disabled, we return the
3975	* PGMREGMMIO2RANGE_F_IS_DIRTY status for all pages. We cannot
3976	* get more accurate data than that after unmapping or disabling.
3977	*/
3978	RT_BZERO(pvBitmap, cbBitmap);
3979	uint32_t iPageNo = 0;
3980	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
3981	{
3982	if (pCur->fFlags & PGMREGMMIO2RANGE_F_IS_DIRTY)
3983	{
3984	ASMBitSetRange(pvBitmap, iPageNo, iPageNo + (pCur->RamRange.cb >> PAGE_SHIFT));
3985	pCur->fFlags &= ~PGMREGMMIO2RANGE_F_IS_DIRTY;
3986	}
3987	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3988	break;
3989	iPageNo += pCur->RamRange.cb >> PAGE_SHIFT;
3990	}
3991	}
3992	}
3993	/*
3994	* No dirty chunks.
3995	*/
3996	else
3997	RT_BZERO(pvBitmap, cbBitmap);
3998	}
3999	/*
4000	* No bitmap. Reset the region if tracking is currently enabled.
4001	*/
4002	else if ( (pFirstRegMmio->fFlags & (PGMREGMMIO2RANGE_F_MAPPED \| PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
4003	== (PGMREGMMIO2RANGE_F_MAPPED \| PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
4004	{
4005	#ifdef VBOX_WITH_PGM_NEM_MODE
4006	if (pFirstRegMmio->pPhysHandlerR3 == NULL)
4007	{
4008	AssertReturn(VM_IS_NEM_ENABLED(pVM), VERR_INTERNAL_ERROR_4);
4009	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
4010	{
4011	int rc2 = NEMR3PhysMmio2QueryAndResetDirtyBitmap(pVM, pCur->RamRange.GCPhys, pCur->RamRange.cb,
4012	pCur->RamRange.uNemRange, NULL, 0);
4013	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
4014	rc = rc2;
4015	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
4016	break;
4017	}
4018	}
4019	else
4020	#endif
4021	{
4022	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio; pCur; pCur = pCur->pNextR3)
4023	{
4024	pCur->fFlags &= ~PGMREGMMIO2RANGE_F_IS_DIRTY;
4025	int rc2 = PGMHandlerPhysicalReset(pVM, pCur->RamRange.GCPhys);
4026	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
4027	rc = rc2;
4028	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
4029	break;
4030	}
4031	}
4032	}
4033
4034	return rc;
4035	}
4036
4037
4038	/**
4039	* Queries the dirty page bitmap and resets the monitoring.
4040	*
4041	* The PGMPHYS_MMIO2_FLAGS_TRACK_DIRTY_PAGES flag must be specified when
4042	* creating the range for this to work.
4043	*
4044	* @returns VBox status code.
4045	* @retval VERR_INVALID_FUNCTION if not created using
4046	* PGMPHYS_MMIO2_FLAGS_TRACK_DIRTY_PAGES.
4047	* @param pVM The cross context VM structure.
4048	* @param pDevIns The device owning the MMIO2 handle.
4049	* @param hMmio2 The region handle.
4050	* @param pvBitmap The output bitmap. Must be 8-byte aligned. Ignored
4051	* when @a cbBitmap is zero.
4052	* @param cbBitmap The size of the bitmap. Must be the size of the whole
4053	* MMIO2 range, rounded up to the nearest 8 bytes.
4054	* When zero only a reset is done.
4055	*/
4056	VMMR3_INT_DECL(int) PGMR3PhysMmio2QueryAndResetDirtyBitmap(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2,
4057	void *pvBitmap, size_t cbBitmap)
4058	{
4059	/*
4060	* Do some basic validation before grapping the PGM lock and continuing.
4061	*/
4062	AssertPtrReturn(pDevIns, VERR_INVALID_POINTER);
4063	AssertReturn(RT_ALIGN_Z(cbBitmap, sizeof(uint64_t)) == cbBitmap, VERR_INVALID_PARAMETER);
4064	int rc = PGM_LOCK(pVM);
4065	if (RT_SUCCESS(rc))
4066	{
4067	STAM_PROFILE_START(&pVM->pgm.s.StatMmio2QueryAndResetDirtyBitmap, a);
4068	rc = pgmR3PhysMmio2QueryAndResetDirtyBitmapLocked(pVM, pDevIns, hMmio2, pvBitmap, cbBitmap);
4069	STAM_PROFILE_STOP(&pVM->pgm.s.StatMmio2QueryAndResetDirtyBitmap, a);
4070	PGM_UNLOCK(pVM);
4071	}
4072	return rc;
4073	}
4074
4075	/**
4076	* Worker for PGMR3PhysMmio2ControlDirtyPageTracking
4077	*
4078	* Called owning the PGM lock.
4079	*/
4080	static int pgmR3PhysMmio2ControlDirtyPageTrackingLocked(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, bool fEnabled)
4081	{
4082	/*
4083	* Continue validation.
4084	*/
4085	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
4086	AssertReturn(pFirstRegMmio, VERR_INVALID_HANDLE);
4087	AssertReturn( (pFirstRegMmio->fFlags & (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK))
4088	== (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK)
4089	, VERR_INVALID_FUNCTION);
4090	AssertReturn(pDevIns == pFirstRegMmio->pDevInsR3, VERR_NOT_OWNER);
4091
4092	#ifdef VBOX_WITH_PGM_NEM_MODE
4093	/*
4094	* This is a nop if NEM is responsible for doing the tracking, we simply
4095	* leave the tracking on all the time there.
4096	*/
4097	if (pFirstRegMmio->pPhysHandlerR3 == NULL)
4098	{
4099	AssertReturn(VM_IS_NEM_ENABLED(pVM), VERR_INTERNAL_ERROR_4);
4100	return VINF_SUCCESS;
4101	}
4102	#endif
4103
4104	/*
4105	* Anyting needing doing?
4106	*/
4107	if (fEnabled != RT_BOOL(pFirstRegMmio->fFlags & PGMREGMMIO2RANGE_F_TRACKING_ENABLED))
4108	{
4109	LogFlowFunc(("fEnabled=%RTbool %s\n", fEnabled, pFirstRegMmio->RamRange.pszDesc));
4110
4111	/*
4112	* Update the PGMREGMMIO2RANGE_F_TRACKING_ENABLED flag.
4113	*/
4114	for (PPGMREGMMIO2RANGE pCur = pFirstRegMmio;;)
4115	{
4116	if (fEnabled)
4117	pCur->fFlags \|= PGMREGMMIO2RANGE_F_TRACKING_ENABLED;
4118	else
4119	pCur->fFlags &= ~PGMREGMMIO2RANGE_F_TRACKING_ENABLED;
4120	if (pCur->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
4121	break;
4122	pCur = pCur->pNextR3;
4123	AssertPtrReturn(pCur, VERR_INTERNAL_ERROR_5);
4124	AssertReturn( (pCur->fFlags & (PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES \| PGMREGMMIO2RANGE_F_FIRST_CHUNK))
4125	== PGMREGMMIO2RANGE_F_TRACK_DIRTY_PAGES
4126	, VERR_INTERNAL_ERROR_4);
4127	}
4128
4129	/*
4130	* Enable/disable handlers if currently mapped.
4131	*
4132	* We ignore status codes here as we've already changed the flags and
4133	* returning a failure status now would be confusing. Besides, the two
4134	* functions will continue past failures. As argued in the mapping code,
4135	* it's in the release log.
4136	*/
4137	if (pFirstRegMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED)
4138	{
4139	if (fEnabled)
4140	pgmR3PhysMmio2EnableDirtyPageTracing(pVM, pFirstRegMmio);
4141	else
4142	pgmR3PhysMmio2DisableDirtyPageTracing(pVM, pFirstRegMmio);
4143	}
4144	}
4145	else
4146	LogFlowFunc(("fEnabled=%RTbool %s - no change\n", fEnabled, pFirstRegMmio->RamRange.pszDesc));
4147
4148	return VINF_SUCCESS;
4149	}
4150
4151
4152	/**
4153	* Controls the dirty page tracking for an MMIO2 range.
4154	*
4155	* @returns VBox status code.
4156	* @param pVM The cross context VM structure.
4157	* @param pDevIns The device owning the MMIO2 memory.
4158	* @param hMmio2 The handle of the region.
4159	* @param fEnabled The new tracking state.
4160	*/
4161	VMMR3_INT_DECL(int) PGMR3PhysMmio2ControlDirtyPageTracking(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, bool fEnabled)
4162	{
4163	/*
4164	* Do some basic validation before grapping the PGM lock and continuing.
4165	*/
4166	AssertPtrReturn(pDevIns, VERR_INVALID_POINTER);
4167	int rc = PGM_LOCK(pVM);
4168	if (RT_SUCCESS(rc))
4169	{
4170	rc = pgmR3PhysMmio2ControlDirtyPageTrackingLocked(pVM, pDevIns, hMmio2, fEnabled);
4171	PGM_UNLOCK(pVM);
4172	}
4173	return rc;
4174	}
4175
4176
4177	/**
4178	* Changes the region number of an MMIO2 region.
4179	*
4180	* This is only for dealing with save state issues, nothing else.
4181	*
4182	* @return VBox status code.
4183	*
4184	* @param pVM The cross context VM structure.
4185	* @param pDevIns The device owning the MMIO2 memory.
4186	* @param hMmio2 The handle of the region.
4187	* @param iNewRegion The new region index.
4188	*
4189	* @thread EMT(0)
4190	* @sa @bugref{9359}
4191	*/
4192	VMMR3_INT_DECL(int) PGMR3PhysMmio2ChangeRegionNo(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, uint32_t iNewRegion)
4193	{
4194	/*
4195	* Validate input.
4196	*/
4197	VM_ASSERT_EMT0_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
4198	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_LOADING, VERR_VM_INVALID_VM_STATE);
4199	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
4200	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
4201	AssertReturn(iNewRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
4202
4203	AssertReturn(pVM->enmVMState == VMSTATE_LOADING, VERR_INVALID_STATE);
4204
4205	int rc = PGM_LOCK(pVM);
4206	AssertRCReturn(rc, rc);
4207
4208	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
4209	AssertReturnStmt(pFirstRegMmio, PGM_UNLOCK(pVM), VERR_NOT_FOUND);
4210	AssertReturnStmt(pgmR3PhysMmio2Find(pVM, pDevIns, pFirstRegMmio->iSubDev, iNewRegion, NIL_PGMMMIO2HANDLE) == NULL,
4211	PGM_UNLOCK(pVM), VERR_RESOURCE_IN_USE);
4212
4213	/*
4214	* Make the change.
4215	*/
4216	pFirstRegMmio->iRegion = (uint8_t)iNewRegion;
4217
4218	PGM_UNLOCK(pVM);
4219	return VINF_SUCCESS;
4220	}
4221
4222
4223
4224	/*********************************************************************************************************************************
4225	* ROM *
4226	*********************************************************************************************************************************/
4227
4228	/**
4229	* Worker for PGMR3PhysRomRegister.
4230	*
4231	* This is here to simplify lock management, i.e. the caller does all the
4232	* locking and we can simply return without needing to remember to unlock
4233	* anything first.
4234	*
4235	* @returns VBox status code.
4236	* @param pVM The cross context VM structure.
4237	* @param pDevIns The device instance owning the ROM.
4238	* @param GCPhys First physical address in the range.
4239	* Must be page aligned!
4240	* @param cb The size of the range (in bytes).
4241	* Must be page aligned!
4242	* @param pvBinary Pointer to the binary data backing the ROM image.
4243	* @param cbBinary The size of the binary data pvBinary points to.
4244	* This must be less or equal to @a cb.
4245	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
4246	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
4247	* @param pszDesc Pointer to description string. This must not be freed.
4248	*/
4249	static int pgmR3PhysRomRegisterLocked(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
4250	const void pvBinary, uint32_t cbBinary, uint8_t fFlags, const char pszDesc)
4251	{
4252	/*
4253	* Validate input.
4254	*/
4255	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
4256	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
4257	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
4258	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
4259	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
4260	AssertPtrReturn(pvBinary, VERR_INVALID_PARAMETER);
4261	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
4262	AssertReturn(!(fFlags & ~PGMPHYS_ROM_FLAGS_VALID_MASK), VERR_INVALID_PARAMETER);
4263	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_CREATING, VERR_VM_INVALID_VM_STATE);
4264
4265	const uint32_t cPages = cb >> PAGE_SHIFT;
4266
4267	/*
4268	* Find the ROM location in the ROM list first.
4269	*/
4270	PPGMROMRANGE pRomPrev = NULL;
4271	PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3;
4272	while (pRom && GCPhysLast >= pRom->GCPhys)
4273	{
4274	if ( GCPhys <= pRom->GCPhysLast
4275	&& GCPhysLast >= pRom->GCPhys)
4276	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
4277	GCPhys, GCPhysLast, pszDesc,
4278	pRom->GCPhys, pRom->GCPhysLast, pRom->pszDesc),
4279	VERR_PGM_RAM_CONFLICT);
4280	/* next */
4281	pRomPrev = pRom;
4282	pRom = pRom->pNextR3;
4283	}
4284
4285	/*
4286	* Find the RAM location and check for conflicts.
4287	*
4288	* Conflict detection is a bit different than for RAM registration since a
4289	* ROM can be located within a RAM range. So, what we have to check for is
4290	* other memory types (other than RAM that is) and that we don't span more
4291	* than one RAM range (lazy).
4292	*/
4293	bool fRamExists = false;
4294	PPGMRAMRANGE pRamPrev = NULL;
4295	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
4296	while (pRam && GCPhysLast >= pRam->GCPhys)
4297	{
4298	if ( GCPhys <= pRam->GCPhysLast
4299	&& GCPhysLast >= pRam->GCPhys)
4300	{
4301	/* completely within? */
4302	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
4303	&& GCPhysLast <= pRam->GCPhysLast,
4304	("%RGp-%RGp (%s) falls partly outside %RGp-%RGp (%s)\n",
4305	GCPhys, GCPhysLast, pszDesc,
4306	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
4307	VERR_PGM_RAM_CONFLICT);
4308	fRamExists = true;
4309	break;
4310	}
4311
4312	/* next */
4313	pRamPrev = pRam;
4314	pRam = pRam->pNextR3;
4315	}
4316	if (fRamExists)
4317	{
4318	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
4319	uint32_t cPagesLeft = cPages;
4320	while (cPagesLeft-- > 0)
4321	{
4322	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
4323	("%RGp (%R[pgmpage]) isn't a RAM page - registering %RGp-%RGp (%s).\n",
4324	pRam->GCPhys + ((RTGCPHYS)(uintptr_t)(pPage - &pRam->aPages[0]) << PAGE_SHIFT),
4325	pPage, GCPhys, GCPhysLast, pszDesc), VERR_PGM_RAM_CONFLICT);
4326	Assert(PGM_PAGE_IS_ZERO(pPage) \|\| PGM_IS_IN_NEM_MODE(pVM));
4327	pPage++;
4328	}
4329	}
4330
4331	/*
4332	* Update the base memory reservation if necessary.
4333	*/
4334	uint32_t cExtraBaseCost = fRamExists ? 0 : cPages;
4335	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4336	cExtraBaseCost += cPages;
4337	if (cExtraBaseCost)
4338	{
4339	int rc = MMR3IncreaseBaseReservation(pVM, cExtraBaseCost);
4340	if (RT_FAILURE(rc))
4341	return rc;
4342	}
4343
4344	#ifdef VBOX_WITH_NATIVE_NEM
4345	/*
4346	* Early NEM notification before we've made any changes or anything.
4347	*/
4348	uint32_t const fNemNotify = (fRamExists ? NEM_NOTIFY_PHYS_ROM_F_REPLACE : 0)
4349	\| (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED ? NEM_NOTIFY_PHYS_ROM_F_SHADOW : 0);
4350	uint8_t u2NemState = UINT8_MAX;
4351	if (VM_IS_NEM_ENABLED(pVM))
4352	{
4353	int rc = NEMR3NotifyPhysRomRegisterEarly(pVM, GCPhys, cPages << PAGE_SHIFT,
4354	fRamExists ? PGM_RAMRANGE_CALC_PAGE_R3PTR(pRam, GCPhys) : NULL,
4355	fNemNotify, &u2NemState);
4356	AssertLogRelRCReturn(rc, rc);
4357	}
4358	#endif
4359
4360	/*
4361	* Allocate memory for the virgin copy of the RAM. In simplified memory mode,
4362	* we allocate memory for any ad-hoc RAM range and for shadow pages.
4363	*/
4364	PGMMALLOCATEPAGESREQ pReq = NULL;
4365	#ifdef VBOX_WITH_PGM_NEM_MODE
4366	void *pvRam = NULL;
4367	void *pvAlt = NULL;
4368	if (pVM->pgm.s.fNemMode)
4369	{
4370	if (!fRamExists)
4371	{
4372	int rc = SUPR3PageAlloc(cPages, &pvRam);
4373	if (RT_FAILURE(rc))
4374	return rc;
4375	}
4376	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4377	{
4378	int rc = SUPR3PageAlloc(cPages, &pvAlt);
4379	if (RT_FAILURE(rc))
4380	{
4381	if (pvRam)
4382	SUPR3PageFree(pvRam, cPages);
4383	return rc;
4384	}
4385	}
4386	}
4387	else
4388	#endif
4389	{
4390	int rc = GMMR3AllocatePagesPrepare(pVM, &pReq, cPages, GMMACCOUNT_BASE);
4391	AssertRCReturn(rc, rc);
4392
4393	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4394	{
4395	pReq->aPages[iPage].HCPhysGCPhys = GCPhys + (iPage << PAGE_SHIFT);
4396	pReq->aPages[iPage].fZeroed = false;
4397	pReq->aPages[iPage].idPage = NIL_GMM_PAGEID;
4398	pReq->aPages[iPage].idSharedPage = NIL_GMM_PAGEID;
4399	}
4400
4401	rc = GMMR3AllocatePagesPerform(pVM, pReq);
4402	if (RT_FAILURE(rc))
4403	{
4404	GMMR3AllocatePagesCleanup(pReq);
4405	return rc;
4406	}
4407	}
4408
4409	/*
4410	* Allocate the new ROM range and RAM range (if necessary).
4411	*/
4412	PPGMROMRANGE pRomNew;
4413	int rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMROMRANGE, aPages[cPages]), 0, MM_TAG_PGM_PHYS, (void **)&pRomNew);
4414	if (RT_SUCCESS(rc))
4415	{
4416	PPGMRAMRANGE pRamNew = NULL;
4417	if (!fRamExists)
4418	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), sizeof(PGMPAGE), MM_TAG_PGM_PHYS, (void **)&pRamNew);
4419	if (RT_SUCCESS(rc))
4420	{
4421	/*
4422	* Initialize and insert the RAM range (if required).
4423	*/
4424	uint32_t const idxFirstRamPage = fRamExists ? (GCPhys - pRam->GCPhys) >> PAGE_SHIFT : 0;
4425	PPGMROMPAGE pRomPage = &pRomNew->aPages[0];
4426	if (!fRamExists)
4427	{
4428	/* New RAM range. */
4429	pRamNew->pSelfR0 = MMHyperCCToR0(pVM, pRamNew);
4430	pRamNew->GCPhys = GCPhys;
4431	pRamNew->GCPhysLast = GCPhysLast;
4432	pRamNew->cb = cb;
4433	pRamNew->pszDesc = pszDesc;
4434	pRamNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_ROM;
4435	pRamNew->pvR3 = NULL;
4436	pRamNew->paLSPages = NULL;
4437
4438	PPGMPAGE pRamPage = &pRamNew->aPages[idxFirstRamPage];
4439	#ifdef VBOX_WITH_PGM_NEM_MODE
4440	if (pVM->pgm.s.fNemMode)
4441	{
4442	AssertPtr(pvRam); Assert(pReq == NULL);
4443	pRamNew->pvR3 = pvRam;
4444	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4445	{
4446	PGM_PAGE_INIT(pRamPage, UINT64_C(0x0000fffffffff000), NIL_GMM_PAGEID,
4447	PGMPAGETYPE_ROM, PGM_PAGE_STATE_ALLOCATED);
4448	pRomPage->Virgin = *pRamPage;
4449	}
4450	}
4451	else
4452	#endif
4453	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4454	{
4455	PGM_PAGE_INIT(pRamPage,
4456	pReq->aPages[iPage].HCPhysGCPhys,
4457	pReq->aPages[iPage].idPage,
4458	PGMPAGETYPE_ROM,
4459	PGM_PAGE_STATE_ALLOCATED);
4460
4461	pRomPage->Virgin = *pRamPage;
4462	}
4463
4464	pVM->pgm.s.cAllPages += cPages;
4465	pVM->pgm.s.cPrivatePages += cPages;
4466	pgmR3PhysLinkRamRange(pVM, pRamNew, pRamPrev);
4467	}
4468	else
4469	{
4470	/* Existing RAM range. */
4471	PPGMPAGE pRamPage = &pRam->aPages[idxFirstRamPage];
4472	#ifdef VBOX_WITH_PGM_NEM_MODE
4473	if (pVM->pgm.s.fNemMode)
4474	{
4475	Assert(pvRam == NULL); Assert(pReq == NULL);
4476	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4477	{
4478	Assert(PGM_PAGE_GET_HCPHYS(pRamPage) == UINT64_C(0x0000fffffffff000));
4479	Assert(PGM_PAGE_GET_PAGEID(pRamPage) == NIL_GMM_PAGEID);
4480	Assert(PGM_PAGE_GET_STATE(pRamPage) == PGM_PAGE_STATE_ALLOCATED);
4481	PGM_PAGE_SET_TYPE(pVM, pRamPage, PGMPAGETYPE_ROM);
4482	PGM_PAGE_SET_STATE(pVM, pRamPage, PGM_PAGE_STATE_ALLOCATED);
4483	PGM_PAGE_SET_PDE_TYPE(pVM, pRamPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4484	PGM_PAGE_SET_PTE_INDEX(pVM, pRamPage, 0);
4485	PGM_PAGE_SET_TRACKING(pVM, pRamPage, 0);
4486
4487	pRomPage->Virgin = *pRamPage;
4488	}
4489	}
4490	else
4491	#endif
4492	{
4493	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4494	{
4495	PGM_PAGE_SET_TYPE(pVM, pRamPage, PGMPAGETYPE_ROM);
4496	PGM_PAGE_SET_HCPHYS(pVM, pRamPage, pReq->aPages[iPage].HCPhysGCPhys);
4497	PGM_PAGE_SET_STATE(pVM, pRamPage, PGM_PAGE_STATE_ALLOCATED);
4498	PGM_PAGE_SET_PAGEID(pVM, pRamPage, pReq->aPages[iPage].idPage);
4499	PGM_PAGE_SET_PDE_TYPE(pVM, pRamPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4500	PGM_PAGE_SET_PTE_INDEX(pVM, pRamPage, 0);
4501	PGM_PAGE_SET_TRACKING(pVM, pRamPage, 0);
4502
4503	pRomPage->Virgin = *pRamPage;
4504	}
4505	pVM->pgm.s.cZeroPages -= cPages;
4506	pVM->pgm.s.cPrivatePages += cPages;
4507	}
4508	pRamNew = pRam;
4509	}
4510
4511	#ifdef VBOX_WITH_NATIVE_NEM
4512	/* Set the NEM state of the pages if needed. */
4513	if (u2NemState != UINT8_MAX)
4514	pgmPhysSetNemStateForPages(&pRamNew->aPages[idxFirstRamPage], cPages, u2NemState);
4515	#endif
4516
4517	/* Flush physical page map TLB. */
4518	pgmPhysInvalidatePageMapTLB(pVM);
4519
4520	/*
4521	* Register the ROM access handler.
4522	*/
4523	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, pVM->pgm.s.hRomPhysHandlerType,
4524	pRomNew, MMHyperCCToR0(pVM, pRomNew), NIL_RTRCPTR, pszDesc);
4525	if (RT_SUCCESS(rc))
4526	{
4527	/*
4528	* Copy the image over to the virgin pages.
4529	* This must be done after linking in the RAM range.
4530	*/
4531	size_t cbBinaryLeft = cbBinary;
4532	PPGMPAGE pRamPage = &pRamNew->aPages[idxFirstRamPage];
4533	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
4534	{
4535	void *pvDstPage;
4536	rc = pgmPhysPageMap(pVM, pRamPage, GCPhys + (iPage << PAGE_SHIFT), &pvDstPage);
4537	if (RT_FAILURE(rc))
4538	{
4539	VMSetError(pVM, rc, RT_SRC_POS, "Failed to map virgin ROM page at %RGp", GCPhys);
4540	break;
4541	}
4542	if (cbBinaryLeft >= PAGE_SIZE)
4543	{
4544	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), PAGE_SIZE);
4545	cbBinaryLeft -= PAGE_SIZE;
4546	}
4547	else
4548	{
4549	ASMMemZeroPage(pvDstPage); /* (shouldn't be necessary, but can't hurt either) */
4550	if (cbBinaryLeft > 0)
4551	{
4552	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), cbBinaryLeft);
4553	cbBinaryLeft = 0;
4554	}
4555	}
4556	}
4557	if (RT_SUCCESS(rc))
4558	{
4559	/*
4560	* Initialize the ROM range.
4561	* Note that the Virgin member of the pages has already been initialized above.
4562	*/
4563	pRomNew->GCPhys = GCPhys;
4564	pRomNew->GCPhysLast = GCPhysLast;
4565	pRomNew->cb = cb;
4566	pRomNew->fFlags = fFlags;
4567	pRomNew->idSavedState = UINT8_MAX;
4568	pRomNew->cbOriginal = cbBinary;
4569	pRomNew->pszDesc = pszDesc;
4570	#ifdef VBOX_WITH_PGM_NEM_MODE
4571	pRomNew->pbR3Alternate = (uint8_t *)pvAlt;
4572	#endif
4573	pRomNew->pvOriginal = fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY
4574	? pvBinary : RTMemDup(pvBinary, cbBinary);
4575	if (pRomNew->pvOriginal)
4576	{
4577	for (unsigned iPage = 0; iPage < cPages; iPage++)
4578	{
4579	PPGMROMPAGE pPage = &pRomNew->aPages[iPage];
4580	pPage->enmProt = PGMROMPROT_READ_ROM_WRITE_IGNORE;
4581	#ifdef VBOX_WITH_PGM_NEM_MODE
4582	if (pVM->pgm.s.fNemMode)
4583	PGM_PAGE_INIT(&pPage->Shadow, UINT64_C(0x0000fffffffff000), NIL_GMM_PAGEID,
4584	PGMPAGETYPE_ROM_SHADOW, PGM_PAGE_STATE_ALLOCATED);
4585	else
4586	#endif
4587	PGM_PAGE_INIT_ZERO(&pPage->Shadow, pVM, PGMPAGETYPE_ROM_SHADOW);
4588	}
4589
4590	/* update the page count stats for the shadow pages. */
4591	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4592	{
4593	#ifdef VBOX_WITH_PGM_NEM_MODE
4594	if (pVM->pgm.s.fNemMode)
4595	pVM->pgm.s.cPrivatePages += cPages;
4596	else
4597	#endif
4598	pVM->pgm.s.cZeroPages += cPages;
4599	pVM->pgm.s.cAllPages += cPages;
4600	}
4601
4602	/*
4603	* Insert the ROM range, tell REM and return successfully.
4604	*/
4605	pRomNew->pNextR3 = pRom;
4606	pRomNew->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4607
4608	if (pRomPrev)
4609	{
4610	pRomPrev->pNextR3 = pRomNew;
4611	pRomPrev->pNextR0 = MMHyperCCToR0(pVM, pRomNew);
4612	}
4613	else
4614	{
4615	pVM->pgm.s.pRomRangesR3 = pRomNew;
4616	pVM->pgm.s.pRomRangesR0 = MMHyperCCToR0(pVM, pRomNew);
4617	}
4618
4619	pgmPhysInvalidatePageMapTLB(pVM);
4620	#ifdef VBOX_WITH_PGM_NEM_MODE
4621	if (!pVM->pgm.s.fNemMode)
4622	#endif
4623	GMMR3AllocatePagesCleanup(pReq);
4624
4625	#ifdef VBOX_WITH_NATIVE_NEM
4626	/*
4627	* Notify NEM again.
4628	*/
4629	if (VM_IS_NEM_ENABLED(pVM))
4630	{
4631	u2NemState = UINT8_MAX;
4632	rc = NEMR3NotifyPhysRomRegisterLate(pVM, GCPhys, cb, PGM_RAMRANGE_CALC_PAGE_R3PTR(pRamNew, GCPhys),
4633	fNemNotify, &u2NemState);
4634	if (u2NemState != UINT8_MAX)
4635	pgmPhysSetNemStateForPages(&pRamNew->aPages[idxFirstRamPage], cPages, u2NemState);
4636	if (RT_SUCCESS(rc))
4637	return rc;
4638	}
4639	else
4640	#endif
4641	return rc;
4642
4643	/*
4644	* bail out
4645	*/
4646	#ifdef VBOX_WITH_NATIVE_NEM
4647	/* unlink */
4648	if (pRomPrev)
4649	{
4650	pRomPrev->pNextR3 = pRom;
4651	pRomPrev->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4652	}
4653	else
4654	{
4655	pVM->pgm.s.pRomRangesR3 = pRom;
4656	pVM->pgm.s.pRomRangesR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4657	}
4658
4659	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4660	{
4661	# ifdef VBOX_WITH_PGM_NEM_MODE
4662	if (pVM->pgm.s.fNemMode)
4663	pVM->pgm.s.cPrivatePages -= cPages;
4664	else
4665	# endif
4666	pVM->pgm.s.cZeroPages -= cPages;
4667	pVM->pgm.s.cAllPages -= cPages;
4668	}
4669	#endif
4670	}
4671	else
4672	rc = VERR_NO_MEMORY;
4673	}
4674
4675	int rc2 = PGMHandlerPhysicalDeregister(pVM, GCPhys);
4676	AssertRC(rc2);
4677	}
4678
4679	if (!fRamExists)
4680	{
4681	pgmR3PhysUnlinkRamRange2(pVM, pRamNew, pRamPrev);
4682	MMHyperFree(pVM, pRamNew);
4683	}
4684	else
4685	{
4686	PPGMPAGE pRamPage = &pRam->aPages[idxFirstRamPage];
4687	#ifdef VBOX_WITH_PGM_NEM_MODE
4688	if (pVM->pgm.s.fNemMode)
4689	{
4690	Assert(pvRam == NULL); Assert(pReq == NULL);
4691	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++, pRomPage++)
4692	{
4693	Assert(PGM_PAGE_GET_HCPHYS(pRamPage) == UINT64_C(0x0000fffffffff000));
4694	Assert(PGM_PAGE_GET_PAGEID(pRamPage) == NIL_GMM_PAGEID);
4695	Assert(PGM_PAGE_GET_STATE(pRamPage) == PGM_PAGE_STATE_ALLOCATED);
4696	PGM_PAGE_SET_TYPE(pVM, pRamPage, PGMPAGETYPE_RAM);
4697	PGM_PAGE_SET_STATE(pVM, pRamPage, PGM_PAGE_STATE_ALLOCATED);
4698	}
4699	}
4700	else
4701	#endif
4702	{
4703	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
4704	PGM_PAGE_INIT_ZERO(pRamPage, pVM, PGMPAGETYPE_RAM);
4705	pVM->pgm.s.cZeroPages += cPages;
4706	pVM->pgm.s.cPrivatePages -= cPages;
4707	}
4708	}
4709	}
4710	MMHyperFree(pVM, pRomNew);
4711	}
4712
4713	/** @todo Purge the mapping cache or something... */
4714	#ifdef VBOX_WITH_PGM_NEM_MODE
4715	if (pVM->pgm.s.fNemMode)
4716	{
4717	Assert(!pReq);
4718	if (pvRam)
4719	SUPR3PageFree(pvRam, cPages);
4720	if (pvAlt)
4721	SUPR3PageFree(pvAlt, cPages);
4722	}
4723	else
4724	#endif
4725	{
4726	GMMR3FreeAllocatedPages(pVM, pReq);
4727	GMMR3AllocatePagesCleanup(pReq);
4728	}
4729	return rc;
4730	}
4731
4732
4733	/**
4734	* Registers a ROM image.
4735	*
4736	* Shadowed ROM images requires double the amount of backing memory, so,
4737	* don't use that unless you have to. Shadowing of ROM images is process
4738	* where we can select where the reads go and where the writes go. On real
4739	* hardware the chipset provides means to configure this. We provide
4740	* PGMR3PhysProtectROM() for this purpose.
4741	*
4742	* A read-only copy of the ROM image will always be kept around while we
4743	* will allocate RAM pages for the changes on demand (unless all memory
4744	* is configured to be preallocated).
4745	*
4746	* @returns VBox status code.
4747	* @param pVM The cross context VM structure.
4748	* @param pDevIns The device instance owning the ROM.
4749	* @param GCPhys First physical address in the range.
4750	* Must be page aligned!
4751	* @param cb The size of the range (in bytes).
4752	* Must be page aligned!
4753	* @param pvBinary Pointer to the binary data backing the ROM image.
4754	* @param cbBinary The size of the binary data pvBinary points to.
4755	* This must be less or equal to @a cb.
4756	* @param fFlags Mask of flags, PGMPHYS_ROM_FLAGS_XXX.
4757	* @param pszDesc Pointer to description string. This must not be freed.
4758	*
4759	* @remark There is no way to remove the rom, automatically on device cleanup or
4760	* manually from the device yet. This isn't difficult in any way, it's
4761	* just not something we expect to be necessary for a while.
4762	*/
4763	VMMR3DECL(int) PGMR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
4764	const void pvBinary, uint32_t cbBinary, uint8_t fFlags, const char pszDesc)
4765	{
4766	Log(("PGMR3PhysRomRegister: pDevIns=%p GCPhys=%RGp(-%RGp) cb=%RGp pvBinary=%p cbBinary=%#x fFlags=%#x pszDesc=%s\n",
4767	pDevIns, GCPhys, GCPhys + cb, cb, pvBinary, cbBinary, fFlags, pszDesc));
4768	PGM_LOCK_VOID(pVM);
4769	int rc = pgmR3PhysRomRegisterLocked(pVM, pDevIns, GCPhys, cb, pvBinary, cbBinary, fFlags, pszDesc);
4770	PGM_UNLOCK(pVM);
4771	return rc;
4772	}
4773
4774
4775	/**
4776	* Called by PGMR3MemSetup to reset the shadow, switch to the virgin, and verify
4777	* that the virgin part is untouched.
4778	*
4779	* This is done after the normal memory has been cleared.
4780	*
4781	* ASSUMES that the caller owns the PGM lock.
4782	*
4783	* @param pVM The cross context VM structure.
4784	*/
4785	int pgmR3PhysRomReset(PVM pVM)
4786	{
4787	PGM_LOCK_ASSERT_OWNER(pVM);
4788	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4789	{
4790	const uint32_t cPages = pRom->cb >> PAGE_SHIFT;
4791
4792	if (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4793	{
4794	/*
4795	* Reset the physical handler.
4796	*/
4797	int rc = PGMR3PhysRomProtect(pVM, pRom->GCPhys, pRom->cb, PGMROMPROT_READ_ROM_WRITE_IGNORE);
4798	AssertRCReturn(rc, rc);
4799
4800	/*
4801	* What we do with the shadow pages depends on the memory
4802	* preallocation option. If not enabled, we'll just throw
4803	* out all the dirty pages and replace them by the zero page.
4804	*/
4805	#ifdef VBOX_WITH_PGM_NEM_MODE
4806	if (pVM->pgm.s.fNemMode)
4807	{
4808	/* Clear all the shadow pages (currently using alternate backing). */
4809	RT_BZERO(pRom->pbR3Alternate, pRom->cb);
4810	}
4811	else
4812	#endif
4813	if (!pVM->pgm.s.fRamPreAlloc)
4814	{
4815	/* Free the dirty pages. */
4816	uint32_t cPendingPages = 0;
4817	PGMMFREEPAGESREQ pReq;
4818	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
4819	AssertRCReturn(rc, rc);
4820
4821	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4822	if ( !PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow)
4823	&& !PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow))
4824	{
4825	Assert(PGM_PAGE_GET_STATE(&pRom->aPages[iPage].Shadow) == PGM_PAGE_STATE_ALLOCATED);
4826	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, &pRom->aPages[iPage].Shadow,
4827	pRom->GCPhys + (iPage << PAGE_SHIFT),
4828	(PGMPAGETYPE)PGM_PAGE_GET_TYPE(&pRom->aPages[iPage].Shadow));
4829	AssertLogRelRCReturn(rc, rc);
4830	}
4831
4832	if (cPendingPages)
4833	{
4834	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
4835	AssertLogRelRCReturn(rc, rc);
4836	}
4837	GMMR3FreePagesCleanup(pReq);
4838	}
4839	else
4840	{
4841	/* clear all the shadow pages. */
4842	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4843	{
4844	if (PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow))
4845	continue;
4846	Assert(!PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow));
4847	void *pvDstPage;
4848	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4849	rc = pgmPhysPageMakeWritableAndMap(pVM, &pRom->aPages[iPage].Shadow, GCPhys, &pvDstPage);
4850	if (RT_FAILURE(rc))
4851	break;
4852	ASMMemZeroPage(pvDstPage);
4853	}
4854	AssertRCReturn(rc, rc);
4855	}
4856	}
4857
4858	/*
4859	* Restore the original ROM pages after a saved state load.
4860	* Also, in strict builds check that ROM pages remain unmodified.
4861	*/
4862	#ifndef VBOX_STRICT
4863	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4864	#endif
4865	{
4866	size_t cbSrcLeft = pRom->cbOriginal;
4867	uint8_t const pbSrcPage = (uint8_t const )pRom->pvOriginal;
4868	uint32_t cRestored = 0;
4869	for (uint32_t iPage = 0; iPage < cPages && cbSrcLeft > 0; iPage++, pbSrcPage += PAGE_SIZE)
4870	{
4871	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4872	PPGMPAGE const pPage = pgmPhysGetPage(pVM, GCPhys);
4873	void const *pvDstPage = NULL;
4874	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhys, &pvDstPage);
4875	if (RT_FAILURE(rc))
4876	break;
4877
4878	if (memcmp(pvDstPage, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE)))
4879	{
4880	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4881	{
4882	void *pvDstPageW = NULL;
4883	rc = pgmPhysPageMap(pVM, pPage, GCPhys, &pvDstPageW);
4884	AssertLogRelRCReturn(rc, rc);
4885	memcpy(pvDstPageW, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE));
4886	cRestored++;
4887	}
4888	else
4889	LogRel(("pgmR3PhysRomReset: %RGp: ROM page changed (%s)\n", GCPhys, pRom->pszDesc));
4890	}
4891	cbSrcLeft -= RT_MIN(cbSrcLeft, PAGE_SIZE);
4892	}
4893	if (cRestored > 0)
4894	LogRel(("PGM: ROM \"%s\": Reloaded %u of %u pages.\n", pRom->pszDesc, cRestored, cPages));
4895	}
4896	}
4897
4898	/* Clear the ROM restore flag now as we only need to do this once after
4899	loading saved state. */
4900	pVM->pgm.s.fRestoreRomPagesOnReset = false;
4901
4902	return VINF_SUCCESS;
4903	}
4904
4905
4906	/**
4907	* Called by PGMR3Term to free resources.
4908	*
4909	* ASSUMES that the caller owns the PGM lock.
4910	*
4911	* @param pVM The cross context VM structure.
4912	*/
4913	void pgmR3PhysRomTerm(PVM pVM)
4914	{
4915	/*
4916	* Free the heap copy of the original bits.
4917	*/
4918	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4919	{
4920	if ( pRom->pvOriginal
4921	&& !(pRom->fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY))
4922	{
4923	RTMemFree((void *)pRom->pvOriginal);
4924	pRom->pvOriginal = NULL;
4925	}
4926	}
4927	}
4928
4929
4930	/**
4931	* Change the shadowing of a range of ROM pages.
4932	*
4933	* This is intended for implementing chipset specific memory registers
4934	* and will not be very strict about the input. It will silently ignore
4935	* any pages that are not the part of a shadowed ROM.
4936	*
4937	* @returns VBox status code.
4938	* @retval VINF_PGM_SYNC_CR3
4939	*
4940	* @param pVM The cross context VM structure.
4941	* @param GCPhys Where to start. Page aligned.
4942	* @param cb How much to change. Page aligned.
4943	* @param enmProt The new ROM protection.
4944	*/
4945	VMMR3DECL(int) PGMR3PhysRomProtect(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMROMPROT enmProt)
4946	{
4947	/*
4948	* Check input
4949	*/
4950	if (!cb)
4951	return VINF_SUCCESS;
4952	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4953	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4954	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
4955	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
4956	AssertReturn(enmProt >= PGMROMPROT_INVALID && enmProt <= PGMROMPROT_END, VERR_INVALID_PARAMETER);
4957
4958	/*
4959	* Process the request.
4960	*/
4961	PGM_LOCK_VOID(pVM);
4962	int rc = VINF_SUCCESS;
4963	bool fFlushTLB = false;
4964	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4965	{
4966	if ( GCPhys <= pRom->GCPhysLast
4967	&& GCPhysLast >= pRom->GCPhys
4968	&& (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED))
4969	{
4970	/*
4971	* Iterate the relevant pages and make necessary the changes.
4972	*/
4973	#ifdef VBOX_WITH_NATIVE_NEM
4974	PPGMRAMRANGE const pRam = pgmPhysGetRange(pVM, GCPhys);
4975	AssertPtrReturn(pRam, VERR_INTERNAL_ERROR_3);
4976	#endif
4977	bool fChanges = false;
4978	uint32_t const cPages = pRom->GCPhysLast <= GCPhysLast
4979	? pRom->cb >> PAGE_SHIFT
4980	: (GCPhysLast - pRom->GCPhys + 1) >> PAGE_SHIFT;
4981	for (uint32_t iPage = (GCPhys - pRom->GCPhys) >> PAGE_SHIFT;
4982	iPage < cPages;
4983	iPage++)
4984	{
4985	PPGMROMPAGE pRomPage = &pRom->aPages[iPage];
4986	if (PGMROMPROT_IS_ROM(pRomPage->enmProt) != PGMROMPROT_IS_ROM(enmProt))
4987	{
4988	fChanges = true;
4989
4990	/* flush references to the page. */
4991	RTGCPHYS const GCPhysPage = pRom->GCPhys + (iPage << PAGE_SHIFT);
4992	PPGMPAGE pRamPage = pgmPhysGetPage(pVM, GCPhysPage);
4993	int rc2 = pgmPoolTrackUpdateGCPhys(pVM, GCPhysPage, pRamPage, true /fFlushPTEs/, &fFlushTLB);
4994	if (rc2 != VINF_SUCCESS && (rc == VINF_SUCCESS \|\| RT_FAILURE(rc2)))
4995	rc = rc2;
4996	#ifdef VBOX_WITH_NATIVE_NEM
4997	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pRamPage);
4998	#endif
4999
5000	PPGMPAGE pOld = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Virgin : &pRomPage->Shadow;
5001	PPGMPAGE pNew = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Shadow : &pRomPage->Virgin;
5002
5003	pOld = pRamPage;
5004	pRamPage = pNew;
5005	/** @todo preserve the volatile flags (handlers) when these have been moved out of HCPhys! */
5006
5007	#ifdef VBOX_WITH_NATIVE_NEM
5008	# ifdef VBOX_WITH_PGM_NEM_MODE
5009	/* In simplified mode we have to switch the page data around too. */
5010	if (pVM->pgm.s.fNemMode)
5011	{
5012	uint8_t abPage[PAGE_SIZE];
5013	uint8_t * const pbRamPage = PGM_RAMRANGE_CALC_PAGE_R3PTR(pRam, GCPhysPage);
5014	memcpy(abPage, &pRom->pbR3Alternate[(size_t)iPage << PAGE_SHIFT], sizeof(abPage));
5015	memcpy(&pRom->pbR3Alternate[(size_t)iPage << PAGE_SHIFT], pbRamPage, sizeof(abPage));
5016	memcpy(pbRamPage, abPage, sizeof(abPage));
5017	}
5018	# endif
5019	/* Tell NEM about the backing and protection change. */
5020	if (VM_IS_NEM_ENABLED(pVM))
5021	{
5022	PGMPAGETYPE enmType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pNew);
5023	NEMHCNotifyPhysPageChanged(pVM, GCPhys, PGM_PAGE_GET_HCPHYS(pOld), PGM_PAGE_GET_HCPHYS(pNew),
5024	PGM_RAMRANGE_CALC_PAGE_R3PTR(pRam, GCPhysPage),
5025	pgmPhysPageCalcNemProtection(pRamPage, enmType), enmType, &u2State);
5026	PGM_PAGE_SET_NEM_STATE(pRamPage, u2State);
5027	}
5028	#endif
5029	}
5030	pRomPage->enmProt = enmProt;
5031	}
5032
5033	/*
5034	* Reset the access handler if we made changes, no need
5035	* to optimize this.
5036	*/
5037	if (fChanges)
5038	{
5039	int rc2 = PGMHandlerPhysicalReset(pVM, pRom->GCPhys);
5040	if (RT_FAILURE(rc2))
5041	{
5042	PGM_UNLOCK(pVM);
5043	AssertRC(rc);
5044	return rc2;
5045	}
5046	}
5047
5048	/* Advance - cb isn't updated. */
5049	GCPhys = pRom->GCPhys + (cPages << PAGE_SHIFT);
5050	}
5051	}
5052	PGM_UNLOCK(pVM);
5053	if (fFlushTLB)
5054	PGM_INVL_ALL_VCPU_TLBS(pVM);
5055
5056	return rc;
5057	}
5058
5059
5060
5061	/*********************************************************************************************************************************
5062	* Ballooning *
5063	*********************************************************************************************************************************/
5064
5065	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
5066
5067	/**
5068	* Rendezvous callback used by PGMR3ChangeMemBalloon that changes the memory balloon size
5069	*
5070	* This is only called on one of the EMTs while the other ones are waiting for
5071	* it to complete this function.
5072	*
5073	* @returns VINF_SUCCESS (VBox strict status code).
5074	* @param pVM The cross context VM structure.
5075	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
5076	* @param pvUser User parameter
5077	*/
5078	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysChangeMemBalloonRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
5079	{
5080	uintptr_t paUser = (uintptr_t )pvUser;
5081	bool fInflate = !!paUser[0];
5082	unsigned cPages = paUser[1];
5083	RTGCPHYS paPhysPage = (RTGCPHYS )paUser[2];
5084	uint32_t cPendingPages = 0;
5085	PGMMFREEPAGESREQ pReq;
5086	int rc;
5087
5088	Log(("pgmR3PhysChangeMemBalloonRendezvous: %s %x pages\n", (fInflate) ? "inflate" : "deflate", cPages));
5089	PGM_LOCK_VOID(pVM);
5090
5091	if (fInflate)
5092	{
5093	/* Flush the PGM pool cache as we might have stale references to pages that we just freed. */
5094	pgmR3PoolClearAllRendezvous(pVM, pVCpu, NULL);
5095
5096	/* Replace pages with ZERO pages. */
5097	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
5098	if (RT_FAILURE(rc))
5099	{
5100	PGM_UNLOCK(pVM);
5101	AssertLogRelRC(rc);
5102	return rc;
5103	}
5104
5105	/* Iterate the pages. */
5106	for (unsigned i = 0; i < cPages; i++)
5107	{
5108	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
5109	if ( pPage == NULL
5110	\|\| PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM)
5111	{
5112	Log(("pgmR3PhysChangeMemBalloonRendezvous: invalid physical page %RGp pPage->u3Type=%d\n", paPhysPage[i], pPage ? PGM_PAGE_GET_TYPE(pPage) : 0));
5113	break;
5114	}
5115
5116	LogFlow(("balloon page: %RGp\n", paPhysPage[i]));
5117
5118	/* Flush the shadow PT if this page was previously used as a guest page table. */
5119	pgmPoolFlushPageByGCPhys(pVM, paPhysPage[i]);
5120
5121	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, paPhysPage[i], (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage));
5122	if (RT_FAILURE(rc))
5123	{
5124	PGM_UNLOCK(pVM);
5125	AssertLogRelRC(rc);
5126	return rc;
5127	}
5128	Assert(PGM_PAGE_IS_ZERO(pPage));
5129	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_BALLOONED);
5130	}
5131
5132	if (cPendingPages)
5133	{
5134	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
5135	if (RT_FAILURE(rc))
5136	{
5137	PGM_UNLOCK(pVM);
5138	AssertLogRelRC(rc);
5139	return rc;
5140	}
5141	}
5142	GMMR3FreePagesCleanup(pReq);
5143	}
5144	else
5145	{
5146	/* Iterate the pages. */
5147	for (unsigned i = 0; i < cPages; i++)
5148	{
5149	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
5150	AssertBreak(pPage && PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM);
5151
5152	LogFlow(("Free ballooned page: %RGp\n", paPhysPage[i]));
5153
5154	Assert(PGM_PAGE_IS_BALLOONED(pPage));
5155
5156	/* Change back to zero page. (NEM does not need to be informed.) */
5157	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
5158	}
5159
5160	/* Note that we currently do not map any ballooned pages in our shadow page tables, so no need to flush the pgm pool. */
5161	}
5162
5163	/* Notify GMM about the balloon change. */
5164	rc = GMMR3BalloonedPages(pVM, (fInflate) ? GMMBALLOONACTION_INFLATE : GMMBALLOONACTION_DEFLATE, cPages);
5165	if (RT_SUCCESS(rc))
5166	{
5167	if (!fInflate)
5168	{
5169	Assert(pVM->pgm.s.cBalloonedPages >= cPages);
5170	pVM->pgm.s.cBalloonedPages -= cPages;
5171	}
5172	else
5173	pVM->pgm.s.cBalloonedPages += cPages;
5174	}
5175
5176	PGM_UNLOCK(pVM);
5177
5178	/* Flush the recompiler's TLB as well. */
5179	for (VMCPUID i = 0; i < pVM->cCpus; i++)
5180	CPUMSetChangedFlags(pVM->apCpusR3[i], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
5181
5182	AssertLogRelRC(rc);
5183	return rc;
5184	}
5185
5186
5187	/**
5188	* Frees a range of ram pages, replacing them with ZERO pages; helper for PGMR3PhysFreeRamPages
5189	*
5190	* @returns VBox status code.
5191	* @param pVM The cross context VM structure.
5192	* @param fInflate Inflate or deflate memory balloon
5193	* @param cPages Number of pages to free
5194	* @param paPhysPage Array of guest physical addresses
5195	*/
5196	static DECLCALLBACK(void) pgmR3PhysChangeMemBalloonHelper(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
5197	{
5198	uintptr_t paUser[3];
5199
5200	paUser[0] = fInflate;
5201	paUser[1] = cPages;
5202	paUser[2] = (uintptr_t)paPhysPage;
5203	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
5204	AssertRC(rc);
5205
5206	/* Made a copy in PGMR3PhysFreeRamPages; free it here. */
5207	RTMemFree(paPhysPage);
5208	}
5209
5210	#endif /* 64-bit host && (Windows \|\| Solaris \|\| Linux \|\| FreeBSD) */
5211
5212	/**
5213	* Inflate or deflate a memory balloon
5214	*
5215	* @returns VBox status code.
5216	* @param pVM The cross context VM structure.
5217	* @param fInflate Inflate or deflate memory balloon
5218	* @param cPages Number of pages to free
5219	* @param paPhysPage Array of guest physical addresses
5220	*/
5221	VMMR3DECL(int) PGMR3PhysChangeMemBalloon(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
5222	{
5223	/* This must match GMMR0Init; currently we only support memory ballooning on all 64-bit hosts except Mac OS X */
5224	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
5225	int rc;
5226
5227	/* Older additions (ancient non-functioning balloon code) pass wrong physical addresses. */
5228	AssertReturn(!(paPhysPage[0] & 0xfff), VERR_INVALID_PARAMETER);
5229
5230	/* We own the IOM lock here and could cause a deadlock by waiting for another VCPU that is blocking on the IOM lock.
5231	* In the SMP case we post a request packet to postpone the job.
5232	*/
5233	if (pVM->cCpus > 1)
5234	{
5235	unsigned cbPhysPage = cPages * sizeof(paPhysPage[0]);
5236	RTGCPHYS paPhysPageCopy = (RTGCPHYS )RTMemAlloc(cbPhysPage);
5237	AssertReturn(paPhysPageCopy, VERR_NO_MEMORY);
5238
5239	memcpy(paPhysPageCopy, paPhysPage, cbPhysPage);
5240
5241	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysChangeMemBalloonHelper, 4, pVM, fInflate, cPages, paPhysPageCopy);
5242	AssertRC(rc);
5243	}
5244	else
5245	{
5246	uintptr_t paUser[3];
5247
5248	paUser[0] = fInflate;
5249	paUser[1] = cPages;
5250	paUser[2] = (uintptr_t)paPhysPage;
5251	rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
5252	AssertRC(rc);
5253	}
5254	return rc;
5255
5256	#else
5257	NOREF(pVM); NOREF(fInflate); NOREF(cPages); NOREF(paPhysPage);
5258	return VERR_NOT_IMPLEMENTED;
5259	#endif
5260	}
5261
5262
5263	/*********************************************************************************************************************************
5264	* Write Monitoring *
5265	*********************************************************************************************************************************/
5266
5267	/**
5268	* Rendezvous callback used by PGMR3WriteProtectRAM that write protects all
5269	* physical RAM.
5270	*
5271	* This is only called on one of the EMTs while the other ones are waiting for
5272	* it to complete this function.
5273	*
5274	* @returns VINF_SUCCESS (VBox strict status code).
5275	* @param pVM The cross context VM structure.
5276	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
5277	* @param pvUser User parameter, unused.
5278	*/
5279	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysWriteProtectRAMRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
5280	{
5281	int rc = VINF_SUCCESS;
5282	NOREF(pvUser); NOREF(pVCpu);
5283
5284	PGM_LOCK_VOID(pVM);
5285	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
5286	pgmPoolResetDirtyPages(pVM);
5287	#endif
5288
5289	/** @todo pointless to write protect the physical page pointed to by RSP. */
5290
5291	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
5292	pRam;
5293	pRam = pRam->CTX_SUFF(pNext))
5294	{
5295	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
5296	for (uint32_t iPage = 0; iPage < cPages; iPage++)
5297	{
5298	PPGMPAGE pPage = &pRam->aPages[iPage];
5299	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
5300
5301	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
5302	\|\| enmPageType == PGMPAGETYPE_MMIO2)
5303	{
5304	/*
5305	* A RAM page.
5306	*/
5307	switch (PGM_PAGE_GET_STATE(pPage))
5308	{
5309	case PGM_PAGE_STATE_ALLOCATED:
5310	/** @todo Optimize this: Don't always re-enable write
5311	* monitoring if the page is known to be very busy. */
5312	if (PGM_PAGE_IS_WRITTEN_TO(pPage))
5313	PGM_PAGE_CLEAR_WRITTEN_TO(pVM, pPage);
5314
5315	pgmPhysPageWriteMonitor(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
5316	break;
5317
5318	case PGM_PAGE_STATE_SHARED:
5319	AssertFailed();
5320	break;
5321
5322	case PGM_PAGE_STATE_WRITE_MONITORED: /* nothing to change. */
5323	default:
5324	break;
5325	}
5326	}
5327	}
5328	}
5329	pgmR3PoolWriteProtectPages(pVM);
5330	PGM_INVL_ALL_VCPU_TLBS(pVM);
5331	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
5332	CPUMSetChangedFlags(pVM->apCpusR3[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
5333
5334	PGM_UNLOCK(pVM);
5335	return rc;
5336	}
5337
5338	/**
5339	* Protect all physical RAM to monitor writes
5340	*
5341	* @returns VBox status code.
5342	* @param pVM The cross context VM structure.
5343	*/
5344	VMMR3DECL(int) PGMR3PhysWriteProtectRAM(PVM pVM)
5345	{
5346	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
5347
5348	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysWriteProtectRAMRendezvous, NULL);
5349	AssertRC(rc);
5350	return rc;
5351	}
5352
5353
5354	/*********************************************************************************************************************************
5355	* Stats. *
5356	*********************************************************************************************************************************/
5357
5358	/**
5359	* Query the amount of free memory inside VMMR0
5360	*
5361	* @returns VBox status code.
5362	* @param pUVM The user mode VM handle.
5363	* @param pcbAllocMem Where to return the amount of memory allocated
5364	* by VMs.
5365	* @param pcbFreeMem Where to return the amount of memory that is
5366	* allocated from the host but not currently used
5367	* by any VMs.
5368	* @param pcbBallonedMem Where to return the sum of memory that is
5369	* currently ballooned by the VMs.
5370	* @param pcbSharedMem Where to return the amount of memory that is
5371	* currently shared.
5372	*/
5373	VMMR3DECL(int) PGMR3QueryGlobalMemoryStats(PUVM pUVM, uint64_t pcbAllocMem, uint64_t pcbFreeMem,
5374	uint64_t pcbBallonedMem, uint64_t pcbSharedMem)
5375	{
5376	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
5377	VM_ASSERT_VALID_EXT_RETURN(pUVM->pVM, VERR_INVALID_VM_HANDLE);
5378
5379	uint64_t cAllocPages = 0;
5380	uint64_t cFreePages = 0;
5381	uint64_t cBalloonPages = 0;
5382	uint64_t cSharedPages = 0;
5383	int rc = GMMR3QueryHypervisorMemoryStats(pUVM->pVM, &cAllocPages, &cFreePages, &cBalloonPages, &cSharedPages);
5384	AssertRCReturn(rc, rc);
5385
5386	if (pcbAllocMem)
5387	pcbAllocMem = cAllocPages _4K;
5388
5389	if (pcbFreeMem)
5390	pcbFreeMem = cFreePages _4K;
5391
5392	if (pcbBallonedMem)
5393	pcbBallonedMem = cBalloonPages _4K;
5394
5395	if (pcbSharedMem)
5396	pcbSharedMem = cSharedPages _4K;
5397
5398	Log(("PGMR3QueryVMMMemoryStats: all=%llx free=%llx ballooned=%llx shared=%llx\n",
5399	cAllocPages, cFreePages, cBalloonPages, cSharedPages));
5400	return VINF_SUCCESS;
5401	}
5402
5403
5404	/**
5405	* Query memory stats for the VM.
5406	*
5407	* @returns VBox status code.
5408	* @param pUVM The user mode VM handle.
5409	* @param pcbTotalMem Where to return total amount memory the VM may
5410	* possibly use.
5411	* @param pcbPrivateMem Where to return the amount of private memory
5412	* currently allocated.
5413	* @param pcbSharedMem Where to return the amount of actually shared
5414	* memory currently used by the VM.
5415	* @param pcbZeroMem Where to return the amount of memory backed by
5416	* zero pages.
5417	*
5418	* @remarks The total mem is normally larger than the sum of the three
5419	* components. There are two reasons for this, first the amount of
5420	* shared memory is what we're sure is shared instead of what could
5421	* possibly be shared with someone. Secondly, because the total may
5422	* include some pure MMIO pages that doesn't go into any of the three
5423	* sub-counts.
5424	*
5425	* @todo Why do we return reused shared pages instead of anything that could
5426	* potentially be shared? Doesn't this mean the first VM gets a much
5427	* lower number of shared pages?
5428	*/
5429	VMMR3DECL(int) PGMR3QueryMemoryStats(PUVM pUVM, uint64_t pcbTotalMem, uint64_t pcbPrivateMem,
5430	uint64_t pcbSharedMem, uint64_t pcbZeroMem)
5431	{
5432	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
5433	PVM pVM = pUVM->pVM;
5434	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
5435
5436	if (pcbTotalMem)
5437	pcbTotalMem = (uint64_t)pVM->pgm.s.cAllPages PAGE_SIZE;
5438
5439	if (pcbPrivateMem)
5440	pcbPrivateMem = (uint64_t)pVM->pgm.s.cPrivatePages PAGE_SIZE;
5441
5442	if (pcbSharedMem)
5443	pcbSharedMem = (uint64_t)pVM->pgm.s.cReusedSharedPages PAGE_SIZE;
5444
5445	if (pcbZeroMem)
5446	pcbZeroMem = (uint64_t)pVM->pgm.s.cZeroPages PAGE_SIZE;
5447
5448	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));
5449	return VINF_SUCCESS;
5450	}
5451
5452
5453
5454	/*********************************************************************************************************************************
5455	* Chunk Mappings and Page Allocation *
5456	*********************************************************************************************************************************/
5457
5458	/**
5459	* Tree enumeration callback for dealing with age rollover.
5460	* It will perform a simple compression of the current age.
5461	*/
5462	static DECLCALLBACK(int) pgmR3PhysChunkAgeingRolloverCallback(PAVLU32NODECORE pNode, void *pvUser)
5463	{
5464	/* Age compression - ASSUMES iNow == 4. */
5465	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
5466	if (pChunk->iLastUsed >= UINT32_C(0xffffff00))
5467	pChunk->iLastUsed = 3;
5468	else if (pChunk->iLastUsed >= UINT32_C(0xfffff000))
5469	pChunk->iLastUsed = 2;
5470	else if (pChunk->iLastUsed)
5471	pChunk->iLastUsed = 1;
5472	else /* iLastUsed = 0 */
5473	pChunk->iLastUsed = 4;
5474
5475	NOREF(pvUser);
5476	return 0;
5477	}
5478
5479
5480	/**
5481	* The structure passed in the pvUser argument of pgmR3PhysChunkUnmapCandidateCallback().
5482	*/
5483	typedef struct PGMR3PHYSCHUNKUNMAPCB
5484	{
5485	PVM pVM; /*< Pointer to the VM. /
5486	PPGMCHUNKR3MAP pChunk; /*< The chunk to unmap. /
5487	} PGMR3PHYSCHUNKUNMAPCB, *PPGMR3PHYSCHUNKUNMAPCB;
5488
5489
5490	/**
5491	* Callback used to find the mapping that's been unused for
5492	* the longest time.
5493	*/
5494	static DECLCALLBACK(int) pgmR3PhysChunkUnmapCandidateCallback(PAVLU32NODECORE pNode, void *pvUser)
5495	{
5496	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
5497	PPGMR3PHYSCHUNKUNMAPCB pArg = (PPGMR3PHYSCHUNKUNMAPCB)pvUser;
5498
5499	/*
5500	* Check for locks and compare when last used.
5501	*/
5502	if (pChunk->cRefs)
5503	return 0;
5504	if (pChunk->cPermRefs)
5505	return 0;
5506	if ( pArg->pChunk
5507	&& pChunk->iLastUsed >= pArg->pChunk->iLastUsed)
5508	return 0;
5509
5510	/*
5511	* Check that it's not in any of the TLBs.
5512	*/
5513	PVM pVM = pArg->pVM;
5514	if ( pVM->pgm.s.ChunkR3Map.Tlb.aEntries[PGM_CHUNKR3MAPTLB_IDX(pChunk->Core.Key)].idChunk
5515	== pChunk->Core.Key)
5516	{
5517	pChunk = NULL;
5518	return 0;
5519	}
5520	#ifdef VBOX_STRICT
5521	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
5522	{
5523	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk != pChunk);
5524	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk != pChunk->Core.Key);
5525	}
5526	#endif
5527
5528	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.PhysTlbR3.aEntries); i++)
5529	if (pVM->pgm.s.PhysTlbR3.aEntries[i].pMap == pChunk)
5530	return 0;
5531
5532	pArg->pChunk = pChunk;
5533	return 0;
5534	}
5535
5536
5537	/**
5538	* Finds a good candidate for unmapping when the ring-3 mapping cache is full.
5539	*
5540	* The candidate will not be part of any TLBs, so no need to flush
5541	* anything afterwards.
5542	*
5543	* @returns Chunk id.
5544	* @param pVM The cross context VM structure.
5545	*/
5546	static int32_t pgmR3PhysChunkFindUnmapCandidate(PVM pVM)
5547	{
5548	PGM_LOCK_ASSERT_OWNER(pVM);
5549
5550	/*
5551	* Enumerate the age tree starting with the left most node.
5552	*/
5553	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
5554	PGMR3PHYSCHUNKUNMAPCB Args;
5555	Args.pVM = pVM;
5556	Args.pChunk = NULL;
5557	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkUnmapCandidateCallback, &Args);
5558	Assert(Args.pChunk);
5559	if (Args.pChunk)
5560	{
5561	Assert(Args.pChunk->cRefs == 0);
5562	Assert(Args.pChunk->cPermRefs == 0);
5563	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
5564	return Args.pChunk->Core.Key;
5565	}
5566
5567	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
5568	return INT32_MAX;
5569	}
5570
5571
5572	/**
5573	* Rendezvous callback used by pgmR3PhysUnmapChunk that unmaps a chunk
5574	*
5575	* This is only called on one of the EMTs while the other ones are waiting for
5576	* it to complete this function.
5577	*
5578	* @returns VINF_SUCCESS (VBox strict status code).
5579	* @param pVM The cross context VM structure.
5580	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
5581	* @param pvUser User pointer. Unused
5582	*
5583	*/
5584	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysUnmapChunkRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
5585	{
5586	int rc = VINF_SUCCESS;
5587	PGM_LOCK_VOID(pVM);
5588	NOREF(pVCpu); NOREF(pvUser);
5589
5590	if (pVM->pgm.s.ChunkR3Map.c >= pVM->pgm.s.ChunkR3Map.cMax)
5591	{
5592	/* Flush the pgm pool cache; call the internal rendezvous handler as we're already in a rendezvous handler here. */
5593	/** @todo also not really efficient to unmap a chunk that contains PD
5594	* or PT pages. */
5595	pgmR3PoolClearAllRendezvous(pVM, pVM->apCpusR3[0], NULL /* no need to flush the REM TLB as we already did that above */);
5596
5597	/*
5598	* Request the ring-0 part to unmap a chunk to make space in the mapping cache.
5599	*/
5600	GMMMAPUNMAPCHUNKREQ Req;
5601	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
5602	Req.Hdr.cbReq = sizeof(Req);
5603	Req.pvR3 = NULL;
5604	Req.idChunkMap = NIL_GMM_CHUNKID;
5605	Req.idChunkUnmap = pgmR3PhysChunkFindUnmapCandidate(pVM);
5606	if (Req.idChunkUnmap != INT32_MAX)
5607	{
5608	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkUnmap, a);
5609	rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
5610	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkUnmap, a);
5611	if (RT_SUCCESS(rc))
5612	{
5613	/*
5614	* Remove the unmapped one.
5615	*/
5616	PPGMCHUNKR3MAP pUnmappedChunk = (PPGMCHUNKR3MAP)RTAvlU32Remove(&pVM->pgm.s.ChunkR3Map.pTree, Req.idChunkUnmap);
5617	AssertRelease(pUnmappedChunk);
5618	AssertRelease(!pUnmappedChunk->cRefs);
5619	AssertRelease(!pUnmappedChunk->cPermRefs);
5620	pUnmappedChunk->pv = NULL;
5621	pUnmappedChunk->Core.Key = UINT32_MAX;
5622	MMR3HeapFree(pUnmappedChunk);
5623	pVM->pgm.s.ChunkR3Map.c--;
5624	pVM->pgm.s.cUnmappedChunks++;
5625
5626	/*
5627	* Flush dangling PGM pointers (R3 & R0 ptrs to GC physical addresses).
5628	*/
5629	/** @todo We should not flush chunks which include cr3 mappings. */
5630	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
5631	{
5632	PPGMCPU pPGM = &pVM->apCpusR3[idCpu]->pgm.s;
5633
5634	pPGM->pGst32BitPdR3 = NULL;
5635	pPGM->pGstPaePdptR3 = NULL;
5636	pPGM->pGstAmd64Pml4R3 = NULL;
5637	pPGM->pGst32BitPdR0 = NIL_RTR0PTR;
5638	pPGM->pGstPaePdptR0 = NIL_RTR0PTR;
5639	pPGM->pGstAmd64Pml4R0 = NIL_RTR0PTR;
5640	for (unsigned i = 0; i < RT_ELEMENTS(pPGM->apGstPaePDsR3); i++)
5641	{
5642	pPGM->apGstPaePDsR3[i] = NULL;
5643	pPGM->apGstPaePDsR0[i] = NIL_RTR0PTR;
5644	}
5645
5646	/* Flush REM TLBs. */
5647	CPUMSetChangedFlags(pVM->apCpusR3[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
5648	}
5649	}
5650	}
5651	}
5652	PGM_UNLOCK(pVM);
5653	return rc;
5654	}
5655
5656	/**
5657	* Unmap a chunk to free up virtual address space (request packet handler for pgmR3PhysChunkMap)
5658	*
5659	* @returns VBox status code.
5660	* @param pVM The cross context VM structure.
5661	*/
5662	static DECLCALLBACK(void) pgmR3PhysUnmapChunk(PVM pVM)
5663	{
5664	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysUnmapChunkRendezvous, NULL);
5665	AssertRC(rc);
5666	}
5667
5668
5669	/**
5670	* Maps the given chunk into the ring-3 mapping cache.
5671	*
5672	* This will call ring-0.
5673	*
5674	* @returns VBox status code.
5675	* @param pVM The cross context VM structure.
5676	* @param idChunk The chunk in question.
5677	* @param ppChunk Where to store the chunk tracking structure.
5678	*
5679	* @remarks Called from within the PGM critical section.
5680	* @remarks Can be called from any thread!
5681	*/
5682	int pgmR3PhysChunkMap(PVM pVM, uint32_t idChunk, PPPGMCHUNKR3MAP ppChunk)
5683	{
5684	int rc;
5685
5686	PGM_LOCK_ASSERT_OWNER(pVM);
5687
5688	/*
5689	* Move the chunk time forward.
5690	*/
5691	pVM->pgm.s.ChunkR3Map.iNow++;
5692	if (pVM->pgm.s.ChunkR3Map.iNow == 0)
5693	{
5694	pVM->pgm.s.ChunkR3Map.iNow = 4;
5695	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkAgeingRolloverCallback, NULL);
5696	}
5697
5698	/*
5699	* Allocate a new tracking structure first.
5700	*/
5701	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3HeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk));
5702	AssertReturn(pChunk, VERR_NO_MEMORY);
5703	pChunk->Core.Key = idChunk;
5704	pChunk->iLastUsed = pVM->pgm.s.ChunkR3Map.iNow;
5705
5706	/*
5707	* Request the ring-0 part to map the chunk in question.
5708	*/
5709	GMMMAPUNMAPCHUNKREQ Req;
5710	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
5711	Req.Hdr.cbReq = sizeof(Req);
5712	Req.pvR3 = NULL;
5713	Req.idChunkMap = idChunk;
5714	Req.idChunkUnmap = NIL_GMM_CHUNKID;
5715
5716	/* Must be callable from any thread, so can't use VMMR3CallR0. */
5717	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkMap, a);
5718	rc = SUPR3CallVMMR0Ex(VMCC_GET_VMR0_FOR_CALL(pVM), NIL_VMCPUID, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
5719	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkMap, a);
5720	if (RT_SUCCESS(rc))
5721	{
5722	pChunk->pv = Req.pvR3;
5723
5724	/*
5725	* If we're running out of virtual address space, then we should
5726	* unmap another chunk.
5727	*
5728	* Currently, an unmap operation requires that all other virtual CPUs
5729	* are idling and not by chance making use of the memory we're
5730	* unmapping. So, we create an async unmap operation here.
5731	*
5732	* Now, when creating or restoring a saved state this wont work very
5733	* well since we may want to restore all guest RAM + a little something.
5734	* So, we have to do the unmap synchronously. Fortunately for us
5735	* though, during these operations the other virtual CPUs are inactive
5736	* and it should be safe to do this.
5737	*/
5738	/** @todo Eventually we should lock all memory when used and do
5739	* map+unmap as one kernel call without any rendezvous or
5740	* other precautions. */
5741	if (pVM->pgm.s.ChunkR3Map.c + 1 >= pVM->pgm.s.ChunkR3Map.cMax)
5742	{
5743	switch (VMR3GetState(pVM))
5744	{
5745	case VMSTATE_LOADING:
5746	case VMSTATE_SAVING:
5747	{
5748	PVMCPU pVCpu = VMMGetCpu(pVM);
5749	if ( pVCpu
5750	&& pVM->pgm.s.cDeprecatedPageLocks == 0)
5751	{
5752	pgmR3PhysUnmapChunkRendezvous(pVM, pVCpu, NULL);
5753	break;
5754	}
5755	}
5756	RT_FALL_THRU();
5757	default:
5758	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysUnmapChunk, 1, pVM);
5759	AssertRC(rc);
5760	break;
5761	}
5762	}
5763
5764	/*
5765	* Update the tree. We must do this after any unmapping to make sure
5766	* the chunk we're going to return isn't unmapped by accident.
5767	*/
5768	AssertPtr(Req.pvR3);
5769	bool fRc = RTAvlU32Insert(&pVM->pgm.s.ChunkR3Map.pTree, &pChunk->Core);
5770	AssertRelease(fRc);
5771	pVM->pgm.s.ChunkR3Map.c++;
5772	pVM->pgm.s.cMappedChunks++;
5773	}
5774	else
5775	{
5776	/** @todo this may fail because of /proc/sys/vm/max_map_count, so we
5777	* should probably restrict ourselves on linux. */
5778	AssertRC(rc);
5779	MMR3HeapFree(pChunk);
5780	pChunk = NULL;
5781	}
5782
5783	*ppChunk = pChunk;
5784	return rc;
5785	}
5786
5787
5788	/**
5789	* Invalidates the TLB for the ring-3 mapping cache.
5790	*
5791	* @param pVM The cross context VM structure.
5792	*/
5793	VMMR3DECL(void) PGMR3PhysChunkInvalidateTLB(PVM pVM)
5794	{
5795	PGM_LOCK_VOID(pVM);
5796	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
5797	{
5798	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk = NIL_GMM_CHUNKID;
5799	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk = NULL;
5800	}
5801	/* The page map TLB references chunks, so invalidate that one too. */
5802	pgmPhysInvalidatePageMapTLB(pVM);
5803	PGM_UNLOCK(pVM);
5804	}
5805
5806
5807	/**
5808	* Response to VM_FF_PGM_NEED_HANDY_PAGES and helper for pgmPhysEnsureHandyPage.
5809	*
5810	* This function will also work the VM_FF_PGM_NO_MEMORY force action flag, to
5811	* signal and clear the out of memory condition. When called, this API is used
5812	* to try clear the condition when the user wants to resume.
5813	*
5814	* @returns The following VBox status codes.
5815	* @retval VINF_SUCCESS on success. FFs cleared.
5816	* @retval VINF_EM_NO_MEMORY if we're out of memory. The FF is not cleared in
5817	* this case and it gets accompanied by VM_FF_PGM_NO_MEMORY.
5818	*
5819	* @param pVM The cross context VM structure.
5820	*
5821	* @remarks The VINF_EM_NO_MEMORY status is for the benefit of the FF processing
5822	* in EM.cpp and shouldn't be propagated outside TRPM, HM, EM and
5823	* pgmPhysEnsureHandyPage. There is one exception to this in the \#PF
5824	* handler.
5825	*/
5826	VMMR3DECL(int) PGMR3PhysAllocateHandyPages(PVM pVM)
5827	{
5828	PGM_LOCK_VOID(pVM);
5829
5830	/*
5831	* Allocate more pages, noting down the index of the first new page.
5832	*/
5833	uint32_t iClear = pVM->pgm.s.cHandyPages;
5834	AssertMsgReturn(iClear <= RT_ELEMENTS(pVM->pgm.s.aHandyPages), ("%d", iClear), VERR_PGM_HANDY_PAGE_IPE);
5835	Log(("PGMR3PhysAllocateHandyPages: %d -> %d\n", iClear, RT_ELEMENTS(pVM->pgm.s.aHandyPages)));
5836	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
5837	/** @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) */
5838	if ( rc == VERR_GMM_HIT_VM_ACCOUNT_LIMIT
5839	&& pVM->pgm.s.cHandyPages > 0)
5840	{
5841	/* Still handy pages left, so don't panic. */
5842	rc = VINF_SUCCESS;
5843	}
5844
5845	if (RT_SUCCESS(rc))
5846	{
5847	AssertMsg(rc == VINF_SUCCESS, ("%Rrc\n", rc));
5848	Assert(pVM->pgm.s.cHandyPages > 0);
5849	#ifdef VBOX_STRICT
5850	uint32_t i;
5851	for (i = iClear; i < pVM->pgm.s.cHandyPages; i++)
5852	if ( pVM->pgm.s.aHandyPages[i].idPage == NIL_GMM_PAGEID
5853	\|\| pVM->pgm.s.aHandyPages[i].idSharedPage != NIL_GMM_PAGEID
5854	\|\| (pVM->pgm.s.aHandyPages[i].HCPhysGCPhys & PAGE_OFFSET_MASK))
5855	break;
5856	if (i != pVM->pgm.s.cHandyPages)
5857	{
5858	RTAssertMsg1Weak(NULL, __LINE__, __FILE__, __FUNCTION__);
5859	RTAssertMsg2Weak("i=%d iClear=%d cHandyPages=%d\n", i, iClear, pVM->pgm.s.cHandyPages);
5860	for (uint32_t j = iClear; j < pVM->pgm.s.cHandyPages; j++)
5861	RTAssertMsg2Add("%03d: idPage=%d HCPhysGCPhys=%RHp idSharedPage=%d%s\n", j,
5862	pVM->pgm.s.aHandyPages[j].idPage,
5863	pVM->pgm.s.aHandyPages[j].HCPhysGCPhys,
5864	pVM->pgm.s.aHandyPages[j].idSharedPage,
5865	j == i ? " <---" : "");
5866	RTAssertPanic();
5867	}
5868	#endif
5869	}
5870	else
5871	{
5872	/*
5873	* We should never get here unless there is a genuine shortage of
5874	* memory (or some internal error). Flag the error so the VM can be
5875	* suspended ASAP and the user informed. If we're totally out of
5876	* handy pages we will return failure.
5877	*/
5878	/* Report the failure. */
5879	LogRel(("PGM: Failed to procure handy pages; rc=%Rrc cHandyPages=%#x\n"
5880	" cAllPages=%#x cPrivatePages=%#x cSharedPages=%#x cZeroPages=%#x\n",
5881	rc, pVM->pgm.s.cHandyPages,
5882	pVM->pgm.s.cAllPages, pVM->pgm.s.cPrivatePages, pVM->pgm.s.cSharedPages, pVM->pgm.s.cZeroPages));
5883
5884	if ( rc != VERR_NO_MEMORY
5885	&& rc != VERR_NO_PHYS_MEMORY
5886	&& rc != VERR_LOCK_FAILED)
5887	for (uint32_t i = 0; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
5888	{
5889	LogRel(("PGM: aHandyPages[#%#04x] = {.HCPhysGCPhys=%RHp, .idPage=%#08x, .idSharedPage=%#08x}\n",
5890	i, pVM->pgm.s.aHandyPages[i].HCPhysGCPhys, pVM->pgm.s.aHandyPages[i].idPage,
5891	pVM->pgm.s.aHandyPages[i].idSharedPage));
5892	uint32_t const idPage = pVM->pgm.s.aHandyPages[i].idPage;
5893	if (idPage != NIL_GMM_PAGEID)
5894	{
5895	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
5896	pRam;
5897	pRam = pRam->pNextR3)
5898	{
5899	uint32_t const cPages = pRam->cb >> PAGE_SHIFT;
5900	for (uint32_t iPage = 0; iPage < cPages; iPage++)
5901	if (PGM_PAGE_GET_PAGEID(&pRam->aPages[iPage]) == idPage)
5902	LogRel(("PGM: Used by %RGp %R[pgmpage] (%s)\n",
5903	pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pRam->aPages[iPage], pRam->pszDesc));
5904	}
5905	}
5906	}
5907
5908	if (rc == VERR_NO_MEMORY)
5909	{
5910	uint64_t cbHostRamAvail = 0;
5911	int rc2 = RTSystemQueryAvailableRam(&cbHostRamAvail);
5912	if (RT_SUCCESS(rc2))
5913	LogRel(("Host RAM: %RU64MB available\n", cbHostRamAvail / _1M));
5914	else
5915	LogRel(("Cannot determine the amount of available host memory\n"));
5916	}
5917
5918	/* Set the FFs and adjust rc. */
5919	VM_FF_SET(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
5920	VM_FF_SET(pVM, VM_FF_PGM_NO_MEMORY);
5921	if ( rc == VERR_NO_MEMORY
5922	\|\| rc == VERR_NO_PHYS_MEMORY
5923	\|\| rc == VERR_LOCK_FAILED)
5924	rc = VINF_EM_NO_MEMORY;
5925	}
5926
5927	PGM_UNLOCK(pVM);
5928	return rc;
5929	}
5930
5931
5932	/*********************************************************************************************************************************
5933	* Other Stuff *
5934	*********************************************************************************************************************************/
5935
5936	/**
5937	* Sets the Address Gate 20 state.
5938	*
5939	* @param pVCpu The cross context virtual CPU structure.
5940	* @param fEnable True if the gate should be enabled.
5941	* False if the gate should be disabled.
5942	*/
5943	VMMDECL(void) PGMR3PhysSetA20(PVMCPU pVCpu, bool fEnable)
5944	{
5945	LogFlow(("PGMR3PhysSetA20 %d (was %d)\n", fEnable, pVCpu->pgm.s.fA20Enabled));
5946	if (pVCpu->pgm.s.fA20Enabled != fEnable)
5947	{
5948	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX
5949	PCCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
5950	if ( CPUMIsGuestInVmxRootMode(pCtx)
5951	&& !fEnable)
5952	{
5953	Log(("Cannot enter A20M mode while in VMX root mode\n"));
5954	return;
5955	}
5956	#endif
5957	pVCpu->pgm.s.fA20Enabled = fEnable;
5958	pVCpu->pgm.s.GCPhysA20Mask = ~((RTGCPHYS)!fEnable << 20);
5959	if (VM_IS_NEM_ENABLED(pVCpu->CTX_SUFF(pVM)))
5960	NEMR3NotifySetA20(pVCpu, fEnable);
5961	#ifdef PGM_WITH_A20
5962	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
5963	pgmR3RefreshShadowModeAfterA20Change(pVCpu);
5964	HMFlushTlb(pVCpu);
5965	#endif
5966	IEMTlbInvalidateAllPhysical(pVCpu);
5967	STAM_REL_COUNTER_INC(&pVCpu->pgm.s.cA20Changes);
5968	}
5969	}
5970

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PGMPhys.cpp@ 92441

Download in other formats: