PGMPhys.cpp@ 91688

Last change on this file since 91688 was 91581, checked in by vboxsync, 3 years ago
VMM/PGMR3PhysRegisterRam: Must release lock when returning VERR_PGM_RAM_CONFLICT.
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 198.3 KB

Line
1	/* $Id: PGMPhys.cpp 91581 2021-10-06 07:22:22Z 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	* PGMR3PhysReadU8-64
60	* PGMR3PhysWriteU8-64
61	*/
62	#define PGMPHYSFN_READNAME PGMR3PhysReadU8
63	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU8
64	#define PGMPHYS_DATASIZE 1
65	#define PGMPHYS_DATATYPE uint8_t
66	#include "PGMPhysRWTmpl.h"
67
68	#define PGMPHYSFN_READNAME PGMR3PhysReadU16
69	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU16
70	#define PGMPHYS_DATASIZE 2
71	#define PGMPHYS_DATATYPE uint16_t
72	#include "PGMPhysRWTmpl.h"
73
74	#define PGMPHYSFN_READNAME PGMR3PhysReadU32
75	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU32
76	#define PGMPHYS_DATASIZE 4
77	#define PGMPHYS_DATATYPE uint32_t
78	#include "PGMPhysRWTmpl.h"
79
80	#define PGMPHYSFN_READNAME PGMR3PhysReadU64
81	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU64
82	#define PGMPHYS_DATASIZE 8
83	#define PGMPHYS_DATATYPE uint64_t
84	#include "PGMPhysRWTmpl.h"
85
86
87	/**
88	* EMT worker for PGMR3PhysReadExternal.
89	*/
90	static DECLCALLBACK(int) pgmR3PhysReadExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, void *pvBuf, size_t cbRead,
91	PGMACCESSORIGIN enmOrigin)
92	{
93	VBOXSTRICTRC rcStrict = PGMPhysRead(pVM, *pGCPhys, pvBuf, cbRead, enmOrigin);
94	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
95	return VINF_SUCCESS;
96	}
97
98
99	/**
100	* Read from physical memory, external users.
101	*
102	* @returns VBox status code.
103	* @retval VINF_SUCCESS.
104	*
105	* @param pVM The cross context VM structure.
106	* @param GCPhys Physical address to read from.
107	* @param pvBuf Where to read into.
108	* @param cbRead How many bytes to read.
109	* @param enmOrigin Who is calling.
110	*
111	* @thread Any but EMTs.
112	*/
113	VMMR3DECL(int) PGMR3PhysReadExternal(PVM pVM, RTGCPHYS GCPhys, void *pvBuf, size_t cbRead, PGMACCESSORIGIN enmOrigin)
114	{
115	VM_ASSERT_OTHER_THREAD(pVM);
116
117	AssertMsgReturn(cbRead > 0, ("don't even think about reading zero bytes!\n"), VINF_SUCCESS);
118	LogFlow(("PGMR3PhysReadExternal: %RGp %d\n", GCPhys, cbRead));
119
120	PGM_LOCK_VOID(pVM);
121
122	/*
123	* Copy loop on ram ranges.
124	*/
125	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
126	for (;;)
127	{
128	/* Inside range or not? */
129	if (pRam && GCPhys >= pRam->GCPhys)
130	{
131	/*
132	* Must work our way thru this page by page.
133	*/
134	RTGCPHYS off = GCPhys - pRam->GCPhys;
135	while (off < pRam->cb)
136	{
137	unsigned iPage = off >> PAGE_SHIFT;
138	PPGMPAGE pPage = &pRam->aPages[iPage];
139
140	/*
141	* If the page has an ALL access handler, we'll have to
142	* delegate the job to EMT.
143	*/
144	if ( PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
145	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
146	{
147	PGM_UNLOCK(pVM);
148
149	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysReadExternalEMT, 5,
150	pVM, &GCPhys, pvBuf, cbRead, enmOrigin);
151	}
152	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
153
154	/*
155	* Simple stuff, go ahead.
156	*/
157	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
158	if (cb > cbRead)
159	cb = cbRead;
160	PGMPAGEMAPLOCK PgMpLck;
161	const void *pvSrc;
162	int rc = pgmPhysGCPhys2CCPtrInternalReadOnly(pVM, pPage, pRam->GCPhys + off, &pvSrc, &PgMpLck);
163	if (RT_SUCCESS(rc))
164	{
165	memcpy(pvBuf, pvSrc, cb);
166	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
167	}
168	else
169	{
170	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternalReadOnly failed on %RGp / %R[pgmpage] -> %Rrc\n",
171	pRam->GCPhys + off, pPage, rc));
172	memset(pvBuf, 0xff, cb);
173	}
174
175	/* next page */
176	if (cb >= cbRead)
177	{
178	PGM_UNLOCK(pVM);
179	return VINF_SUCCESS;
180	}
181	cbRead -= cb;
182	off += cb;
183	GCPhys += cb;
184	pvBuf = (char *)pvBuf + cb;
185	} /* walk pages in ram range. */
186	}
187	else
188	{
189	LogFlow(("PGMPhysRead: Unassigned %RGp size=%u\n", GCPhys, cbRead));
190
191	/*
192	* Unassigned address space.
193	*/
194	size_t cb = pRam ? pRam->GCPhys - GCPhys : ~(size_t)0;
195	if (cb >= cbRead)
196	{
197	memset(pvBuf, 0xff, cbRead);
198	break;
199	}
200	memset(pvBuf, 0xff, cb);
201
202	cbRead -= cb;
203	pvBuf = (char *)pvBuf + cb;
204	GCPhys += cb;
205	}
206
207	/* Advance range if necessary. */
208	while (pRam && GCPhys > pRam->GCPhysLast)
209	pRam = pRam->CTX_SUFF(pNext);
210	} /* Ram range walk */
211
212	PGM_UNLOCK(pVM);
213
214	return VINF_SUCCESS;
215	}
216
217
218	/**
219	* EMT worker for PGMR3PhysWriteExternal.
220	*/
221	static DECLCALLBACK(int) pgmR3PhysWriteExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, const void *pvBuf, size_t cbWrite,
222	PGMACCESSORIGIN enmOrigin)
223	{
224	/** @todo VERR_EM_NO_MEMORY */
225	VBOXSTRICTRC rcStrict = PGMPhysWrite(pVM, *pGCPhys, pvBuf, cbWrite, enmOrigin);
226	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
227	return VINF_SUCCESS;
228	}
229
230
231	/**
232	* Write to physical memory, external users.
233	*
234	* @returns VBox status code.
235	* @retval VINF_SUCCESS.
236	* @retval VERR_EM_NO_MEMORY.
237	*
238	* @param pVM The cross context VM structure.
239	* @param GCPhys Physical address to write to.
240	* @param pvBuf What to write.
241	* @param cbWrite How many bytes to write.
242	* @param enmOrigin Who is calling.
243	*
244	* @thread Any but EMTs.
245	*/
246	VMMDECL(int) PGMR3PhysWriteExternal(PVM pVM, RTGCPHYS GCPhys, const void *pvBuf, size_t cbWrite, PGMACCESSORIGIN enmOrigin)
247	{
248	VM_ASSERT_OTHER_THREAD(pVM);
249
250	AssertMsg(!pVM->pgm.s.fNoMorePhysWrites,
251	("Calling PGMR3PhysWriteExternal after pgmR3Save()! GCPhys=%RGp cbWrite=%#x enmOrigin=%d\n",
252	GCPhys, cbWrite, enmOrigin));
253	AssertMsgReturn(cbWrite > 0, ("don't even think about writing zero bytes!\n"), VINF_SUCCESS);
254	LogFlow(("PGMR3PhysWriteExternal: %RGp %d\n", GCPhys, cbWrite));
255
256	PGM_LOCK_VOID(pVM);
257
258	/*
259	* Copy loop on ram ranges, stop when we hit something difficult.
260	*/
261	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
262	for (;;)
263	{
264	/* Inside range or not? */
265	if (pRam && GCPhys >= pRam->GCPhys)
266	{
267	/*
268	* Must work our way thru this page by page.
269	*/
270	RTGCPTR off = GCPhys - pRam->GCPhys;
271	while (off < pRam->cb)
272	{
273	RTGCPTR iPage = off >> PAGE_SHIFT;
274	PPGMPAGE pPage = &pRam->aPages[iPage];
275
276	/*
277	* Is the page problematic, we have to do the work on the EMT.
278	*
279	* Allocating writable pages and access handlers are
280	* problematic, write monitored pages are simple and can be
281	* dealt with here.
282	*/
283	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
284	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
285	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
286	{
287	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
288	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
289	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, GCPhys);
290	else
291	{
292	PGM_UNLOCK(pVM);
293
294	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysWriteExternalEMT, 5,
295	pVM, &GCPhys, pvBuf, cbWrite, enmOrigin);
296	}
297	}
298	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
299
300	/*
301	* Simple stuff, go ahead.
302	*/
303	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
304	if (cb > cbWrite)
305	cb = cbWrite;
306	PGMPAGEMAPLOCK PgMpLck;
307	void *pvDst;
308	int rc = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, pRam->GCPhys + off, &pvDst, &PgMpLck);
309	if (RT_SUCCESS(rc))
310	{
311	memcpy(pvDst, pvBuf, cb);
312	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
313	}
314	else
315	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternal failed on %RGp / %R[pgmpage] -> %Rrc\n",
316	pRam->GCPhys + off, pPage, rc));
317
318	/* next page */
319	if (cb >= cbWrite)
320	{
321	PGM_UNLOCK(pVM);
322	return VINF_SUCCESS;
323	}
324
325	cbWrite -= cb;
326	off += cb;
327	GCPhys += cb;
328	pvBuf = (const char *)pvBuf + cb;
329	} /* walk pages in ram range */
330	}
331	else
332	{
333	/*
334	* Unassigned address space, skip it.
335	*/
336	if (!pRam)
337	break;
338	size_t cb = pRam->GCPhys - GCPhys;
339	if (cb >= cbWrite)
340	break;
341	cbWrite -= cb;
342	pvBuf = (const char *)pvBuf + cb;
343	GCPhys += cb;
344	}
345
346	/* Advance range if necessary. */
347	while (pRam && GCPhys > pRam->GCPhysLast)
348	pRam = pRam->CTX_SUFF(pNext);
349	} /* Ram range walk */
350
351	PGM_UNLOCK(pVM);
352	return VINF_SUCCESS;
353	}
354
355
356	/**
357	* VMR3ReqCall worker for PGMR3PhysGCPhys2CCPtrExternal to make pages writable.
358	*
359	* @returns see PGMR3PhysGCPhys2CCPtrExternal
360	* @param pVM The cross context VM structure.
361	* @param pGCPhys Pointer to the guest physical address.
362	* @param ppv Where to store the mapping address.
363	* @param pLock Where to store the lock.
364	*/
365	static DECLCALLBACK(int) pgmR3PhysGCPhys2CCPtrDelegated(PVM pVM, PRTGCPHYS pGCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
366	{
367	/*
368	* Just hand it to PGMPhysGCPhys2CCPtr and check that it's not a page with
369	* an access handler after it succeeds.
370	*/
371	int rc = PGM_LOCK(pVM);
372	AssertRCReturn(rc, rc);
373
374	rc = PGMPhysGCPhys2CCPtr(pVM, *pGCPhys, ppv, pLock);
375	if (RT_SUCCESS(rc))
376	{
377	PPGMPAGEMAPTLBE pTlbe;
378	int rc2 = pgmPhysPageQueryTlbe(pVM, *pGCPhys, &pTlbe);
379	AssertFatalRC(rc2);
380	PPGMPAGE pPage = pTlbe->pPage;
381	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
382	{
383	PGMPhysReleasePageMappingLock(pVM, pLock);
384	rc = VERR_PGM_PHYS_PAGE_RESERVED;
385	}
386	else if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
387	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
388	\|\| pgmPoolIsDirtyPage(pVM, *pGCPhys)
389	#endif
390	)
391	{
392	/* We must flush any corresponding pgm pool page here, otherwise we'll
393	* not be informed about writes and keep bogus gst->shw mappings around.
394	*/
395	pgmPoolFlushPageByGCPhys(pVM, *pGCPhys);
396	Assert(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage));
397	/** @todo r=bird: return VERR_PGM_PHYS_PAGE_RESERVED here if it still has
398	* active handlers, see the PGMR3PhysGCPhys2CCPtrExternal docs. */
399	}
400	}
401
402	PGM_UNLOCK(pVM);
403	return rc;
404	}
405
406
407	/**
408	* Requests the mapping of a guest page into ring-3, external threads.
409	*
410	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
411	* release it.
412	*
413	* This API will assume your intention is to write to the page, and will
414	* therefore replace shared and zero pages. If you do not intend to modify the
415	* page, use the PGMR3PhysGCPhys2CCPtrReadOnlyExternal() API.
416	*
417	* @returns VBox status code.
418	* @retval VINF_SUCCESS on success.
419	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
420	* backing or if the page has any active access handlers. The caller
421	* must fall back on using PGMR3PhysWriteExternal.
422	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
423	*
424	* @param pVM The cross context VM structure.
425	* @param GCPhys The guest physical address of the page that should be mapped.
426	* @param ppv Where to store the address corresponding to GCPhys.
427	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
428	*
429	* @remark Avoid calling this API from within critical sections (other than the
430	* PGM one) because of the deadlock risk when we have to delegating the
431	* task to an EMT.
432	* @thread Any.
433	*/
434	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrExternal(PVM pVM, RTGCPHYS GCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
435	{
436	AssertPtr(ppv);
437	AssertPtr(pLock);
438
439	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
440
441	int rc = PGM_LOCK(pVM);
442	AssertRCReturn(rc, rc);
443
444	/*
445	* Query the Physical TLB entry for the page (may fail).
446	*/
447	PPGMPAGEMAPTLBE pTlbe;
448	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
449	if (RT_SUCCESS(rc))
450	{
451	PPGMPAGE pPage = pTlbe->pPage;
452	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
453	rc = VERR_PGM_PHYS_PAGE_RESERVED;
454	else
455	{
456	/*
457	* If the page is shared, the zero page, or being write monitored
458	* it must be converted to an page that's writable if possible.
459	* We can only deal with write monitored pages here, the rest have
460	* to be on an EMT.
461	*/
462	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
463	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
464	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
465	\|\| pgmPoolIsDirtyPage(pVM, GCPhys)
466	#endif
467	)
468	{
469	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
470	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
471	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
472	&& !pgmPoolIsDirtyPage(pVM, GCPhys) /** @todo we're very likely doing this twice. */
473	#endif
474	)
475	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, GCPhys);
476	else
477	{
478	PGM_UNLOCK(pVM);
479
480	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
481	pVM, &GCPhys, ppv, pLock);
482	}
483	}
484
485	/*
486	* Now, just perform the locking and calculate the return address.
487	*/
488	PPGMPAGEMAP pMap = pTlbe->pMap;
489	if (pMap)
490	pMap->cRefs++;
491
492	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
493	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
494	{
495	if (cLocks == 0)
496	pVM->pgm.s.cWriteLockedPages++;
497	PGM_PAGE_INC_WRITE_LOCKS(pPage);
498	}
499	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
500	{
501	PGM_PAGE_INC_WRITE_LOCKS(pPage);
502	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", GCPhys, pPage));
503	if (pMap)
504	pMap->cRefs++; /* Extra ref to prevent it from going away. */
505	}
506
507	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
508	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
509	pLock->pvMap = pMap;
510	}
511	}
512
513	PGM_UNLOCK(pVM);
514	return rc;
515	}
516
517
518	/**
519	* Requests the mapping of a guest page into ring-3, external threads.
520	*
521	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
522	* release it.
523	*
524	* @returns VBox status code.
525	* @retval VINF_SUCCESS on success.
526	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
527	* backing or if the page as an active ALL access handler. The caller
528	* must fall back on using PGMPhysRead.
529	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
530	*
531	* @param pVM The cross context VM structure.
532	* @param GCPhys The guest physical address of the page that should be mapped.
533	* @param ppv Where to store the address corresponding to GCPhys.
534	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
535	*
536	* @remark Avoid calling this API from within critical sections (other than
537	* the PGM one) because of the deadlock risk.
538	* @thread Any.
539	*/
540	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrReadOnlyExternal(PVM pVM, RTGCPHYS GCPhys, void const **ppv, PPGMPAGEMAPLOCK pLock)
541	{
542	int rc = PGM_LOCK(pVM);
543	AssertRCReturn(rc, rc);
544
545	/*
546	* Query the Physical TLB entry for the page (may fail).
547	*/
548	PPGMPAGEMAPTLBE pTlbe;
549	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
550	if (RT_SUCCESS(rc))
551	{
552	PPGMPAGE pPage = pTlbe->pPage;
553	#if 1
554	/* MMIO pages doesn't have any readable backing. */
555	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
556	rc = VERR_PGM_PHYS_PAGE_RESERVED;
557	#else
558	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
559	rc = VERR_PGM_PHYS_PAGE_RESERVED;
560	#endif
561	else
562	{
563	/*
564	* Now, just perform the locking and calculate the return address.
565	*/
566	PPGMPAGEMAP pMap = pTlbe->pMap;
567	if (pMap)
568	pMap->cRefs++;
569
570	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
571	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
572	{
573	if (cLocks == 0)
574	pVM->pgm.s.cReadLockedPages++;
575	PGM_PAGE_INC_READ_LOCKS(pPage);
576	}
577	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
578	{
579	PGM_PAGE_INC_READ_LOCKS(pPage);
580	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", GCPhys, pPage));
581	if (pMap)
582	pMap->cRefs++; /* Extra ref to prevent it from going away. */
583	}
584
585	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
586	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
587	pLock->pvMap = pMap;
588	}
589	}
590
591	PGM_UNLOCK(pVM);
592	return rc;
593	}
594
595
596	/**
597	* Requests the mapping of multiple guest page into ring-3, external threads.
598	*
599	* When you're done with the pages, call PGMPhysBulkReleasePageMappingLock()
600	* ASAP to release them.
601	*
602	* This API will assume your intention is to write to the pages, and will
603	* therefore replace shared and zero pages. If you do not intend to modify the
604	* pages, use the PGMR3PhysBulkGCPhys2CCPtrReadOnlyExternal() API.
605	*
606	* @returns VBox status code.
607	* @retval VINF_SUCCESS on success.
608	* @retval VERR_PGM_PHYS_PAGE_RESERVED if any of the pages has no physical
609	* backing or if any of the pages the page has any active access
610	* handlers. The caller must fall back on using PGMR3PhysWriteExternal.
611	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if @a paGCPhysPages contains
612	* an invalid physical address.
613	*
614	* @param pVM The cross context VM structure.
615	* @param cPages Number of pages to lock.
616	* @param paGCPhysPages The guest physical address of the pages that
617	* should be mapped (@a cPages entries).
618	* @param papvPages Where to store the ring-3 mapping addresses
619	* corresponding to @a paGCPhysPages.
620	* @param paLocks Where to store the locking information that
621	* pfnPhysBulkReleasePageMappingLock needs (@a cPages
622	* in length).
623	*
624	* @remark Avoid calling this API from within critical sections (other than the
625	* PGM one) because of the deadlock risk when we have to delegating the
626	* task to an EMT.
627	* @thread Any.
628	*/
629	VMMR3DECL(int) PGMR3PhysBulkGCPhys2CCPtrExternal(PVM pVM, uint32_t cPages, PCRTGCPHYS paGCPhysPages,
630	void **papvPages, PPGMPAGEMAPLOCK paLocks)
631	{
632	Assert(cPages > 0);
633	AssertPtr(papvPages);
634	AssertPtr(paLocks);
635
636	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
637
638	int rc = PGM_LOCK(pVM);
639	AssertRCReturn(rc, rc);
640
641	/*
642	* Lock the pages one by one.
643	* The loop body is similar to PGMR3PhysGCPhys2CCPtrExternal.
644	*/
645	int32_t cNextYield = 128;
646	uint32_t iPage;
647	for (iPage = 0; iPage < cPages; iPage++)
648	{
649	if (--cNextYield > 0)
650	{ /* likely */ }
651	else
652	{
653	PGM_UNLOCK(pVM);
654	ASMNopPause();
655	PGM_LOCK_VOID(pVM);
656	cNextYield = 128;
657	}
658
659	/*
660	* Query the Physical TLB entry for the page (may fail).
661	*/
662	PPGMPAGEMAPTLBE pTlbe;
663	rc = pgmPhysPageQueryTlbe(pVM, paGCPhysPages[iPage], &pTlbe);
664	if (RT_SUCCESS(rc))
665	{ }
666	else
667	break;
668	PPGMPAGE pPage = pTlbe->pPage;
669
670	/*
671	* No MMIO or active access handlers.
672	*/
673	if ( !PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage)
674	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
675	{ }
676	else
677	{
678	rc = VERR_PGM_PHYS_PAGE_RESERVED;
679	break;
680	}
681
682	/*
683	* The page must be in the allocated state and not be a dirty pool page.
684	* We can handle converting a write monitored page to an allocated one, but
685	* anything more complicated must be delegated to an EMT.
686	*/
687	bool fDelegateToEmt = false;
688	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_ALLOCATED)
689	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
690	fDelegateToEmt = pgmPoolIsDirtyPage(pVM, paGCPhysPages[iPage]);
691	#else
692	fDelegateToEmt = false;
693	#endif
694	else if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
695	{
696	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
697	if (!pgmPoolIsDirtyPage(pVM, paGCPhysPages[iPage]))
698	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, paGCPhysPages[iPage]);
699	else
700	fDelegateToEmt = true;
701	#endif
702	}
703	else
704	fDelegateToEmt = true;
705	if (!fDelegateToEmt)
706	{ }
707	else
708	{
709	/* We could do this delegation in bulk, but considered too much work vs gain. */
710	PGM_UNLOCK(pVM);
711	rc = VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
712	pVM, &paGCPhysPages[iPage], &papvPages[iPage], &paLocks[iPage]);
713	PGM_LOCK_VOID(pVM);
714	if (RT_FAILURE(rc))
715	break;
716	cNextYield = 128;
717	}
718
719	/*
720	* Now, just perform the locking and address calculation.
721	*/
722	PPGMPAGEMAP pMap = pTlbe->pMap;
723	if (pMap)
724	pMap->cRefs++;
725
726	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
727	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
728	{
729	if (cLocks == 0)
730	pVM->pgm.s.cWriteLockedPages++;
731	PGM_PAGE_INC_WRITE_LOCKS(pPage);
732	}
733	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
734	{
735	PGM_PAGE_INC_WRITE_LOCKS(pPage);
736	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", paGCPhysPages[iPage], pPage));
737	if (pMap)
738	pMap->cRefs++; /* Extra ref to prevent it from going away. */
739	}
740
741	papvPages[iPage] = (void *)((uintptr_t)pTlbe->pv \| (uintptr_t)(paGCPhysPages[iPage] & PAGE_OFFSET_MASK));
742	paLocks[iPage].uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
743	paLocks[iPage].pvMap = pMap;
744	}
745
746	PGM_UNLOCK(pVM);
747
748	/*
749	* On failure we must unlock any pages we managed to get already.
750	*/
751	if (RT_FAILURE(rc) && iPage > 0)
752	PGMPhysBulkReleasePageMappingLocks(pVM, iPage, paLocks);
753
754	return rc;
755	}
756
757
758	/**
759	* Requests the mapping of multiple guest page into ring-3, for reading only,
760	* external threads.
761	*
762	* When you're done with the pages, call PGMPhysReleasePageMappingLock() ASAP
763	* to release them.
764	*
765	* @returns VBox status code.
766	* @retval VINF_SUCCESS on success.
767	* @retval VERR_PGM_PHYS_PAGE_RESERVED if any of the pages has no physical
768	* backing or if any of the pages the page has an active ALL access
769	* handler. The caller must fall back on using PGMR3PhysWriteExternal.
770	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if @a paGCPhysPages contains
771	* an invalid physical address.
772	*
773	* @param pVM The cross context VM structure.
774	* @param cPages Number of pages to lock.
775	* @param paGCPhysPages The guest physical address of the pages that
776	* should be mapped (@a cPages entries).
777	* @param papvPages Where to store the ring-3 mapping addresses
778	* corresponding to @a paGCPhysPages.
779	* @param paLocks Where to store the lock information that
780	* pfnPhysReleasePageMappingLock needs (@a cPages
781	* in length).
782	*
783	* @remark Avoid calling this API from within critical sections (other than
784	* the PGM one) because of the deadlock risk.
785	* @thread Any.
786	*/
787	VMMR3DECL(int) PGMR3PhysBulkGCPhys2CCPtrReadOnlyExternal(PVM pVM, uint32_t cPages, PCRTGCPHYS paGCPhysPages,
788	void const **papvPages, PPGMPAGEMAPLOCK paLocks)
789	{
790	Assert(cPages > 0);
791	AssertPtr(papvPages);
792	AssertPtr(paLocks);
793
794	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
795
796	int rc = PGM_LOCK(pVM);
797	AssertRCReturn(rc, rc);
798
799	/*
800	* Lock the pages one by one.
801	* The loop body is similar to PGMR3PhysGCPhys2CCPtrReadOnlyExternal.
802	*/
803	int32_t cNextYield = 256;
804	uint32_t iPage;
805	for (iPage = 0; iPage < cPages; iPage++)
806	{
807	if (--cNextYield > 0)
808	{ /* likely */ }
809	else
810	{
811	PGM_UNLOCK(pVM);
812	ASMNopPause();
813	PGM_LOCK_VOID(pVM);
814	cNextYield = 256;
815	}
816
817	/*
818	* Query the Physical TLB entry for the page (may fail).
819	*/
820	PPGMPAGEMAPTLBE pTlbe;
821	rc = pgmPhysPageQueryTlbe(pVM, paGCPhysPages[iPage], &pTlbe);
822	if (RT_SUCCESS(rc))
823	{ }
824	else
825	break;
826	PPGMPAGE pPage = pTlbe->pPage;
827
828	/*
829	* No MMIO or active all access handlers, everything else can be accessed.
830	*/
831	if ( !PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage)
832	&& !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
833	{ }
834	else
835	{
836	rc = VERR_PGM_PHYS_PAGE_RESERVED;
837	break;
838	}
839
840	/*
841	* Now, just perform the locking and address calculation.
842	*/
843	PPGMPAGEMAP pMap = pTlbe->pMap;
844	if (pMap)
845	pMap->cRefs++;
846
847	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
848	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
849	{
850	if (cLocks == 0)
851	pVM->pgm.s.cReadLockedPages++;
852	PGM_PAGE_INC_READ_LOCKS(pPage);
853	}
854	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
855	{
856	PGM_PAGE_INC_READ_LOCKS(pPage);
857	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", paGCPhysPages[iPage], pPage));
858	if (pMap)
859	pMap->cRefs++; /* Extra ref to prevent it from going away. */
860	}
861
862	papvPages[iPage] = (void *)((uintptr_t)pTlbe->pv \| (uintptr_t)(paGCPhysPages[iPage] & PAGE_OFFSET_MASK));
863	paLocks[iPage].uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
864	paLocks[iPage].pvMap = pMap;
865	}
866
867	PGM_UNLOCK(pVM);
868
869	/*
870	* On failure we must unlock any pages we managed to get already.
871	*/
872	if (RT_FAILURE(rc) && iPage > 0)
873	PGMPhysBulkReleasePageMappingLocks(pVM, iPage, paLocks);
874
875	return rc;
876	}
877
878
879	#define MAKE_LEAF(a_pNode) \
880	do { \
881	(a_pNode)->pLeftR3 = NIL_RTR3PTR; \
882	(a_pNode)->pRightR3 = NIL_RTR3PTR; \
883	(a_pNode)->pLeftR0 = NIL_RTR0PTR; \
884	(a_pNode)->pRightR0 = NIL_RTR0PTR; \
885	} while (0)
886
887	#define INSERT_LEFT(a_pParent, a_pNode) \
888	do { \
889	(a_pParent)->pLeftR3 = (a_pNode); \
890	(a_pParent)->pLeftR0 = (a_pNode)->pSelfR0; \
891	} while (0)
892	#define INSERT_RIGHT(a_pParent, a_pNode) \
893	do { \
894	(a_pParent)->pRightR3 = (a_pNode); \
895	(a_pParent)->pRightR0 = (a_pNode)->pSelfR0; \
896	} while (0)
897
898
899	/**
900	* Recursive tree builder.
901	*
902	* @param ppRam Pointer to the iterator variable.
903	* @param iDepth The current depth. Inserts a leaf node if 0.
904	*/
905	static PPGMRAMRANGE pgmR3PhysRebuildRamRangeSearchTreesRecursively(PPGMRAMRANGE *ppRam, int iDepth)
906	{
907	PPGMRAMRANGE pRam;
908	if (iDepth <= 0)
909	{
910	/*
911	* Leaf node.
912	*/
913	pRam = *ppRam;
914	if (pRam)
915	{
916	*ppRam = pRam->pNextR3;
917	MAKE_LEAF(pRam);
918	}
919	}
920	else
921	{
922
923	/*
924	* Intermediate node.
925	*/
926	PPGMRAMRANGE pLeft = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
927
928	pRam = *ppRam;
929	if (!pRam)
930	return pLeft;
931	*ppRam = pRam->pNextR3;
932	MAKE_LEAF(pRam);
933	INSERT_LEFT(pRam, pLeft);
934
935	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
936	if (pRight)
937	INSERT_RIGHT(pRam, pRight);
938	}
939	return pRam;
940	}
941
942
943	/**
944	* Rebuilds the RAM range search trees.
945	*
946	* @param pVM The cross context VM structure.
947	*/
948	static void pgmR3PhysRebuildRamRangeSearchTrees(PVM pVM)
949	{
950
951	/*
952	* Create the reasonably balanced tree in a sequential fashion.
953	* For simplicity (laziness) we use standard recursion here.
954	*/
955	int iDepth = 0;
956	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
957	PPGMRAMRANGE pRoot = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, 0);
958	while (pRam)
959	{
960	PPGMRAMRANGE pLeft = pRoot;
961
962	pRoot = pRam;
963	pRam = pRam->pNextR3;
964	MAKE_LEAF(pRoot);
965	INSERT_LEFT(pRoot, pLeft);
966
967	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, iDepth);
968	if (pRight)
969	INSERT_RIGHT(pRoot, pRight);
970	/** @todo else: rotate the tree. */
971
972	iDepth++;
973	}
974
975	pVM->pgm.s.pRamRangeTreeR3 = pRoot;
976	pVM->pgm.s.pRamRangeTreeR0 = pRoot ? pRoot->pSelfR0 : NIL_RTR0PTR;
977
978	#ifdef VBOX_STRICT
979	/*
980	* Verify that the above code works.
981	*/
982	unsigned cRanges = 0;
983	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
984	cRanges++;
985	Assert(cRanges > 0);
986
987	unsigned cMaxDepth = ASMBitLastSetU32(cRanges);
988	if ((1U << cMaxDepth) < cRanges)
989	cMaxDepth++;
990
991	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
992	{
993	unsigned cDepth = 0;
994	PPGMRAMRANGE pRam2 = pVM->pgm.s.pRamRangeTreeR3;
995	for (;;)
996	{
997	if (pRam == pRam2)
998	break;
999	Assert(pRam2);
1000	if (pRam->GCPhys < pRam2->GCPhys)
1001	pRam2 = pRam2->pLeftR3;
1002	else
1003	pRam2 = pRam2->pRightR3;
1004	}
1005	AssertMsg(cDepth <= cMaxDepth, ("cDepth=%d cMaxDepth=%d\n", cDepth, cMaxDepth));
1006	}
1007	#endif /* VBOX_STRICT */
1008	}
1009
1010	#undef MAKE_LEAF
1011	#undef INSERT_LEFT
1012	#undef INSERT_RIGHT
1013
1014	/**
1015	* Relinks the RAM ranges using the pSelfRC and pSelfR0 pointers.
1016	*
1017	* Called when anything was relocated.
1018	*
1019	* @param pVM The cross context VM structure.
1020	*/
1021	void pgmR3PhysRelinkRamRanges(PVM pVM)
1022	{
1023	PPGMRAMRANGE pCur;
1024
1025	#ifdef VBOX_STRICT
1026	for (pCur = pVM->pgm.s.pRamRangesXR3; pCur; pCur = pCur->pNextR3)
1027	{
1028	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfR0 == MMHyperCCToR0(pVM, pCur));
1029	Assert((pCur->GCPhys & PAGE_OFFSET_MASK) == 0);
1030	Assert((pCur->GCPhysLast & PAGE_OFFSET_MASK) == PAGE_OFFSET_MASK);
1031	Assert((pCur->cb & PAGE_OFFSET_MASK) == 0);
1032	Assert(pCur->cb == pCur->GCPhysLast - pCur->GCPhys + 1);
1033	for (PPGMRAMRANGE pCur2 = pVM->pgm.s.pRamRangesXR3; pCur2; pCur2 = pCur2->pNextR3)
1034	Assert( pCur2 == pCur
1035	\|\| strcmp(pCur2->pszDesc, pCur->pszDesc)); /** @todo fix MMIO ranges!! */
1036	}
1037	#endif
1038
1039	pCur = pVM->pgm.s.pRamRangesXR3;
1040	if (pCur)
1041	{
1042	pVM->pgm.s.pRamRangesXR0 = pCur->pSelfR0;
1043
1044	for (; pCur->pNextR3; pCur = pCur->pNextR3)
1045	pCur->pNextR0 = pCur->pNextR3->pSelfR0;
1046
1047	Assert(pCur->pNextR0 == NIL_RTR0PTR);
1048	}
1049	else
1050	{
1051	Assert(pVM->pgm.s.pRamRangesXR0 == NIL_RTR0PTR);
1052	}
1053	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1054
1055	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1056	}
1057
1058
1059	/**
1060	* Links a new RAM range into the list.
1061	*
1062	* @param pVM The cross context VM structure.
1063	* @param pNew Pointer to the new list entry.
1064	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
1065	*/
1066	static void pgmR3PhysLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, PPGMRAMRANGE pPrev)
1067	{
1068	AssertMsg(pNew->pszDesc, ("%RGp-%RGp\n", pNew->GCPhys, pNew->GCPhysLast));
1069	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfR0 == MMHyperCCToR0(pVM, pNew));
1070
1071	PGM_LOCK_VOID(pVM);
1072
1073	PPGMRAMRANGE pRam = pPrev ? pPrev->pNextR3 : pVM->pgm.s.pRamRangesXR3;
1074	pNew->pNextR3 = pRam;
1075	pNew->pNextR0 = pRam ? pRam->pSelfR0 : NIL_RTR0PTR;
1076
1077	if (pPrev)
1078	{
1079	pPrev->pNextR3 = pNew;
1080	pPrev->pNextR0 = pNew->pSelfR0;
1081	}
1082	else
1083	{
1084	pVM->pgm.s.pRamRangesXR3 = pNew;
1085	pVM->pgm.s.pRamRangesXR0 = pNew->pSelfR0;
1086	}
1087	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1088
1089	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1090	PGM_UNLOCK(pVM);
1091	}
1092
1093
1094	/**
1095	* Unlink an existing RAM range from the list.
1096	*
1097	* @param pVM The cross context VM structure.
1098	* @param pRam Pointer to the new list entry.
1099	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
1100	*/
1101	static void pgmR3PhysUnlinkRamRange2(PVM pVM, PPGMRAMRANGE pRam, PPGMRAMRANGE pPrev)
1102	{
1103	Assert(pPrev ? pPrev->pNextR3 == pRam : pVM->pgm.s.pRamRangesXR3 == pRam);
1104	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfR0 == MMHyperCCToR0(pVM, pRam));
1105
1106	PGM_LOCK_VOID(pVM);
1107
1108	PPGMRAMRANGE pNext = pRam->pNextR3;
1109	if (pPrev)
1110	{
1111	pPrev->pNextR3 = pNext;
1112	pPrev->pNextR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
1113	}
1114	else
1115	{
1116	Assert(pVM->pgm.s.pRamRangesXR3 == pRam);
1117	pVM->pgm.s.pRamRangesXR3 = pNext;
1118	pVM->pgm.s.pRamRangesXR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
1119	}
1120	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1121
1122	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1123	PGM_UNLOCK(pVM);
1124	}
1125
1126
1127	/**
1128	* Unlink an existing RAM range from the list.
1129	*
1130	* @param pVM The cross context VM structure.
1131	* @param pRam Pointer to the new list entry.
1132	*/
1133	static void pgmR3PhysUnlinkRamRange(PVM pVM, PPGMRAMRANGE pRam)
1134	{
1135	PGM_LOCK_VOID(pVM);
1136
1137	/* find prev. */
1138	PPGMRAMRANGE pPrev = NULL;
1139	PPGMRAMRANGE pCur = pVM->pgm.s.pRamRangesXR3;
1140	while (pCur != pRam)
1141	{
1142	pPrev = pCur;
1143	pCur = pCur->pNextR3;
1144	}
1145	AssertFatal(pCur);
1146
1147	pgmR3PhysUnlinkRamRange2(pVM, pRam, pPrev);
1148	PGM_UNLOCK(pVM);
1149	}
1150
1151
1152	/**
1153	* Frees a range of pages, replacing them with ZERO pages of the specified type.
1154	*
1155	* @returns VBox status code.
1156	* @param pVM The cross context VM structure.
1157	* @param pRam The RAM range in which the pages resides.
1158	* @param GCPhys The address of the first page.
1159	* @param GCPhysLast The address of the last page.
1160	* @param enmType The page type to replace then with.
1161	*/
1162	static int pgmR3PhysFreePageRange(PVM pVM, PPGMRAMRANGE pRam, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast, PGMPAGETYPE enmType)
1163	{
1164	PGM_LOCK_ASSERT_OWNER(pVM);
1165	uint32_t cPendingPages = 0;
1166	PGMMFREEPAGESREQ pReq;
1167	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1168	AssertLogRelRCReturn(rc, rc);
1169
1170	/* Iterate the pages. */
1171	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
1172	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
1173	while (cPagesLeft-- > 0)
1174	{
1175	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys, enmType);
1176	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
1177
1178	PGM_PAGE_SET_TYPE(pVM, pPageDst, enmType);
1179
1180	GCPhys += PAGE_SIZE;
1181	pPageDst++;
1182	}
1183
1184	if (cPendingPages)
1185	{
1186	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1187	AssertLogRelRCReturn(rc, rc);
1188	}
1189	GMMR3FreePagesCleanup(pReq);
1190
1191	return rc;
1192	}
1193
1194	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
1195
1196	/**
1197	* Rendezvous callback used by PGMR3ChangeMemBalloon that changes the memory balloon size
1198	*
1199	* This is only called on one of the EMTs while the other ones are waiting for
1200	* it to complete this function.
1201	*
1202	* @returns VINF_SUCCESS (VBox strict status code).
1203	* @param pVM The cross context VM structure.
1204	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
1205	* @param pvUser User parameter
1206	*/
1207	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysChangeMemBalloonRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
1208	{
1209	uintptr_t paUser = (uintptr_t )pvUser;
1210	bool fInflate = !!paUser[0];
1211	unsigned cPages = paUser[1];
1212	RTGCPHYS paPhysPage = (RTGCPHYS )paUser[2];
1213	uint32_t cPendingPages = 0;
1214	PGMMFREEPAGESREQ pReq;
1215	int rc;
1216
1217	Log(("pgmR3PhysChangeMemBalloonRendezvous: %s %x pages\n", (fInflate) ? "inflate" : "deflate", cPages));
1218	PGM_LOCK_VOID(pVM);
1219
1220	if (fInflate)
1221	{
1222	/* Flush the PGM pool cache as we might have stale references to pages that we just freed. */
1223	pgmR3PoolClearAllRendezvous(pVM, pVCpu, NULL);
1224
1225	/* Replace pages with ZERO pages. */
1226	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1227	if (RT_FAILURE(rc))
1228	{
1229	PGM_UNLOCK(pVM);
1230	AssertLogRelRC(rc);
1231	return rc;
1232	}
1233
1234	/* Iterate the pages. */
1235	for (unsigned i = 0; i < cPages; i++)
1236	{
1237	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
1238	if ( pPage == NULL
1239	\|\| PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM)
1240	{
1241	Log(("pgmR3PhysChangeMemBalloonRendezvous: invalid physical page %RGp pPage->u3Type=%d\n", paPhysPage[i], pPage ? PGM_PAGE_GET_TYPE(pPage) : 0));
1242	break;
1243	}
1244
1245	LogFlow(("balloon page: %RGp\n", paPhysPage[i]));
1246
1247	/* Flush the shadow PT if this page was previously used as a guest page table. */
1248	pgmPoolFlushPageByGCPhys(pVM, paPhysPage[i]);
1249
1250	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, paPhysPage[i], (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage));
1251	if (RT_FAILURE(rc))
1252	{
1253	PGM_UNLOCK(pVM);
1254	AssertLogRelRC(rc);
1255	return rc;
1256	}
1257	Assert(PGM_PAGE_IS_ZERO(pPage));
1258	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_BALLOONED);
1259	}
1260
1261	if (cPendingPages)
1262	{
1263	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1264	if (RT_FAILURE(rc))
1265	{
1266	PGM_UNLOCK(pVM);
1267	AssertLogRelRC(rc);
1268	return rc;
1269	}
1270	}
1271	GMMR3FreePagesCleanup(pReq);
1272	}
1273	else
1274	{
1275	/* Iterate the pages. */
1276	for (unsigned i = 0; i < cPages; i++)
1277	{
1278	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
1279	AssertBreak(pPage && PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM);
1280
1281	LogFlow(("Free ballooned page: %RGp\n", paPhysPage[i]));
1282
1283	Assert(PGM_PAGE_IS_BALLOONED(pPage));
1284
1285	/* Change back to zero page. (NEM does not need to be informed.) */
1286	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
1287	}
1288
1289	/* Note that we currently do not map any ballooned pages in our shadow page tables, so no need to flush the pgm pool. */
1290	}
1291
1292	/* Notify GMM about the balloon change. */
1293	rc = GMMR3BalloonedPages(pVM, (fInflate) ? GMMBALLOONACTION_INFLATE : GMMBALLOONACTION_DEFLATE, cPages);
1294	if (RT_SUCCESS(rc))
1295	{
1296	if (!fInflate)
1297	{
1298	Assert(pVM->pgm.s.cBalloonedPages >= cPages);
1299	pVM->pgm.s.cBalloonedPages -= cPages;
1300	}
1301	else
1302	pVM->pgm.s.cBalloonedPages += cPages;
1303	}
1304
1305	PGM_UNLOCK(pVM);
1306
1307	/* Flush the recompiler's TLB as well. */
1308	for (VMCPUID i = 0; i < pVM->cCpus; i++)
1309	CPUMSetChangedFlags(pVM->apCpusR3[i], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1310
1311	AssertLogRelRC(rc);
1312	return rc;
1313	}
1314
1315
1316	/**
1317	* Frees a range of ram pages, replacing them with ZERO pages; helper for PGMR3PhysFreeRamPages
1318	*
1319	* @returns VBox status code.
1320	* @param pVM The cross context VM structure.
1321	* @param fInflate Inflate or deflate memory balloon
1322	* @param cPages Number of pages to free
1323	* @param paPhysPage Array of guest physical addresses
1324	*/
1325	static DECLCALLBACK(void) pgmR3PhysChangeMemBalloonHelper(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1326	{
1327	uintptr_t paUser[3];
1328
1329	paUser[0] = fInflate;
1330	paUser[1] = cPages;
1331	paUser[2] = (uintptr_t)paPhysPage;
1332	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1333	AssertRC(rc);
1334
1335	/* Made a copy in PGMR3PhysFreeRamPages; free it here. */
1336	RTMemFree(paPhysPage);
1337	}
1338
1339	#endif /* 64-bit host && (Windows \|\| Solaris \|\| Linux \|\| FreeBSD) */
1340
1341	/**
1342	* Inflate or deflate a memory balloon
1343	*
1344	* @returns VBox status code.
1345	* @param pVM The cross context VM structure.
1346	* @param fInflate Inflate or deflate memory balloon
1347	* @param cPages Number of pages to free
1348	* @param paPhysPage Array of guest physical addresses
1349	*/
1350	VMMR3DECL(int) PGMR3PhysChangeMemBalloon(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1351	{
1352	/* This must match GMMR0Init; currently we only support memory ballooning on all 64-bit hosts except Mac OS X */
1353	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
1354	int rc;
1355
1356	/* Older additions (ancient non-functioning balloon code) pass wrong physical addresses. */
1357	AssertReturn(!(paPhysPage[0] & 0xfff), VERR_INVALID_PARAMETER);
1358
1359	/* We own the IOM lock here and could cause a deadlock by waiting for another VCPU that is blocking on the IOM lock.
1360	* In the SMP case we post a request packet to postpone the job.
1361	*/
1362	if (pVM->cCpus > 1)
1363	{
1364	unsigned cbPhysPage = cPages * sizeof(paPhysPage[0]);
1365	RTGCPHYS paPhysPageCopy = (RTGCPHYS )RTMemAlloc(cbPhysPage);
1366	AssertReturn(paPhysPageCopy, VERR_NO_MEMORY);
1367
1368	memcpy(paPhysPageCopy, paPhysPage, cbPhysPage);
1369
1370	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysChangeMemBalloonHelper, 4, pVM, fInflate, cPages, paPhysPageCopy);
1371	AssertRC(rc);
1372	}
1373	else
1374	{
1375	uintptr_t paUser[3];
1376
1377	paUser[0] = fInflate;
1378	paUser[1] = cPages;
1379	paUser[2] = (uintptr_t)paPhysPage;
1380	rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1381	AssertRC(rc);
1382	}
1383	return rc;
1384
1385	#else
1386	NOREF(pVM); NOREF(fInflate); NOREF(cPages); NOREF(paPhysPage);
1387	return VERR_NOT_IMPLEMENTED;
1388	#endif
1389	}
1390
1391
1392	/**
1393	* Rendezvous callback used by PGMR3WriteProtectRAM that write protects all
1394	* physical RAM.
1395	*
1396	* This is only called on one of the EMTs while the other ones are waiting for
1397	* it to complete this function.
1398	*
1399	* @returns VINF_SUCCESS (VBox strict status code).
1400	* @param pVM The cross context VM structure.
1401	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
1402	* @param pvUser User parameter, unused.
1403	*/
1404	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysWriteProtectRAMRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
1405	{
1406	int rc = VINF_SUCCESS;
1407	NOREF(pvUser); NOREF(pVCpu);
1408
1409	PGM_LOCK_VOID(pVM);
1410	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
1411	pgmPoolResetDirtyPages(pVM);
1412	#endif
1413
1414	/** @todo pointless to write protect the physical page pointed to by RSP. */
1415
1416	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
1417	pRam;
1418	pRam = pRam->CTX_SUFF(pNext))
1419	{
1420	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
1421	for (uint32_t iPage = 0; iPage < cPages; iPage++)
1422	{
1423	PPGMPAGE pPage = &pRam->aPages[iPage];
1424	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
1425
1426	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
1427	\|\| enmPageType == PGMPAGETYPE_MMIO2)
1428	{
1429	/*
1430	* A RAM page.
1431	*/
1432	switch (PGM_PAGE_GET_STATE(pPage))
1433	{
1434	case PGM_PAGE_STATE_ALLOCATED:
1435	/** @todo Optimize this: Don't always re-enable write
1436	* monitoring if the page is known to be very busy. */
1437	if (PGM_PAGE_IS_WRITTEN_TO(pPage))
1438	PGM_PAGE_CLEAR_WRITTEN_TO(pVM, pPage);
1439
1440	pgmPhysPageWriteMonitor(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
1441	break;
1442
1443	case PGM_PAGE_STATE_SHARED:
1444	AssertFailed();
1445	break;
1446
1447	case PGM_PAGE_STATE_WRITE_MONITORED: /* nothing to change. */
1448	default:
1449	break;
1450	}
1451	}
1452	}
1453	}
1454	pgmR3PoolWriteProtectPages(pVM);
1455	PGM_INVL_ALL_VCPU_TLBS(pVM);
1456	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
1457	CPUMSetChangedFlags(pVM->apCpusR3[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1458
1459	PGM_UNLOCK(pVM);
1460	return rc;
1461	}
1462
1463	/**
1464	* Protect all physical RAM to monitor writes
1465	*
1466	* @returns VBox status code.
1467	* @param pVM The cross context VM structure.
1468	*/
1469	VMMR3DECL(int) PGMR3PhysWriteProtectRAM(PVM pVM)
1470	{
1471	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1472
1473	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysWriteProtectRAMRendezvous, NULL);
1474	AssertRC(rc);
1475	return rc;
1476	}
1477
1478
1479	/**
1480	* Gets the number of ram ranges.
1481	*
1482	* @returns Number of ram ranges. Returns UINT32_MAX if @a pVM is invalid.
1483	* @param pVM The cross context VM structure.
1484	*/
1485	VMMR3DECL(uint32_t) PGMR3PhysGetRamRangeCount(PVM pVM)
1486	{
1487	VM_ASSERT_VALID_EXT_RETURN(pVM, UINT32_MAX);
1488
1489	PGM_LOCK_VOID(pVM);
1490	uint32_t cRamRanges = 0;
1491	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext))
1492	cRamRanges++;
1493	PGM_UNLOCK(pVM);
1494	return cRamRanges;
1495	}
1496
1497
1498	/**
1499	* Get information about a range.
1500	*
1501	* @returns VINF_SUCCESS or VERR_OUT_OF_RANGE.
1502	* @param pVM The cross context VM structure.
1503	* @param iRange The ordinal of the range.
1504	* @param pGCPhysStart Where to return the start of the range. Optional.
1505	* @param pGCPhysLast Where to return the address of the last byte in the
1506	* range. Optional.
1507	* @param ppszDesc Where to return the range description. Optional.
1508	* @param pfIsMmio Where to indicate that this is a pure MMIO range.
1509	* Optional.
1510	*/
1511	VMMR3DECL(int) PGMR3PhysGetRange(PVM pVM, uint32_t iRange, PRTGCPHYS pGCPhysStart, PRTGCPHYS pGCPhysLast,
1512	const char *ppszDesc, bool pfIsMmio)
1513	{
1514	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1515
1516	PGM_LOCK_VOID(pVM);
1517	uint32_t iCurRange = 0;
1518	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext), iCurRange++)
1519	if (iCurRange == iRange)
1520	{
1521	if (pGCPhysStart)
1522	*pGCPhysStart = pCur->GCPhys;
1523	if (pGCPhysLast)
1524	*pGCPhysLast = pCur->GCPhysLast;
1525	if (ppszDesc)
1526	*ppszDesc = pCur->pszDesc;
1527	if (pfIsMmio)
1528	*pfIsMmio = !!(pCur->fFlags & PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO);
1529
1530	PGM_UNLOCK(pVM);
1531	return VINF_SUCCESS;
1532	}
1533	PGM_UNLOCK(pVM);
1534	return VERR_OUT_OF_RANGE;
1535	}
1536
1537
1538	/**
1539	* Query the amount of free memory inside VMMR0
1540	*
1541	* @returns VBox status code.
1542	* @param pUVM The user mode VM handle.
1543	* @param pcbAllocMem Where to return the amount of memory allocated
1544	* by VMs.
1545	* @param pcbFreeMem Where to return the amount of memory that is
1546	* allocated from the host but not currently used
1547	* by any VMs.
1548	* @param pcbBallonedMem Where to return the sum of memory that is
1549	* currently ballooned by the VMs.
1550	* @param pcbSharedMem Where to return the amount of memory that is
1551	* currently shared.
1552	*/
1553	VMMR3DECL(int) PGMR3QueryGlobalMemoryStats(PUVM pUVM, uint64_t pcbAllocMem, uint64_t pcbFreeMem,
1554	uint64_t pcbBallonedMem, uint64_t pcbSharedMem)
1555	{
1556	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1557	VM_ASSERT_VALID_EXT_RETURN(pUVM->pVM, VERR_INVALID_VM_HANDLE);
1558
1559	uint64_t cAllocPages = 0;
1560	uint64_t cFreePages = 0;
1561	uint64_t cBalloonPages = 0;
1562	uint64_t cSharedPages = 0;
1563	int rc = GMMR3QueryHypervisorMemoryStats(pUVM->pVM, &cAllocPages, &cFreePages, &cBalloonPages, &cSharedPages);
1564	AssertRCReturn(rc, rc);
1565
1566	if (pcbAllocMem)
1567	pcbAllocMem = cAllocPages _4K;
1568
1569	if (pcbFreeMem)
1570	pcbFreeMem = cFreePages _4K;
1571
1572	if (pcbBallonedMem)
1573	pcbBallonedMem = cBalloonPages _4K;
1574
1575	if (pcbSharedMem)
1576	pcbSharedMem = cSharedPages _4K;
1577
1578	Log(("PGMR3QueryVMMMemoryStats: all=%llx free=%llx ballooned=%llx shared=%llx\n",
1579	cAllocPages, cFreePages, cBalloonPages, cSharedPages));
1580	return VINF_SUCCESS;
1581	}
1582
1583
1584	/**
1585	* Query memory stats for the VM.
1586	*
1587	* @returns VBox status code.
1588	* @param pUVM The user mode VM handle.
1589	* @param pcbTotalMem Where to return total amount memory the VM may
1590	* possibly use.
1591	* @param pcbPrivateMem Where to return the amount of private memory
1592	* currently allocated.
1593	* @param pcbSharedMem Where to return the amount of actually shared
1594	* memory currently used by the VM.
1595	* @param pcbZeroMem Where to return the amount of memory backed by
1596	* zero pages.
1597	*
1598	* @remarks The total mem is normally larger than the sum of the three
1599	* components. There are two reasons for this, first the amount of
1600	* shared memory is what we're sure is shared instead of what could
1601	* possibly be shared with someone. Secondly, because the total may
1602	* include some pure MMIO pages that doesn't go into any of the three
1603	* sub-counts.
1604	*
1605	* @todo Why do we return reused shared pages instead of anything that could
1606	* potentially be shared? Doesn't this mean the first VM gets a much
1607	* lower number of shared pages?
1608	*/
1609	VMMR3DECL(int) PGMR3QueryMemoryStats(PUVM pUVM, uint64_t pcbTotalMem, uint64_t pcbPrivateMem,
1610	uint64_t pcbSharedMem, uint64_t pcbZeroMem)
1611	{
1612	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1613	PVM pVM = pUVM->pVM;
1614	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1615
1616	if (pcbTotalMem)
1617	pcbTotalMem = (uint64_t)pVM->pgm.s.cAllPages PAGE_SIZE;
1618
1619	if (pcbPrivateMem)
1620	pcbPrivateMem = (uint64_t)pVM->pgm.s.cPrivatePages PAGE_SIZE;
1621
1622	if (pcbSharedMem)
1623	pcbSharedMem = (uint64_t)pVM->pgm.s.cReusedSharedPages PAGE_SIZE;
1624
1625	if (pcbZeroMem)
1626	pcbZeroMem = (uint64_t)pVM->pgm.s.cZeroPages PAGE_SIZE;
1627
1628	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));
1629	return VINF_SUCCESS;
1630	}
1631
1632
1633	/**
1634	* PGMR3PhysRegisterRam worker that initializes and links a RAM range.
1635	*
1636	* @param pVM The cross context VM structure.
1637	* @param pNew The new RAM range.
1638	* @param GCPhys The address of the RAM range.
1639	* @param GCPhysLast The last address of the RAM range.
1640	* @param RCPtrNew The RC address if the range is floating. NIL_RTRCPTR
1641	* if in HMA.
1642	* @param R0PtrNew Ditto for R0.
1643	* @param pszDesc The description.
1644	* @param pPrev The previous RAM range (for linking).
1645	*/
1646	static void pgmR3PhysInitAndLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast,
1647	RTRCPTR RCPtrNew, RTR0PTR R0PtrNew, const char *pszDesc, PPGMRAMRANGE pPrev)
1648	{
1649	/*
1650	* Initialize the range.
1651	*/
1652	pNew->pSelfR0 = R0PtrNew != NIL_RTR0PTR ? R0PtrNew : MMHyperCCToR0(pVM, pNew);
1653	pNew->GCPhys = GCPhys;
1654	pNew->GCPhysLast = GCPhysLast;
1655	pNew->cb = GCPhysLast - GCPhys + 1;
1656	pNew->pszDesc = pszDesc;
1657	pNew->fFlags = RCPtrNew != NIL_RTRCPTR ? PGM_RAM_RANGE_FLAGS_FLOATING : 0;
1658	pNew->pvR3 = NULL;
1659	pNew->paLSPages = NULL;
1660
1661	uint32_t const cPages = pNew->cb >> PAGE_SHIFT;
1662	RTGCPHYS iPage = cPages;
1663	while (iPage-- > 0)
1664	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_RAM);
1665
1666	/* Update the page count stats. */
1667	pVM->pgm.s.cZeroPages += cPages;
1668	pVM->pgm.s.cAllPages += cPages;
1669
1670	/*
1671	* Link it.
1672	*/
1673	pgmR3PhysLinkRamRange(pVM, pNew, pPrev);
1674	}
1675
1676
1677	#ifndef PGM_WITHOUT_MAPPINGS
1678	/**
1679	* @callback_method_impl{FNPGMRELOCATE, Relocate a floating RAM range.}
1680	* @sa pgmR3PhysMMIO2ExRangeRelocate
1681	*/
1682	static DECLCALLBACK(bool) pgmR3PhysRamRangeRelocate(PVM pVM, RTGCPTR GCPtrOld, RTGCPTR GCPtrNew,
1683	PGMRELOCATECALL enmMode, void *pvUser)
1684	{
1685	PPGMRAMRANGE pRam = (PPGMRAMRANGE)pvUser;
1686	Assert(pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING);
1687	Assert(pRam->pSelfRC == GCPtrOld + PAGE_SIZE); RT_NOREF_PV(GCPtrOld);
1688
1689	switch (enmMode)
1690	{
1691	case PGMRELOCATECALL_SUGGEST:
1692	return true;
1693
1694	case PGMRELOCATECALL_RELOCATE:
1695	{
1696	/*
1697	* Update myself, then relink all the ranges and flush the RC TLB.
1698	*/
1699	PGM_LOCK_VOID(pVM);
1700
1701	pRam->pSelfRC = (RTRCPTR)(GCPtrNew + PAGE_SIZE);
1702
1703	pgmR3PhysRelinkRamRanges(pVM);
1704	for (unsigned i = 0; i < PGM_RAMRANGE_TLB_ENTRIES; i++)
1705	pVM->pgm.s.apRamRangesTlbRC[i] = NIL_RTRCPTR;
1706
1707	PGM_UNLOCK(pVM);
1708	return true;
1709	}
1710
1711	default:
1712	AssertFailedReturn(false);
1713	}
1714	}
1715	#endif /* !PGM_WITHOUT_MAPPINGS */
1716
1717
1718	/**
1719	* PGMR3PhysRegisterRam worker that registers a high chunk.
1720	*
1721	* @returns VBox status code.
1722	* @param pVM The cross context VM structure.
1723	* @param GCPhys The address of the RAM.
1724	* @param cRamPages The number of RAM pages to register.
1725	* @param cbChunk The size of the PGMRAMRANGE guest mapping.
1726	* @param iChunk The chunk number.
1727	* @param pszDesc The RAM range description.
1728	* @param ppPrev Previous RAM range pointer. In/Out.
1729	*/
1730	static int pgmR3PhysRegisterHighRamChunk(PVM pVM, RTGCPHYS GCPhys, uint32_t cRamPages,
1731	uint32_t cbChunk, uint32_t iChunk, const char *pszDesc,
1732	PPGMRAMRANGE *ppPrev)
1733	{
1734	const char *pszDescChunk = iChunk == 0
1735	? pszDesc
1736	: MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s (#%u)", pszDesc, iChunk + 1);
1737	AssertReturn(pszDescChunk, VERR_NO_MEMORY);
1738
1739	/*
1740	* Allocate memory for the new chunk.
1741	*/
1742	size_t const cChunkPages = RT_ALIGN_Z(RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cRamPages]), PAGE_SIZE) >> PAGE_SHIFT;
1743	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
1744	AssertReturn(paChunkPages, VERR_NO_TMP_MEMORY);
1745	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
1746	void *pvChunk = NULL;
1747	int rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk, &R0PtrChunk, paChunkPages);
1748	if (RT_SUCCESS(rc))
1749	{
1750	Assert(R0PtrChunk != NIL_RTR0PTR);
1751	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
1752
1753	PPGMRAMRANGE pNew = (PPGMRAMRANGE)pvChunk;
1754
1755	/*
1756	* Create a mapping and map the pages into it.
1757	* We push these in below the HMA.
1758	*/
1759	RTGCPTR GCPtrChunkMap = pVM->pgm.s.GCPtrPrevRamRangeMapping - cbChunk;
1760	#ifndef PGM_WITHOUT_MAPPINGS
1761	rc = PGMR3MapPT(pVM, GCPtrChunkMap, cbChunk, 0 /fFlags/, pgmR3PhysRamRangeRelocate, pNew, pszDescChunk);
1762	if (RT_SUCCESS(rc))
1763	#endif /* !PGM_WITHOUT_MAPPINGS */
1764	{
1765	pVM->pgm.s.GCPtrPrevRamRangeMapping = GCPtrChunkMap;
1766
1767	RTGCPTR const GCPtrChunk = GCPtrChunkMap + PAGE_SIZE;
1768	#ifndef PGM_WITHOUT_MAPPINGS
1769	RTGCPTR GCPtrPage = GCPtrChunk;
1770	for (uint32_t iPage = 0; iPage < cChunkPages && RT_SUCCESS(rc); iPage++, GCPtrPage += PAGE_SIZE)
1771	rc = PGMMap(pVM, GCPtrPage, paChunkPages[iPage].Phys, PAGE_SIZE, 0);
1772	if (RT_SUCCESS(rc))
1773	#endif /* !PGM_WITHOUT_MAPPINGS */
1774	{
1775	/*
1776	* Ok, init and link the range.
1777	*/
1778	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhys + ((RTGCPHYS)cRamPages << PAGE_SHIFT) - 1,
1779	(RTRCPTR)GCPtrChunk, R0PtrChunk, pszDescChunk, *ppPrev);
1780	*ppPrev = pNew;
1781	}
1782	}
1783
1784	if (RT_FAILURE(rc))
1785	SUPR3PageFreeEx(pvChunk, cChunkPages);
1786	}
1787
1788	RTMemTmpFree(paChunkPages);
1789	return rc;
1790	}
1791
1792
1793	/**
1794	* Sets up a range RAM.
1795	*
1796	* This will check for conflicting registrations, make a resource
1797	* reservation for the memory (with GMM), and setup the per-page
1798	* tracking structures (PGMPAGE).
1799	*
1800	* @returns VBox status code.
1801	* @param pVM The cross context VM structure.
1802	* @param GCPhys The physical address of the RAM.
1803	* @param cb The size of the RAM.
1804	* @param pszDesc The description - not copied, so, don't free or change it.
1805	*/
1806	VMMR3DECL(int) PGMR3PhysRegisterRam(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, const char *pszDesc)
1807	{
1808	/*
1809	* Validate input.
1810	*/
1811	Log(("PGMR3PhysRegisterRam: GCPhys=%RGp cb=%RGp pszDesc=%s\n", GCPhys, cb, pszDesc));
1812	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
1813	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
1814	AssertReturn(cb > 0, VERR_INVALID_PARAMETER);
1815	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
1816	AssertMsgReturn(GCPhysLast > GCPhys, ("The range wraps! GCPhys=%RGp cb=%RGp\n", GCPhys, cb), VERR_INVALID_PARAMETER);
1817	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
1818	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1819
1820	PGM_LOCK_VOID(pVM);
1821
1822	/*
1823	* Find range location and check for conflicts.
1824	*/
1825	PPGMRAMRANGE pPrev = NULL;
1826	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
1827	while (pRam && GCPhysLast >= pRam->GCPhys)
1828	{
1829	AssertLogRelMsgReturnStmt( GCPhysLast < pRam->GCPhys
1830	\|\| GCPhys > pRam->GCPhysLast,
1831	("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
1832	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
1833	PGM_UNLOCK(pVM), VERR_PGM_RAM_CONFLICT);
1834
1835	/* next */
1836	pPrev = pRam;
1837	pRam = pRam->pNextR3;
1838	}
1839
1840	/*
1841	* Register it with GMM (the API bitches).
1842	*/
1843	const RTGCPHYS cPages = cb >> PAGE_SHIFT;
1844	int rc = MMR3IncreaseBaseReservation(pVM, cPages);
1845	if (RT_FAILURE(rc))
1846	{
1847	PGM_UNLOCK(pVM);
1848	return rc;
1849	}
1850
1851	if ( GCPhys >= _4G
1852	&& cPages > 256)
1853	{
1854	/*
1855	* The PGMRAMRANGE structures for the high memory can get very big.
1856	* In order to avoid SUPR3PageAllocEx allocation failures due to the
1857	* allocation size limit there and also to avoid being unable to find
1858	* guest mapping space for them, we split this memory up into 4MB in
1859	* (potential) raw-mode configs and 16MB chunks in forced AMD-V/VT-x
1860	* mode.
1861	*
1862	* The first and last page of each mapping are guard pages and marked
1863	* not-present. So, we've got 4186112 and 16769024 bytes available for
1864	* the PGMRAMRANGE structure.
1865	*
1866	* Note! The sizes used here will influence the saved state.
1867	*/
1868	uint32_t cbChunk = 16U*_1M;
1869	uint32_t cPagesPerChunk = 1048048; /* max ~1048059 */
1870	AssertCompile(sizeof(PGMRAMRANGE) + sizeof(PGMPAGE) * 1048048 < 16U_1M - PAGE_SIZE 2);
1871	AssertRelease(RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPagesPerChunk]) + PAGE_SIZE * 2 <= cbChunk);
1872
1873	RTGCPHYS cPagesLeft = cPages;
1874	RTGCPHYS GCPhysChunk = GCPhys;
1875	uint32_t iChunk = 0;
1876	while (cPagesLeft > 0)
1877	{
1878	uint32_t cPagesInChunk = cPagesLeft;
1879	if (cPagesInChunk > cPagesPerChunk)
1880	cPagesInChunk = cPagesPerChunk;
1881
1882	rc = pgmR3PhysRegisterHighRamChunk(pVM, GCPhysChunk, cPagesInChunk, cbChunk, iChunk, pszDesc, &pPrev);
1883	AssertRCReturn(rc, rc);
1884
1885	/* advance */
1886	GCPhysChunk += (RTGCPHYS)cPagesInChunk << PAGE_SHIFT;
1887	cPagesLeft -= cPagesInChunk;
1888	iChunk++;
1889	}
1890	}
1891	else
1892	{
1893	/*
1894	* Allocate, initialize and link the new RAM range.
1895	*/
1896	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
1897	PPGMRAMRANGE pNew;
1898	rc = MMR3HyperAllocOnceNoRel(pVM, cbRamRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
1899	AssertLogRelMsgRCReturn(rc, ("cbRamRange=%zu\n", cbRamRange), rc);
1900
1901	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhysLast, NIL_RTRCPTR, NIL_RTR0PTR, pszDesc, pPrev);
1902	}
1903	pgmPhysInvalidatePageMapTLB(pVM);
1904
1905	/*
1906	* Notify NEM while holding the lock (experimental) and REM without (like always).
1907	*/
1908	rc = NEMR3NotifyPhysRamRegister(pVM, GCPhys, cb);
1909	PGM_UNLOCK(pVM);
1910	return rc;
1911	}
1912
1913
1914	/**
1915	* Worker called by PGMR3InitFinalize if we're configured to pre-allocate RAM.
1916	*
1917	* We do this late in the init process so that all the ROM and MMIO ranges have
1918	* been registered already and we don't go wasting memory on them.
1919	*
1920	* @returns VBox status code.
1921	*
1922	* @param pVM The cross context VM structure.
1923	*/
1924	int pgmR3PhysRamPreAllocate(PVM pVM)
1925	{
1926	Assert(pVM->pgm.s.fRamPreAlloc);
1927	Log(("pgmR3PhysRamPreAllocate: enter\n"));
1928
1929	/*
1930	* Walk the RAM ranges and allocate all RAM pages, halt at
1931	* the first allocation error.
1932	*/
1933	uint64_t cPages = 0;
1934	uint64_t NanoTS = RTTimeNanoTS();
1935	PGM_LOCK_VOID(pVM);
1936	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1937	{
1938	PPGMPAGE pPage = &pRam->aPages[0];
1939	RTGCPHYS GCPhys = pRam->GCPhys;
1940	uint32_t cLeft = pRam->cb >> PAGE_SHIFT;
1941	while (cLeft-- > 0)
1942	{
1943	if (PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1944	{
1945	switch (PGM_PAGE_GET_STATE(pPage))
1946	{
1947	case PGM_PAGE_STATE_ZERO:
1948	{
1949	int rc = pgmPhysAllocPage(pVM, pPage, GCPhys);
1950	if (RT_FAILURE(rc))
1951	{
1952	LogRel(("PGM: RAM Pre-allocation failed at %RGp (in %s) with rc=%Rrc\n", GCPhys, pRam->pszDesc, rc));
1953	PGM_UNLOCK(pVM);
1954	return rc;
1955	}
1956	cPages++;
1957	break;
1958	}
1959
1960	case PGM_PAGE_STATE_BALLOONED:
1961	case PGM_PAGE_STATE_ALLOCATED:
1962	case PGM_PAGE_STATE_WRITE_MONITORED:
1963	case PGM_PAGE_STATE_SHARED:
1964	/* nothing to do here. */
1965	break;
1966	}
1967	}
1968
1969	/* next */
1970	pPage++;
1971	GCPhys += PAGE_SIZE;
1972	}
1973	}
1974	PGM_UNLOCK(pVM);
1975	NanoTS = RTTimeNanoTS() - NanoTS;
1976
1977	LogRel(("PGM: Pre-allocated %llu pages in %llu ms\n", cPages, NanoTS / 1000000));
1978	Log(("pgmR3PhysRamPreAllocate: returns VINF_SUCCESS\n"));
1979	return VINF_SUCCESS;
1980	}
1981
1982
1983	/**
1984	* Checks shared page checksums.
1985	*
1986	* @param pVM The cross context VM structure.
1987	*/
1988	void pgmR3PhysAssertSharedPageChecksums(PVM pVM)
1989	{
1990	#ifdef VBOX_STRICT
1991	PGM_LOCK_VOID(pVM);
1992
1993	if (pVM->pgm.s.cSharedPages > 0)
1994	{
1995	/*
1996	* Walk the ram ranges.
1997	*/
1998	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1999	{
2000	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2001	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2002
2003	while (iPage-- > 0)
2004	{
2005	PPGMPAGE pPage = &pRam->aPages[iPage];
2006	if (PGM_PAGE_IS_SHARED(pPage))
2007	{
2008	uint32_t u32Checksum = pPage->s.u2Unused0/* \| ((uint32_t)pPage->s.u2Unused1 << 8)*/;
2009	if (!u32Checksum)
2010	{
2011	RTGCPHYS GCPhysPage = pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT);
2012	void const *pvPage;
2013	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhysPage, &pvPage);
2014	if (RT_SUCCESS(rc))
2015	{
2016	uint32_t u32Checksum2 = RTCrc32(pvPage, PAGE_SIZE);
2017	# if 0
2018	AssertMsg((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum, ("GCPhysPage=%RGp\n", GCPhysPage));
2019	# else
2020	if ((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum)
2021	LogFlow(("shpg %#x @ %RGp %#x [OK]\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
2022	else
2023	AssertMsgFailed(("shpg %#x @ %RGp %#x\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
2024	# endif
2025	}
2026	else
2027	AssertRC(rc);
2028	}
2029	}
2030
2031	} /* for each page */
2032
2033	} /* for each ram range */
2034	}
2035
2036	PGM_UNLOCK(pVM);
2037	#endif /* VBOX_STRICT */
2038	NOREF(pVM);
2039	}
2040
2041
2042	/**
2043	* Resets the physical memory state.
2044	*
2045	* ASSUMES that the caller owns the PGM lock.
2046	*
2047	* @returns VBox status code.
2048	* @param pVM The cross context VM structure.
2049	*/
2050	int pgmR3PhysRamReset(PVM pVM)
2051	{
2052	PGM_LOCK_ASSERT_OWNER(pVM);
2053
2054	/* Reset the memory balloon. */
2055	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2056	AssertRC(rc);
2057
2058	#ifdef VBOX_WITH_PAGE_SHARING
2059	/* Clear all registered shared modules. */
2060	pgmR3PhysAssertSharedPageChecksums(pVM);
2061	rc = GMMR3ResetSharedModules(pVM);
2062	AssertRC(rc);
2063	#endif
2064	/* Reset counters. */
2065	pVM->pgm.s.cReusedSharedPages = 0;
2066	pVM->pgm.s.cBalloonedPages = 0;
2067
2068	return VINF_SUCCESS;
2069	}
2070
2071
2072	/**
2073	* Resets (zeros) the RAM after all devices and components have been reset.
2074	*
2075	* ASSUMES that the caller owns the PGM lock.
2076	*
2077	* @returns VBox status code.
2078	* @param pVM The cross context VM structure.
2079	*/
2080	int pgmR3PhysRamZeroAll(PVM pVM)
2081	{
2082	PGM_LOCK_ASSERT_OWNER(pVM);
2083
2084	/*
2085	* We batch up pages that should be freed instead of calling GMM for
2086	* each and every one of them.
2087	*/
2088	uint32_t cPendingPages = 0;
2089	PGMMFREEPAGESREQ pReq;
2090	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2091	AssertLogRelRCReturn(rc, rc);
2092
2093	/*
2094	* Walk the ram ranges.
2095	*/
2096	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2097	{
2098	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2099	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2100
2101	if ( !pVM->pgm.s.fRamPreAlloc
2102	&& pVM->pgm.s.fZeroRamPagesOnReset)
2103	{
2104	/* Replace all RAM pages by ZERO pages. */
2105	while (iPage-- > 0)
2106	{
2107	PPGMPAGE pPage = &pRam->aPages[iPage];
2108	switch (PGM_PAGE_GET_TYPE(pPage))
2109	{
2110	case PGMPAGETYPE_RAM:
2111	/* Do not replace pages part of a 2 MB continuous range
2112	with zero pages, but zero them instead. */
2113	if ( PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE
2114	\|\| PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE_DISABLED)
2115	{
2116	void *pvPage;
2117	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2118	AssertLogRelRCReturn(rc, rc);
2119	ASMMemZeroPage(pvPage);
2120	}
2121	else if (PGM_PAGE_IS_BALLOONED(pPage))
2122	{
2123	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2124	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2125	}
2126	else if (!PGM_PAGE_IS_ZERO(pPage))
2127	{
2128	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2129	PGMPAGETYPE_RAM);
2130	AssertLogRelRCReturn(rc, rc);
2131	}
2132	break;
2133
2134	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2135	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2136	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2137	true /fDoAccounting/);
2138	break;
2139
2140	case PGMPAGETYPE_MMIO2:
2141	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2142	case PGMPAGETYPE_ROM:
2143	case PGMPAGETYPE_MMIO:
2144	break;
2145	default:
2146	AssertFailed();
2147	}
2148	} /* for each page */
2149	}
2150	else
2151	{
2152	/* Zero the memory. */
2153	while (iPage-- > 0)
2154	{
2155	PPGMPAGE pPage = &pRam->aPages[iPage];
2156	switch (PGM_PAGE_GET_TYPE(pPage))
2157	{
2158	case PGMPAGETYPE_RAM:
2159	switch (PGM_PAGE_GET_STATE(pPage))
2160	{
2161	case PGM_PAGE_STATE_ZERO:
2162	break;
2163
2164	case PGM_PAGE_STATE_BALLOONED:
2165	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2166	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2167	break;
2168
2169	case PGM_PAGE_STATE_SHARED:
2170	case PGM_PAGE_STATE_WRITE_MONITORED:
2171	rc = pgmPhysPageMakeWritable(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2172	AssertLogRelRCReturn(rc, rc);
2173	RT_FALL_THRU();
2174
2175	case PGM_PAGE_STATE_ALLOCATED:
2176	if (pVM->pgm.s.fZeroRamPagesOnReset)
2177	{
2178	void *pvPage;
2179	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2180	AssertLogRelRCReturn(rc, rc);
2181	ASMMemZeroPage(pvPage);
2182	}
2183	break;
2184	}
2185	break;
2186
2187	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2188	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2189	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2190	true /fDoAccounting/);
2191	break;
2192
2193	case PGMPAGETYPE_MMIO2:
2194	case PGMPAGETYPE_ROM_SHADOW:
2195	case PGMPAGETYPE_ROM:
2196	case PGMPAGETYPE_MMIO:
2197	break;
2198	default:
2199	AssertFailed();
2200
2201	}
2202	} /* for each page */
2203	}
2204
2205	}
2206
2207	/*
2208	* Finish off any pages pending freeing.
2209	*/
2210	if (cPendingPages)
2211	{
2212	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2213	AssertLogRelRCReturn(rc, rc);
2214	}
2215	GMMR3FreePagesCleanup(pReq);
2216	return VINF_SUCCESS;
2217	}
2218
2219
2220	/**
2221	* Frees all RAM during VM termination
2222	*
2223	* ASSUMES that the caller owns the PGM lock.
2224	*
2225	* @returns VBox status code.
2226	* @param pVM The cross context VM structure.
2227	*/
2228	int pgmR3PhysRamTerm(PVM pVM)
2229	{
2230	PGM_LOCK_ASSERT_OWNER(pVM);
2231
2232	/* Reset the memory balloon. */
2233	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2234	AssertRC(rc);
2235
2236	#ifdef VBOX_WITH_PAGE_SHARING
2237	/*
2238	* Clear all registered shared modules.
2239	*/
2240	pgmR3PhysAssertSharedPageChecksums(pVM);
2241	rc = GMMR3ResetSharedModules(pVM);
2242	AssertRC(rc);
2243
2244	/*
2245	* Flush the handy pages updates to make sure no shared pages are hiding
2246	* in there. (No unlikely if the VM shuts down, apparently.)
2247	*/
2248	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_FLUSH_HANDY_PAGES, 0, NULL);
2249	#endif
2250
2251	/*
2252	* We batch up pages that should be freed instead of calling GMM for
2253	* each and every one of them.
2254	*/
2255	uint32_t cPendingPages = 0;
2256	PGMMFREEPAGESREQ pReq;
2257	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2258	AssertLogRelRCReturn(rc, rc);
2259
2260	/*
2261	* Walk the ram ranges.
2262	*/
2263	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2264	{
2265	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2266	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2267
2268	while (iPage-- > 0)
2269	{
2270	PPGMPAGE pPage = &pRam->aPages[iPage];
2271	switch (PGM_PAGE_GET_TYPE(pPage))
2272	{
2273	case PGMPAGETYPE_RAM:
2274	/* Free all shared pages. Private pages are automatically freed during GMM VM cleanup. */
2275	/** @todo change this to explicitly free private pages here. */
2276	if (PGM_PAGE_IS_SHARED(pPage))
2277	{
2278	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2279	PGMPAGETYPE_RAM);
2280	AssertLogRelRCReturn(rc, rc);
2281	}
2282	break;
2283
2284	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2285	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO:
2286	case PGMPAGETYPE_MMIO2:
2287	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2288	case PGMPAGETYPE_ROM:
2289	case PGMPAGETYPE_MMIO:
2290	break;
2291	default:
2292	AssertFailed();
2293	}
2294	} /* for each page */
2295	}
2296
2297	/*
2298	* Finish off any pages pending freeing.
2299	*/
2300	if (cPendingPages)
2301	{
2302	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2303	AssertLogRelRCReturn(rc, rc);
2304	}
2305	GMMR3FreePagesCleanup(pReq);
2306	return VINF_SUCCESS;
2307	}
2308
2309
2310	/**
2311	* This is the interface IOM is using to register an MMIO region.
2312	*
2313	* It will check for conflicts and ensure that a RAM range structure
2314	* is present before calling the PGMR3HandlerPhysicalRegister API to
2315	* register the callbacks.
2316	*
2317	* @returns VBox status code.
2318	*
2319	* @param pVM The cross context VM structure.
2320	* @param GCPhys The start of the MMIO region.
2321	* @param cb The size of the MMIO region.
2322	* @param hType The physical access handler type registration.
2323	* @param pvUserR3 The user argument for R3.
2324	* @param pvUserR0 The user argument for R0.
2325	* @param pvUserRC The user argument for RC.
2326	* @param pszDesc The description of the MMIO region.
2327	*/
2328	VMMR3DECL(int) PGMR3PhysMMIORegister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMPHYSHANDLERTYPE hType,
2329	RTR3PTR pvUserR3, RTR0PTR pvUserR0, RTRCPTR pvUserRC, const char *pszDesc)
2330	{
2331	/*
2332	* Assert on some assumption.
2333	*/
2334	VM_ASSERT_EMT(pVM);
2335	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2336	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2337	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2338	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2339	Assert(((PPGMPHYSHANDLERTYPEINT)MMHyperHeapOffsetToPtr(pVM, hType))->enmKind == PGMPHYSHANDLERKIND_MMIO);
2340
2341	int rc = PGM_LOCK(pVM);
2342	AssertRCReturn(rc, rc);
2343
2344	/*
2345	* Make sure there's a RAM range structure for the region.
2346	*/
2347	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2348	bool fRamExists = false;
2349	PPGMRAMRANGE pRamPrev = NULL;
2350	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2351	while (pRam && GCPhysLast >= pRam->GCPhys)
2352	{
2353	if ( GCPhysLast >= pRam->GCPhys
2354	&& GCPhys <= pRam->GCPhysLast)
2355	{
2356	/* Simplification: all within the same range. */
2357	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
2358	&& GCPhysLast <= pRam->GCPhysLast,
2359	("%RGp-%RGp (MMIO/%s) falls partly outside %RGp-%RGp (%s)\n",
2360	GCPhys, GCPhysLast, pszDesc,
2361	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2362	PGM_UNLOCK(pVM),
2363	VERR_PGM_RAM_CONFLICT);
2364
2365	/* Check that it's all RAM or MMIO pages. */
2366	PCPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2367	uint32_t cLeft = cb >> PAGE_SHIFT;
2368	while (cLeft-- > 0)
2369	{
2370	AssertLogRelMsgReturnStmt( PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM
2371	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO,
2372	("%RGp-%RGp (MMIO/%s): %RGp is not a RAM or MMIO page - type=%d desc=%s\n",
2373	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, PGM_PAGE_GET_TYPE(pPage), pRam->pszDesc),
2374	PGM_UNLOCK(pVM),
2375	VERR_PGM_RAM_CONFLICT);
2376	pPage++;
2377	}
2378
2379	/* Looks good. */
2380	fRamExists = true;
2381	break;
2382	}
2383
2384	/* next */
2385	pRamPrev = pRam;
2386	pRam = pRam->pNextR3;
2387	}
2388	PPGMRAMRANGE pNew;
2389	if (fRamExists)
2390	{
2391	pNew = NULL;
2392
2393	/*
2394	* Make all the pages in the range MMIO/ZERO pages, freeing any
2395	* RAM pages currently mapped here. This might not be 100% correct
2396	* for PCI memory, but we're doing the same thing for MMIO2 pages.
2397	*/
2398	rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, PGMPAGETYPE_MMIO);
2399	AssertRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
2400
2401	/* Force a PGM pool flush as guest ram references have been changed. */
2402	/** @todo not entirely SMP safe; assuming for now the guest takes
2403	* care of this internally (not touch mapped mmio while changing the
2404	* mapping). */
2405	PVMCPU pVCpu = VMMGetCpu(pVM);
2406	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2407	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2408	}
2409	else
2410	{
2411
2412	/*
2413	* No RAM range, insert an ad hoc one.
2414	*
2415	* Note that we don't have to tell REM about this range because
2416	* PGMHandlerPhysicalRegisterEx will do that for us.
2417	*/
2418	Log(("PGMR3PhysMMIORegister: Adding ad hoc MMIO range for %RGp-%RGp %s\n", GCPhys, GCPhysLast, pszDesc));
2419
2420	const uint32_t cPages = cb >> PAGE_SHIFT;
2421	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
2422	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), 16, MM_TAG_PGM_PHYS, (void **)&pNew);
2423	AssertLogRelMsgRCReturnStmt(rc, ("cbRamRange=%zu\n", cbRamRange), PGM_UNLOCK(pVM), rc);
2424
2425	/* Initialize the range. */
2426	pNew->pSelfR0 = MMHyperCCToR0(pVM, pNew);
2427	pNew->GCPhys = GCPhys;
2428	pNew->GCPhysLast = GCPhysLast;
2429	pNew->cb = cb;
2430	pNew->pszDesc = pszDesc;
2431	pNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO;
2432	pNew->pvR3 = NULL;
2433	pNew->paLSPages = NULL;
2434
2435	uint32_t iPage = cPages;
2436	while (iPage-- > 0)
2437	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_MMIO);
2438	Assert(PGM_PAGE_GET_TYPE(&pNew->aPages[0]) == PGMPAGETYPE_MMIO);
2439
2440	/* update the page count stats. */
2441	pVM->pgm.s.cPureMmioPages += cPages;
2442	pVM->pgm.s.cAllPages += cPages;
2443
2444	/* link it */
2445	pgmR3PhysLinkRamRange(pVM, pNew, pRamPrev);
2446	}
2447
2448	/*
2449	* Register the access handler.
2450	*/
2451	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, hType, pvUserR3, pvUserR0, pvUserRC, pszDesc);
2452	if ( RT_FAILURE(rc)
2453	&& !fRamExists)
2454	{
2455	pVM->pgm.s.cPureMmioPages -= cb >> PAGE_SHIFT;
2456	pVM->pgm.s.cAllPages -= cb >> PAGE_SHIFT;
2457
2458	/* remove the ad hoc range. */
2459	pgmR3PhysUnlinkRamRange2(pVM, pNew, pRamPrev);
2460	pNew->cb = pNew->GCPhys = pNew->GCPhysLast = NIL_RTGCPHYS;
2461	MMHyperFree(pVM, pRam);
2462	}
2463	pgmPhysInvalidatePageMapTLB(pVM);
2464
2465	PGM_UNLOCK(pVM);
2466	return rc;
2467	}
2468
2469
2470	/**
2471	* This is the interface IOM is using to register an MMIO region.
2472	*
2473	* It will take care of calling PGMHandlerPhysicalDeregister and clean up
2474	* any ad hoc PGMRAMRANGE left behind.
2475	*
2476	* @returns VBox status code.
2477	* @param pVM The cross context VM structure.
2478	* @param GCPhys The start of the MMIO region.
2479	* @param cb The size of the MMIO region.
2480	*/
2481	VMMR3DECL(int) PGMR3PhysMMIODeregister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb)
2482	{
2483	VM_ASSERT_EMT(pVM);
2484
2485	int rc = PGM_LOCK(pVM);
2486	AssertRCReturn(rc, rc);
2487
2488	/*
2489	* First deregister the handler, then check if we should remove the ram range.
2490	*/
2491	rc = PGMHandlerPhysicalDeregister(pVM, GCPhys);
2492	if (RT_SUCCESS(rc))
2493	{
2494	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2495	PPGMRAMRANGE pRamPrev = NULL;
2496	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2497	while (pRam && GCPhysLast >= pRam->GCPhys)
2498	{
2499	/** @todo We're being a bit too careful here. rewrite. */
2500	if ( GCPhysLast == pRam->GCPhysLast
2501	&& GCPhys == pRam->GCPhys)
2502	{
2503	Assert(pRam->cb == cb);
2504
2505	/*
2506	* See if all the pages are dead MMIO pages.
2507	*/
2508	uint32_t const cPages = cb >> PAGE_SHIFT;
2509	bool fAllMMIO = true;
2510	uint32_t iPage = 0;
2511	uint32_t cLeft = cPages;
2512	while (cLeft-- > 0)
2513	{
2514	PPGMPAGE pPage = &pRam->aPages[iPage];
2515	if ( !PGM_PAGE_IS_MMIO_OR_ALIAS(pPage)
2516	/\|\| not-out-of-action later /)
2517	{
2518	fAllMMIO = false;
2519	AssertMsgFailed(("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2520	break;
2521	}
2522	Assert( PGM_PAGE_IS_ZERO(pPage)
2523	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2524	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO);
2525	pPage++;
2526	}
2527	if (fAllMMIO)
2528	{
2529	/*
2530	* Ad-hoc range, unlink and free it.
2531	*/
2532	Log(("PGMR3PhysMMIODeregister: Freeing ad hoc MMIO range for %RGp-%RGp %s\n",
2533	GCPhys, GCPhysLast, pRam->pszDesc));
2534
2535	pVM->pgm.s.cAllPages -= cPages;
2536	pVM->pgm.s.cPureMmioPages -= cPages;
2537
2538	pgmR3PhysUnlinkRamRange2(pVM, pRam, pRamPrev);
2539	pRam->cb = pRam->GCPhys = pRam->GCPhysLast = NIL_RTGCPHYS;
2540	MMHyperFree(pVM, pRam);
2541	break;
2542	}
2543	}
2544
2545	/*
2546	* Range match? It will all be within one range (see PGMAllHandler.cpp).
2547	*/
2548	if ( GCPhysLast >= pRam->GCPhys
2549	&& GCPhys <= pRam->GCPhysLast)
2550	{
2551	Assert(GCPhys >= pRam->GCPhys);
2552	Assert(GCPhysLast <= pRam->GCPhysLast);
2553
2554	/*
2555	* Turn the pages back into RAM pages.
2556	*/
2557	uint32_t iPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2558	uint32_t cLeft = cb >> PAGE_SHIFT;
2559	while (cLeft--)
2560	{
2561	PPGMPAGE pPage = &pRam->aPages[iPage];
2562	AssertMsg( (PGM_PAGE_IS_MMIO(pPage) && PGM_PAGE_IS_ZERO(pPage))
2563	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2564	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO,
2565	("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2566	if (PGM_PAGE_IS_MMIO_OR_ALIAS(pPage))
2567	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_RAM);
2568	}
2569	break;
2570	}
2571
2572	/* next */
2573	pRamPrev = pRam;
2574	pRam = pRam->pNextR3;
2575	}
2576	}
2577
2578	/* Force a PGM pool flush as guest ram references have been changed. */
2579	/** @todo Not entirely SMP safe; assuming for now the guest takes care of
2580	* this internally (not touch mapped mmio while changing the mapping). */
2581	PVMCPU pVCpu = VMMGetCpu(pVM);
2582	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2583	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2584
2585	pgmPhysInvalidatePageMapTLB(pVM);
2586	pgmPhysInvalidRamRangeTlbs(pVM);
2587	PGM_UNLOCK(pVM);
2588	return rc;
2589	}
2590
2591
2592	/**
2593	* Locate a MMIO2 range.
2594	*
2595	* @returns Pointer to the MMIO2 range.
2596	* @param pVM The cross context VM structure.
2597	* @param pDevIns The device instance owning the region.
2598	* @param iSubDev The sub-device number.
2599	* @param iRegion The region.
2600	* @param hMmio2 Handle to look up. If NIL, use the @a iSubDev and
2601	* @a iRegion.
2602	*/
2603	DECLINLINE(PPGMREGMMIO2RANGE) pgmR3PhysMmio2Find(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev,
2604	uint32_t iRegion, PGMMMIO2HANDLE hMmio2)
2605	{
2606	if (hMmio2 != NIL_PGMMMIO2HANDLE)
2607	{
2608	if (hMmio2 <= RT_ELEMENTS(pVM->pgm.s.apMmio2RangesR3) && hMmio2 != 0)
2609	{
2610	PPGMREGMMIO2RANGE pCur = pVM->pgm.s.apMmio2RangesR3[hMmio2 - 1];
2611	if (pCur && pCur->pDevInsR3 == pDevIns)
2612	{
2613	Assert(pCur->idMmio2 == hMmio2);
2614	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_MMIO2, NULL);
2615	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, NULL);
2616	return pCur;
2617	}
2618	Assert(!pCur);
2619	}
2620	for (PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2621	if (pCur->idMmio2 == hMmio2)
2622	{
2623	AssertBreak(pCur->pDevInsR3 == pDevIns);
2624	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_MMIO2, NULL);
2625	AssertReturn(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, NULL);
2626	return pCur;
2627	}
2628	}
2629	else
2630	{
2631	/*
2632	* Search the list. There shouldn't be many entries.
2633	*/
2634	/** @todo Optimize this lookup! There may now be many entries and it'll
2635	* become really slow when doing MMR3HyperMapMMIO2 and similar. */
2636	for (PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2637	if ( pCur->pDevInsR3 == pDevIns
2638	&& pCur->iRegion == iRegion
2639	&& pCur->iSubDev == iSubDev)
2640	return pCur;
2641	}
2642	return NULL;
2643	}
2644
2645
2646	#ifndef PGM_WITHOUT_MAPPINGS
2647	/**
2648	* @callback_method_impl{FNPGMRELOCATE, Relocate a floating MMIO/MMIO2 range.}
2649	* @sa pgmR3PhysRamRangeRelocate
2650	*/
2651	static DECLCALLBACK(bool) pgmR3PhysMmio2RangeRelocate(PVM pVM, RTGCPTR GCPtrOld, RTGCPTR GCPtrNew,
2652	PGMRELOCATECALL enmMode, void *pvUser)
2653	{
2654	PPGMREGMMIO2RANGE pMmio = (PPGMREGMMIO2RANGE)pvUser;
2655	Assert(pMmio->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING);
2656	Assert(pMmio->RamRange.pSelfRC == GCPtrOld + PAGE_SIZE + RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange)); RT_NOREF_PV(GCPtrOld);
2657
2658	switch (enmMode)
2659	{
2660	case PGMRELOCATECALL_SUGGEST:
2661	return true;
2662
2663	case PGMRELOCATECALL_RELOCATE:
2664	{
2665	/*
2666	* Update myself, then relink all the ranges and flush the RC TLB.
2667	*/
2668	PGM_LOCK_VOID(pVM);
2669
2670	pMmio->RamRange.pSelfRC = (RTRCPTR)(GCPtrNew + PAGE_SIZE + RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange));
2671
2672	pgmR3PhysRelinkRamRanges(pVM);
2673	for (unsigned i = 0; i < PGM_RAMRANGE_TLB_ENTRIES; i++)
2674	pVM->pgm.s.apRamRangesTlbRC[i] = NIL_RTRCPTR;
2675
2676	PGM_UNLOCK(pVM);
2677	return true;
2678	}
2679
2680	default:
2681	AssertFailedReturn(false);
2682	}
2683	}
2684	#endif /* !PGM_WITHOUT_MAPPINGS */
2685
2686
2687	/**
2688	* Calculates the number of chunks
2689	*
2690	* @returns Number of registration chunk needed.
2691	* @param pVM The cross context VM structure.
2692	* @param cb The size of the MMIO/MMIO2 range.
2693	* @param pcPagesPerChunk Where to return the number of pages tracked by each
2694	* chunk. Optional.
2695	* @param pcbChunk Where to return the guest mapping size for a chunk.
2696	*/
2697	static uint16_t pgmR3PhysMmio2CalcChunkCount(PVM pVM, RTGCPHYS cb, uint32_t pcPagesPerChunk, uint32_t pcbChunk)
2698	{
2699	RT_NOREF_PV(pVM); /* without raw mode */
2700
2701	/*
2702	* This is the same calculation as PGMR3PhysRegisterRam does, except we'll be
2703	* needing a few bytes extra the PGMREGMMIO2RANGE structure.
2704	*
2705	* Note! In additions, we've got a 24 bit sub-page range for MMIO2 ranges, leaving
2706	* us with an absolute maximum of 16777215 pages per chunk (close to 64 GB).
2707	*/
2708	uint32_t cbChunk = 16U*_1M;
2709	uint32_t cPagesPerChunk = 1048048; /* max ~1048059 */
2710	AssertCompile(sizeof(PGMREGMMIO2RANGE) + sizeof(PGMPAGE) * 1048048 < 16U_1M - PAGE_SIZE 2);
2711	AssertRelease(cPagesPerChunk <= PGM_MMIO2_MAX_PAGE_COUNT); /* See above note. */
2712	AssertRelease(RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPagesPerChunk]) + PAGE_SIZE * 2 <= cbChunk);
2713	if (pcbChunk)
2714	*pcbChunk = cbChunk;
2715	if (pcPagesPerChunk)
2716	*pcPagesPerChunk = cPagesPerChunk;
2717
2718	/* Calc the number of chunks we need. */
2719	RTGCPHYS const cPages = cb >> X86_PAGE_SHIFT;
2720	uint16_t cChunks = (uint16_t)((cPages + cPagesPerChunk - 1) / cPagesPerChunk);
2721	AssertRelease((RTGCPHYS)cChunks * cPagesPerChunk >= cPages);
2722	return cChunks;
2723	}
2724
2725
2726	/**
2727	* Worker for PGMR3PhysMMIO2Register that allocates and the PGMREGMMIO2RANGE
2728	* structures and does basic initialization.
2729	*
2730	* Caller must set type specfic members and initialize the PGMPAGE structures.
2731	*
2732	* This was previously also used by PGMR3PhysMmio2PreRegister, a function for
2733	* pre-registering MMIO that was later (6.1) replaced by a new handle based IOM
2734	* interface. The reference to caller and type above is purely historical.
2735	*
2736	* @returns VBox status code.
2737	* @param pVM The cross context VM structure.
2738	* @param pDevIns The device instance owning the region.
2739	* @param iSubDev The sub-device number (internal PCI config number).
2740	* @param iRegion The region number. If the MMIO2 memory is a PCI
2741	* I/O region this number has to be the number of that
2742	* region. Otherwise it can be any number safe
2743	* UINT8_MAX.
2744	* @param cb The size of the region. Must be page aligned.
2745	* @param pszDesc The description.
2746	* @param ppHeadRet Where to return the pointer to the first
2747	* registration chunk.
2748	*
2749	* @thread EMT
2750	*/
2751	static int pgmR3PhysMmio2Create(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb,
2752	const char pszDesc, PPGMREGMMIO2RANGE ppHeadRet)
2753	{
2754	/*
2755	* Figure out how many chunks we need and of which size.
2756	*/
2757	uint32_t cPagesPerChunk;
2758	uint16_t cChunks = pgmR3PhysMmio2CalcChunkCount(pVM, cb, &cPagesPerChunk, NULL);
2759	AssertReturn(cChunks, VERR_PGM_PHYS_MMIO_EX_IPE);
2760
2761	/*
2762	* Allocate the chunks.
2763	*/
2764	PPGMREGMMIO2RANGE *ppNext = ppHeadRet;
2765	*ppNext = NULL;
2766
2767	int rc = VINF_SUCCESS;
2768	uint32_t cPagesLeft = cb >> X86_PAGE_SHIFT;
2769	for (uint16_t iChunk = 0; iChunk < cChunks && RT_SUCCESS(rc); iChunk++)
2770	{
2771	/*
2772	* We currently do a single RAM range for the whole thing. This will
2773	* probably have to change once someone needs really large MMIO regions,
2774	* as we will be running into SUPR3PageAllocEx limitations and such.
2775	*/
2776	const uint32_t cPagesTrackedByChunk = RT_MIN(cPagesLeft, cPagesPerChunk);
2777	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPagesTrackedByChunk]);
2778	PPGMREGMMIO2RANGE pNew = NULL;
2779	if ( iChunk + 1 < cChunks
2780	\|\| cbRange >= _1M)
2781	{
2782	/*
2783	* Allocate memory for the registration structure.
2784	*/
2785	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2786	size_t const cbChunk = (1 + cChunkPages + 1) << PAGE_SHIFT;
2787	AssertLogRelBreakStmt(cbChunk == (uint32_t)cbChunk, rc = VERR_OUT_OF_RANGE);
2788	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
2789	AssertBreakStmt(paChunkPages, rc = VERR_NO_TMP_MEMORY);
2790	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
2791	void *pvChunk = NULL;
2792	rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk, &R0PtrChunk, paChunkPages);
2793	AssertLogRelMsgRCBreakStmt(rc, ("rc=%Rrc, cChunkPages=%#zx\n", rc, cChunkPages), RTMemTmpFree(paChunkPages));
2794
2795	Assert(R0PtrChunk != NIL_RTR0PTR);
2796	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
2797
2798	pNew = (PPGMREGMMIO2RANGE)pvChunk;
2799	pNew->RamRange.fFlags = PGM_RAM_RANGE_FLAGS_FLOATING;
2800	pNew->RamRange.pSelfR0 = R0PtrChunk + RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange);
2801
2802	RTMemTmpFree(paChunkPages);
2803	}
2804	/*
2805	* Not so big, do a one time hyper allocation.
2806	*/
2807	else
2808	{
2809	rc = MMR3HyperAllocOnceNoRel(pVM, cbRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
2810	AssertLogRelMsgRCBreak(rc, ("cbRange=%zu\n", cbRange));
2811
2812	/*
2813	* Initialize allocation specific items.
2814	*/
2815	//pNew->RamRange.fFlags = 0;
2816	pNew->RamRange.pSelfR0 = MMHyperCCToR0(pVM, &pNew->RamRange);
2817	}
2818
2819	/*
2820	* Initialize the registration structure (caller does specific bits).
2821	*/
2822	pNew->pDevInsR3 = pDevIns;
2823	//pNew->pvR3 = NULL;
2824	//pNew->pNext = NULL;
2825	//pNew->fFlags = 0;
2826	if (iChunk == 0)
2827	pNew->fFlags \|= PGMREGMMIO2RANGE_F_FIRST_CHUNK;
2828	if (iChunk + 1 == cChunks)
2829	pNew->fFlags \|= PGMREGMMIO2RANGE_F_LAST_CHUNK;
2830	pNew->iSubDev = iSubDev;
2831	pNew->iRegion = iRegion;
2832	pNew->idSavedState = UINT8_MAX;
2833	pNew->idMmio2 = UINT8_MAX;
2834	//pNew->pPhysHandlerR3 = NULL;
2835	//pNew->paLSPages = NULL;
2836	pNew->RamRange.GCPhys = NIL_RTGCPHYS;
2837	pNew->RamRange.GCPhysLast = NIL_RTGCPHYS;
2838	pNew->RamRange.pszDesc = pszDesc;
2839	pNew->RamRange.cb = pNew->cbReal = (RTGCPHYS)cPagesTrackedByChunk << X86_PAGE_SHIFT;
2840	pNew->RamRange.fFlags \|= PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO_EX;
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	Assert(cPagesLeft == 0);
2850
2851	if (RT_SUCCESS(rc))
2852	{
2853	Assert((*ppHeadRet)->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
2854	return VINF_SUCCESS;
2855	}
2856
2857	/*
2858	* Free floating ranges.
2859	*/
2860	while (*ppHeadRet)
2861	{
2862	PPGMREGMMIO2RANGE pFree = *ppHeadRet;
2863	*ppHeadRet = pFree->pNextR3;
2864
2865	if (pFree->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
2866	{
2867	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[pFree->RamRange.cb >> X86_PAGE_SHIFT]);
2868	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2869	SUPR3PageFreeEx(pFree, cChunkPages);
2870	}
2871	}
2872
2873	return rc;
2874	}
2875
2876
2877	/**
2878	* Common worker PGMR3PhysMmio2PreRegister & PGMR3PhysMMIO2Register that links a
2879	* complete registration entry into the lists and lookup tables.
2880	*
2881	* @param pVM The cross context VM structure.
2882	* @param pNew The new MMIO / MMIO2 registration to link.
2883	*/
2884	static void pgmR3PhysMmio2Link(PVM pVM, PPGMREGMMIO2RANGE pNew)
2885	{
2886	/*
2887	* Link it into the list (order doesn't matter, so insert it at the head).
2888	*
2889	* Note! The range we're linking may consist of multiple chunks, so we
2890	* have to find the last one.
2891	*/
2892	PPGMREGMMIO2RANGE pLast = pNew;
2893	for (pLast = pNew; ; pLast = pLast->pNextR3)
2894	{
2895	if (pLast->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2896	break;
2897	Assert(pLast->pNextR3);
2898	Assert(pLast->pNextR3->pDevInsR3 == pNew->pDevInsR3);
2899	Assert(pLast->pNextR3->iSubDev == pNew->iSubDev);
2900	Assert(pLast->pNextR3->iRegion == pNew->iRegion);
2901	Assert((pLast->pNextR3->fFlags & PGMREGMMIO2RANGE_F_MMIO2) == (pNew->fFlags & PGMREGMMIO2RANGE_F_MMIO2));
2902	Assert(pLast->pNextR3->idMmio2 == (pLast->fFlags & PGMREGMMIO2RANGE_F_MMIO2 ? pLast->idMmio2 + 1 : UINT8_MAX));
2903	}
2904
2905	PGM_LOCK_VOID(pVM);
2906
2907	/* Link in the chain of ranges at the head of the list. */
2908	pLast->pNextR3 = pVM->pgm.s.pRegMmioRangesR3;
2909	pVM->pgm.s.pRegMmioRangesR3 = pNew;
2910
2911	/* If MMIO, insert the MMIO2 range/page IDs. */
2912	uint8_t idMmio2 = pNew->idMmio2;
2913	if (idMmio2 != UINT8_MAX)
2914	{
2915	for (;;)
2916	{
2917	Assert(pNew->fFlags & PGMREGMMIO2RANGE_F_MMIO2);
2918	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == NULL);
2919	Assert(pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] == NIL_RTR0PTR);
2920	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = pNew;
2921	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = pNew->RamRange.pSelfR0 - RT_UOFFSETOF(PGMREGMMIO2RANGE, RamRange);
2922	if (pNew->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
2923	break;
2924	pNew = pNew->pNextR3;
2925	idMmio2++;
2926	}
2927	}
2928	else
2929	Assert(!(pNew->fFlags & PGMREGMMIO2RANGE_F_MMIO2));
2930
2931	pgmPhysInvalidatePageMapTLB(pVM);
2932	PGM_UNLOCK(pVM);
2933	}
2934
2935
2936	/**
2937	* Allocate and register an MMIO2 region.
2938	*
2939	* As mentioned elsewhere, MMIO2 is just RAM spelled differently. It's RAM
2940	* associated with a device. It is also non-shared memory with a permanent
2941	* ring-3 mapping and page backing (presently).
2942	*
2943	* A MMIO2 range may overlap with base memory if a lot of RAM is configured for
2944	* the VM, in which case we'll drop the base memory pages. Presently we will
2945	* make no attempt to preserve anything that happens to be present in the base
2946	* memory that is replaced, this is of course incorrect but it's too much
2947	* effort.
2948	*
2949	* @returns VBox status code.
2950	* @retval VINF_SUCCESS on success, *ppv pointing to the R3 mapping of the
2951	* memory.
2952	* @retval VERR_ALREADY_EXISTS if the region already exists.
2953	*
2954	* @param pVM The cross context VM structure.
2955	* @param pDevIns The device instance owning the region.
2956	* @param iSubDev The sub-device number.
2957	* @param iRegion The region number. If the MMIO2 memory is a PCI
2958	* I/O region this number has to be the number of that
2959	* region. Otherwise it can be any number save
2960	* UINT8_MAX.
2961	* @param cb The size of the region. Must be page aligned.
2962	* @param fFlags Reserved for future use, must be zero.
2963	* @param pszDesc The description.
2964	* @param ppv Where to store the pointer to the ring-3 mapping of
2965	* the memory.
2966	* @param phRegion Where to return the MMIO2 region handle. Optional.
2967	* @thread EMT
2968	*/
2969	VMMR3_INT_DECL(int) PGMR3PhysMmio2Register(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb,
2970	uint32_t fFlags, const char pszDesc, void ppv, PGMMMIO2HANDLE phRegion)
2971	{
2972	/*
2973	* Validate input.
2974	*/
2975	AssertPtrReturn(ppv, VERR_INVALID_POINTER);
2976	*ppv = NULL;
2977	if (phRegion)
2978	{
2979	AssertPtrReturn(phRegion, VERR_INVALID_POINTER);
2980	*phRegion = NIL_PGMMMIO2HANDLE;
2981	}
2982	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2983	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2984	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
2985	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
2986	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2987	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2988	AssertReturn(pgmR3PhysMmio2Find(pVM, pDevIns, iSubDev, iRegion, NIL_PGMMMIO2HANDLE) == NULL, VERR_ALREADY_EXISTS);
2989	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2990	AssertReturn(cb, VERR_INVALID_PARAMETER);
2991	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2992
2993	const uint32_t cPages = cb >> PAGE_SHIFT;
2994	AssertLogRelReturn(((RTGCPHYS)cPages << PAGE_SHIFT) == cb, VERR_INVALID_PARAMETER);
2995	AssertLogRelReturn(cPages <= (MM_MMIO_64_MAX >> X86_PAGE_SHIFT), VERR_OUT_OF_RANGE);
2996	AssertLogRelReturn(cPages <= PGM_MMIO2_MAX_PAGE_COUNT, VERR_OUT_OF_RANGE);
2997
2998	/*
2999	* For the 2nd+ instance, mangle the description string so it's unique.
3000	*/
3001	if (pDevIns->iInstance > 0) /** @todo Move to PDMDevHlp.cpp and use a real string cache. */
3002	{
3003	pszDesc = MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s [%u]", pszDesc, pDevIns->iInstance);
3004	if (!pszDesc)
3005	return VERR_NO_MEMORY;
3006	}
3007
3008	/*
3009	* Allocate an MMIO2 range ID (not freed on failure).
3010	*
3011	* The zero ID is not used as it could be confused with NIL_GMM_PAGEID, so
3012	* the IDs goes from 1 thru PGM_MMIO2_MAX_RANGES.
3013	*/
3014	unsigned cChunks = pgmR3PhysMmio2CalcChunkCount(pVM, cb, NULL, NULL);
3015	PGM_LOCK_VOID(pVM);
3016	uint8_t idMmio2 = pVM->pgm.s.cMmio2Regions + 1;
3017	unsigned cNewMmio2Regions = pVM->pgm.s.cMmio2Regions + cChunks;
3018	if (cNewMmio2Regions > PGM_MMIO2_MAX_RANGES)
3019	{
3020	PGM_UNLOCK(pVM);
3021	AssertLogRelFailedReturn(VERR_PGM_TOO_MANY_MMIO2_RANGES);
3022	}
3023	pVM->pgm.s.cMmio2Regions = cNewMmio2Regions;
3024	PGM_UNLOCK(pVM);
3025
3026	/*
3027	* Try reserve and allocate the backing memory first as this is what is
3028	* most likely to fail.
3029	*/
3030	int rc = MMR3AdjustFixedReservation(pVM, cPages, pszDesc);
3031	if (RT_SUCCESS(rc))
3032	{
3033	PSUPPAGE paPages = (PSUPPAGE)RTMemTmpAlloc(cPages * sizeof(SUPPAGE));
3034	if (RT_SUCCESS(rc))
3035	{
3036	void *pvPages;
3037	#ifndef VBOX_WITH_LINEAR_HOST_PHYS_MEM
3038	RTR0PTR pvPagesR0;
3039	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, &pvPagesR0, paPages);
3040	#else
3041	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, NULL /pR0Ptr/, paPages);
3042	#endif
3043	if (RT_SUCCESS(rc))
3044	{
3045	memset(pvPages, 0, cPages * PAGE_SIZE);
3046
3047	/*
3048	* Create the registered MMIO range record for it.
3049	*/
3050	PPGMREGMMIO2RANGE pNew;
3051	rc = pgmR3PhysMmio2Create(pVM, pDevIns, iSubDev, iRegion, cb, pszDesc, &pNew);
3052	if (RT_SUCCESS(rc))
3053	{
3054	if (phRegion)
3055	phRegion = idMmio2; / The ID of the first chunk. */
3056
3057	uint32_t iSrcPage = 0;
3058	uint8_t pbCurPages = (uint8_t )pvPages;
3059	for (PPGMREGMMIO2RANGE pCur = pNew; pCur; pCur = pCur->pNextR3)
3060	{
3061	pCur->pvR3 = pbCurPages;
3062	#ifndef VBOX_WITH_LINEAR_HOST_PHYS_MEM
3063	pCur->pvR0 = pvPagesR0 + (iSrcPage << PAGE_SHIFT);
3064	#endif
3065	pCur->RamRange.pvR3 = pbCurPages;
3066	pCur->idMmio2 = idMmio2;
3067	pCur->fFlags \|= PGMREGMMIO2RANGE_F_MMIO2;
3068
3069	uint32_t iDstPage = pCur->RamRange.cb >> X86_PAGE_SHIFT;
3070	while (iDstPage-- > 0)
3071	{
3072	PGM_PAGE_INIT(&pNew->RamRange.aPages[iDstPage],
3073	paPages[iDstPage + iSrcPage].Phys,
3074	PGM_MMIO2_PAGEID_MAKE(idMmio2, iDstPage),
3075	PGMPAGETYPE_MMIO2, PGM_PAGE_STATE_ALLOCATED);
3076	}
3077
3078	/* advance. */
3079	iSrcPage += pCur->RamRange.cb >> X86_PAGE_SHIFT;
3080	pbCurPages += pCur->RamRange.cb;
3081	idMmio2++;
3082	}
3083
3084	RTMemTmpFree(paPages);
3085
3086	/*
3087	* Update the page count stats, link the registration and we're done.
3088	*/
3089	pVM->pgm.s.cAllPages += cPages;
3090	pVM->pgm.s.cPrivatePages += cPages;
3091
3092	pgmR3PhysMmio2Link(pVM, pNew);
3093
3094	*ppv = pvPages;
3095	return VINF_SUCCESS;
3096	}
3097
3098	SUPR3PageFreeEx(pvPages, cPages);
3099	}
3100	}
3101	RTMemTmpFree(paPages);
3102	MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pszDesc);
3103	}
3104	if (pDevIns->iInstance > 0)
3105	MMR3HeapFree((void *)pszDesc);
3106	return rc;
3107	}
3108
3109
3110	/**
3111	* Deregisters and frees an MMIO2 region.
3112	*
3113	* Any physical access handlers registered for the region must be deregistered
3114	* before calling this function.
3115	*
3116	* @returns VBox status code.
3117	* @param pVM The cross context VM structure.
3118	* @param pDevIns The device instance owning the region.
3119	* @param hMmio2 The MMIO2 handle to deregister, or NIL if all
3120	* regions for the given device is to be deregistered.
3121	*/
3122	VMMR3_INT_DECL(int) PGMR3PhysMmio2Deregister(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3123	{
3124	/*
3125	* Validate input.
3126	*/
3127	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3128	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3129
3130	/*
3131	* The loop here scanning all registrations will make sure that multi-chunk ranges
3132	* get properly deregistered, though it's original purpose was the wildcard iRegion.
3133	*/
3134	PGM_LOCK_VOID(pVM);
3135	int rc = VINF_SUCCESS;
3136	unsigned cFound = 0;
3137	PPGMREGMMIO2RANGE pPrev = NULL;
3138	PPGMREGMMIO2RANGE pCur = pVM->pgm.s.pRegMmioRangesR3;
3139	while (pCur)
3140	{
3141	uint32_t const fFlags = pCur->fFlags;
3142	if ( pCur->pDevInsR3 == pDevIns
3143	&& ( hMmio2 == NIL_PGMMMIO2HANDLE
3144	\|\| pCur->idMmio2 == hMmio2))
3145	{
3146	Assert(fFlags & PGMREGMMIO2RANGE_F_MMIO2);
3147	cFound++;
3148
3149	/*
3150	* Unmap it if it's mapped.
3151	*/
3152	if (fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3153	{
3154	int rc2 = PGMR3PhysMmio2Unmap(pVM, pCur->pDevInsR3, pCur->idMmio2, pCur->RamRange.GCPhys);
3155	AssertRC(rc2);
3156	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3157	rc = rc2;
3158	}
3159
3160	/*
3161	* Unlink it
3162	*/
3163	PPGMREGMMIO2RANGE pNext = pCur->pNextR3;
3164	if (pPrev)
3165	pPrev->pNextR3 = pNext;
3166	else
3167	pVM->pgm.s.pRegMmioRangesR3 = pNext;
3168	pCur->pNextR3 = NULL;
3169
3170	uint8_t idMmio2 = pCur->idMmio2;
3171	if (idMmio2 != UINT8_MAX)
3172	{
3173	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == pCur);
3174	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = NULL;
3175	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = NIL_RTR0PTR;
3176	}
3177
3178	/*
3179	* Free the memory.
3180	*/
3181	const bool fIsMmio2 = RT_BOOL(fFlags & PGMREGMMIO2RANGE_F_MMIO2);
3182	uint32_t const cPages = pCur->cbReal >> PAGE_SHIFT;
3183	if (fIsMmio2)
3184	{
3185	int rc2 = SUPR3PageFreeEx(pCur->pvR3, cPages);
3186	AssertRC(rc2);
3187	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3188	rc = rc2;
3189
3190	rc2 = MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pCur->RamRange.pszDesc);
3191	AssertRC(rc2);
3192	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3193	rc = rc2;
3194	}
3195
3196	/* we're leaking hyper memory here if done at runtime. */
3197	#ifdef VBOX_STRICT
3198	VMSTATE const enmState = VMR3GetState(pVM);
3199	AssertMsg( enmState == VMSTATE_POWERING_OFF
3200	\|\| enmState == VMSTATE_POWERING_OFF_LS
3201	\|\| enmState == VMSTATE_OFF
3202	\|\| enmState == VMSTATE_OFF_LS
3203	\|\| enmState == VMSTATE_DESTROYING
3204	\|\| enmState == VMSTATE_TERMINATED
3205	\|\| enmState == VMSTATE_CREATING
3206	, ("%s\n", VMR3GetStateName(enmState)));
3207	#endif
3208
3209	if (pCur->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
3210	{
3211	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIO2RANGE, RamRange.aPages[cPages]);
3212	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
3213	SUPR3PageFreeEx(pCur, cChunkPages);
3214	}
3215	/*else
3216	{
3217	rc = MMHyperFree(pVM, pCur); - does not work, see the alloc call.
3218	AssertRCReturn(rc, rc);
3219	} */
3220
3221
3222	/* update page count stats */
3223	pVM->pgm.s.cAllPages -= cPages;
3224	if (fIsMmio2)
3225	pVM->pgm.s.cPrivatePages -= cPages;
3226	else
3227	pVM->pgm.s.cPureMmioPages -= cPages;
3228
3229	/* next */
3230	pCur = pNext;
3231	if (hMmio2 != NIL_PGMMMIO2HANDLE)
3232	{
3233	if (fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3234	break;
3235	hMmio2++;
3236	Assert(pCur->idMmio2 == hMmio2);
3237	Assert(pCur->pDevInsR3 == pDevIns);
3238	Assert(!(pCur->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK));
3239	}
3240	}
3241	else
3242	{
3243	pPrev = pCur;
3244	pCur = pCur->pNextR3;
3245	}
3246	}
3247	pgmPhysInvalidatePageMapTLB(pVM);
3248	PGM_UNLOCK(pVM);
3249	return !cFound && hMmio2 != NIL_PGMMMIO2HANDLE ? VERR_NOT_FOUND : rc;
3250	}
3251
3252
3253	/**
3254	* Maps a MMIO2 region.
3255	*
3256	* This is typically done when a guest / the bios / state loading changes the
3257	* PCI config. The replacing of base memory has the same restrictions as during
3258	* registration, of course.
3259	*
3260	* @returns VBox status code.
3261	*
3262	* @param pVM The cross context VM structure.
3263	* @param pDevIns The device instance owning the region.
3264	* @param hMmio2 The handle of the region to map.
3265	* @param GCPhys The guest-physical address to be remapped.
3266	*/
3267	VMMR3_INT_DECL(int) PGMR3PhysMmio2Map(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS GCPhys)
3268	{
3269	/*
3270	* Validate input.
3271	*
3272	* Note! It's safe to walk the MMIO/MMIO2 list since registrations only
3273	* happens during VM construction.
3274	*/
3275	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3276	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3277	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
3278	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3279	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3280	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3281
3282	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3283	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3284	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3285
3286	PPGMREGMMIO2RANGE pLastMmio = pFirstMmio;
3287	RTGCPHYS cbRange = 0;
3288	for (;;)
3289	{
3290	AssertReturn(!(pLastMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED), VERR_WRONG_ORDER);
3291	Assert(pLastMmio->RamRange.GCPhys == NIL_RTGCPHYS);
3292	Assert(pLastMmio->RamRange.GCPhysLast == NIL_RTGCPHYS);
3293	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3294	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3295	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3296	cbRange += pLastMmio->RamRange.cb;
3297	if (pLastMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3298	break;
3299	pLastMmio = pLastMmio->pNextR3;
3300	}
3301
3302	RTGCPHYS GCPhysLast = GCPhys + cbRange - 1;
3303	AssertLogRelReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3304
3305	/*
3306	* Find our location in the ram range list, checking for restriction
3307	* we don't bother implementing yet (partially overlapping, multiple
3308	* ram ranges).
3309	*/
3310	PGM_LOCK_VOID(pVM);
3311
3312	AssertReturnStmt(!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED), PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3313
3314	bool fRamExists = false;
3315	PPGMRAMRANGE pRamPrev = NULL;
3316	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3317	while (pRam && GCPhysLast >= pRam->GCPhys)
3318	{
3319	if ( GCPhys <= pRam->GCPhysLast
3320	&& GCPhysLast >= pRam->GCPhys)
3321	{
3322	/* Completely within? */
3323	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
3324	&& GCPhysLast <= pRam->GCPhysLast,
3325	("%RGp-%RGp (MMIOEx/%s) falls partly outside %RGp-%RGp (%s)\n",
3326	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc,
3327	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3328	PGM_UNLOCK(pVM),
3329	VERR_PGM_RAM_CONFLICT);
3330
3331	/* Check that all the pages are RAM pages. */
3332	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3333	uint32_t cPagesLeft = cbRange >> PAGE_SHIFT;
3334	while (cPagesLeft-- > 0)
3335	{
3336	AssertLogRelMsgReturnStmt(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
3337	("%RGp isn't a RAM page (%d) - mapping %RGp-%RGp (MMIO2/%s).\n",
3338	GCPhys, PGM_PAGE_GET_TYPE(pPage), GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc),
3339	PGM_UNLOCK(pVM),
3340	VERR_PGM_RAM_CONFLICT);
3341	pPage++;
3342	}
3343
3344	/* There can only be one MMIO/MMIO2 chunk matching here! */
3345	AssertLogRelMsgReturnStmt(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK,
3346	("%RGp-%RGp (MMIOEx/%s, flags %#X) consists of multiple chunks whereas the RAM somehow doesn't!\n",
3347	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3348	PGM_UNLOCK(pVM),
3349	VERR_PGM_PHYS_MMIO_EX_IPE);
3350
3351	fRamExists = true;
3352	break;
3353	}
3354
3355	/* next */
3356	pRamPrev = pRam;
3357	pRam = pRam->pNextR3;
3358	}
3359	Log(("PGMR3PhysMmio2Map: %RGp-%RGp fRamExists=%RTbool %s\n", GCPhys, GCPhysLast, fRamExists, pFirstMmio->RamRange.pszDesc));
3360
3361
3362	/*
3363	* Make the changes.
3364	*/
3365	RTGCPHYS GCPhysCur = GCPhys;
3366	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3367	{
3368	pCurMmio->RamRange.GCPhys = GCPhysCur;
3369	pCurMmio->RamRange.GCPhysLast = GCPhysCur + pCurMmio->RamRange.cb - 1;
3370	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3371	{
3372	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3373	break;
3374	}
3375	GCPhysCur += pCurMmio->RamRange.cb;
3376	}
3377
3378	if (fRamExists)
3379	{
3380	/*
3381	* Make all the pages in the range MMIO/ZERO pages, freeing any
3382	* RAM pages currently mapped here. This might not be 100% correct
3383	* for PCI memory, but we're doing the same thing for MMIO2 pages.
3384	*
3385	* We replace this MMIO/ZERO pages with real pages in the MMIO2 case.
3386	*/
3387	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK); /* Only one chunk */
3388
3389	int rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, PGMPAGETYPE_MMIO);
3390	AssertRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
3391
3392	if (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MMIO2)
3393	{
3394	/* replace the pages, freeing all present RAM pages. */
3395	PPGMPAGE pPageSrc = &pFirstMmio->RamRange.aPages[0];
3396	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3397	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3398	while (cPagesLeft-- > 0)
3399	{
3400	Assert(PGM_PAGE_IS_MMIO(pPageDst));
3401
3402	RTHCPHYS const HCPhys = PGM_PAGE_GET_HCPHYS(pPageSrc);
3403	uint32_t const idPage = PGM_PAGE_GET_PAGEID(pPageSrc);
3404	PGM_PAGE_SET_PAGEID(pVM, pPageDst, idPage);
3405	PGM_PAGE_SET_HCPHYS(pVM, pPageDst, HCPhys);
3406	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO2);
3407	PGM_PAGE_SET_STATE(pVM, pPageDst, PGM_PAGE_STATE_ALLOCATED);
3408	PGM_PAGE_SET_PDE_TYPE(pVM, pPageDst, PGM_PAGE_PDE_TYPE_DONTCARE);
3409	PGM_PAGE_SET_PTE_INDEX(pVM, pPageDst, 0);
3410	PGM_PAGE_SET_TRACKING(pVM, pPageDst, 0);
3411	/* (We tell NEM at the end of the function.) */
3412
3413	pVM->pgm.s.cZeroPages--;
3414	GCPhys += PAGE_SIZE;
3415	pPageSrc++;
3416	pPageDst++;
3417	}
3418	}
3419
3420	/* Flush physical page map TLB. */
3421	pgmPhysInvalidatePageMapTLB(pVM);
3422
3423	/* Force a PGM pool flush as guest ram references have been changed. */
3424	/** @todo not entirely SMP safe; assuming for now the guest takes care of
3425	* this internally (not touch mapped mmio while changing the mapping). */
3426	PVMCPU pVCpu = VMMGetCpu(pVM);
3427	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3428	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3429	}
3430	else
3431	{
3432	/*
3433	* No RAM range, insert the ones prepared during registration.
3434	*/
3435	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3436	{
3437	/* Clear the tracking data of pages we're going to reactivate. */
3438	PPGMPAGE pPageSrc = &pCurMmio->RamRange.aPages[0];
3439	uint32_t cPagesLeft = pCurMmio->RamRange.cb >> PAGE_SHIFT;
3440	while (cPagesLeft-- > 0)
3441	{
3442	PGM_PAGE_SET_TRACKING(pVM, pPageSrc, 0);
3443	PGM_PAGE_SET_PTE_INDEX(pVM, pPageSrc, 0);
3444	pPageSrc++;
3445	}
3446
3447	/* link in the ram range */
3448	pgmR3PhysLinkRamRange(pVM, &pCurMmio->RamRange, pRamPrev);
3449
3450	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3451	{
3452	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3453	break;
3454	}
3455	pRamPrev = &pCurMmio->RamRange;
3456	}
3457	}
3458
3459	/*
3460	* Register the access handler if plain MMIO.
3461	*
3462	* We must register access handlers for each range since the access handler
3463	* code refuses to deal with multiple ranges (and we can).
3464	*/
3465	if (!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MMIO2))
3466	{
3467	AssertFailed();
3468	int rc = VINF_SUCCESS;
3469	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3470	{
3471	Assert(!(pCurMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED));
3472	rc = pgmHandlerPhysicalExRegister(pVM, pCurMmio->pPhysHandlerR3, pCurMmio->RamRange.GCPhys,
3473	pCurMmio->RamRange.GCPhysLast);
3474	if (RT_FAILURE(rc))
3475	break;
3476	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_MAPPED; /* Use this to mark that the handler is registered. */
3477	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3478	break;
3479	}
3480	if (RT_FAILURE(rc))
3481	{
3482	/* Almost impossible, but try clean up properly and get out of here. */
3483	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3484	{
3485	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3486	{
3487	pCurMmio->fFlags &= ~PGMREGMMIO2RANGE_F_MAPPED;
3488	pgmHandlerPhysicalExDeregister(pVM, pCurMmio->pPhysHandlerR3, fRamExists);
3489	}
3490
3491	if (!fRamExists)
3492	pgmR3PhysUnlinkRamRange(pVM, &pCurMmio->RamRange);
3493	else
3494	{
3495	Assert(pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK); /* Only one chunk */
3496
3497	uint32_t cPagesLeft = pCurMmio->RamRange.cb >> PAGE_SHIFT;
3498	PPGMPAGE pPageDst = &pRam->aPages[(pCurMmio->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3499	while (cPagesLeft-- > 0)
3500	{
3501	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
3502	pPageDst++;
3503	}
3504	}
3505
3506	pCurMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3507	pCurMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3508	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3509	break;
3510	}
3511
3512	PGM_UNLOCK(pVM);
3513	return rc;
3514	}
3515	}
3516
3517	/*
3518	* We're good, set the flags and invalid the mapping TLB.
3519	*/
3520	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3521	{
3522	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_MAPPED;
3523	if (fRamExists)
3524	pCurMmio->fFlags \|= PGMREGMMIO2RANGE_F_OVERLAPPING;
3525	else
3526	pCurMmio->fFlags &= ~PGMREGMMIO2RANGE_F_OVERLAPPING;
3527	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3528	break;
3529	}
3530	pgmPhysInvalidatePageMapTLB(pVM);
3531
3532	/*
3533	* Notify NEM while holding the lock (experimental) and REM without (like always).
3534	*/
3535	uint32_t const fNemNotify = (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MMIO2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0)
3536	\| (pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_OVERLAPPING ? NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE : 0);
3537	int rc = NEMR3NotifyPhysMmioExMap(pVM, GCPhys, cbRange, fNemNotify, pFirstMmio->pvR3);
3538
3539	PGM_UNLOCK(pVM);
3540
3541	return rc;
3542	}
3543
3544
3545	/**
3546	* Unmaps an MMIO2 region.
3547	*
3548	* This is typically done when a guest / the bios / state loading changes the
3549	* PCI config. The replacing of base memory has the same restrictions as during
3550	* registration, of course.
3551	*/
3552	VMMR3_INT_DECL(int) PGMR3PhysMmio2Unmap(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS GCPhys)
3553	{
3554	/*
3555	* Validate input
3556	*/
3557	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3558	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3559	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3560	if (GCPhys != NIL_RTGCPHYS)
3561	{
3562	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3563	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3564	}
3565
3566	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3567	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3568	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3569
3570	int rc = PGM_LOCK(pVM);
3571	AssertRCReturn(rc, rc);
3572
3573	PPGMREGMMIO2RANGE pLastMmio = pFirstMmio;
3574	RTGCPHYS cbRange = 0;
3575	for (;;)
3576	{
3577	AssertReturnStmt(pLastMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED, PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3578	AssertReturnStmt(pLastMmio->RamRange.GCPhys == GCPhys + cbRange \|\| GCPhys == NIL_RTGCPHYS, PGM_UNLOCK(pVM), VERR_INVALID_PARAMETER);
3579	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3580	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3581	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3582	cbRange += pLastMmio->RamRange.cb;
3583	if (pLastMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3584	break;
3585	pLastMmio = pLastMmio->pNextR3;
3586	}
3587
3588	Log(("PGMR3PhysMmio2Unmap: %RGp-%RGp %s\n",
3589	pFirstMmio->RamRange.GCPhys, pLastMmio->RamRange.GCPhysLast, pFirstMmio->RamRange.pszDesc));
3590
3591	uint16_t const fOldFlags = pFirstMmio->fFlags;
3592	AssertReturnStmt(fOldFlags & PGMREGMMIO2RANGE_F_MAPPED, PGM_UNLOCK(pVM), VERR_WRONG_ORDER);
3593
3594	/*
3595	* If plain MMIO, we must deregister the handlers first.
3596	*/
3597	if (!(fOldFlags & PGMREGMMIO2RANGE_F_MMIO2))
3598	{
3599	AssertFailed();
3600
3601	PPGMREGMMIO2RANGE pCurMmio = pFirstMmio;
3602	rc = pgmHandlerPhysicalExDeregister(pVM, pFirstMmio->pPhysHandlerR3, RT_BOOL(fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING));
3603	AssertRCReturnStmt(rc, PGM_UNLOCK(pVM), rc);
3604	while (!(pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK))
3605	{
3606	pCurMmio = pCurMmio->pNextR3;
3607	rc = pgmHandlerPhysicalExDeregister(pVM, pCurMmio->pPhysHandlerR3, RT_BOOL(fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING));
3608	AssertRCReturnStmt(rc, PGM_UNLOCK(pVM), VERR_PGM_PHYS_MMIO_EX_IPE);
3609	}
3610	}
3611
3612	/*
3613	* Unmap it.
3614	*/
3615	RTGCPHYS const GCPhysRangeNotify = pFirstMmio->RamRange.GCPhys;
3616	if (fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING)
3617	{
3618	/*
3619	* We've replaced RAM, replace with zero pages.
3620	*
3621	* Note! This is where we might differ a little from a real system, because
3622	* it's likely to just show the RAM pages as they were before the
3623	* MMIO/MMIO2 region was mapped here.
3624	*/
3625	/* Only one chunk allowed when overlapping! */
3626	Assert(fOldFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK);
3627
3628	/* Restore the RAM pages we've replaced. */
3629	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3630	while (pRam->GCPhys > pFirstMmio->RamRange.GCPhysLast)
3631	pRam = pRam->pNextR3;
3632
3633	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3634	if (fOldFlags & PGMREGMMIO2RANGE_F_MMIO2)
3635	pVM->pgm.s.cZeroPages += cPagesLeft;
3636
3637	PPGMPAGE pPageDst = &pRam->aPages[(pFirstMmio->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3638	while (cPagesLeft-- > 0)
3639	{
3640	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
3641	pPageDst++;
3642	}
3643
3644	/* Flush physical page map TLB. */
3645	pgmPhysInvalidatePageMapTLB(pVM);
3646
3647	/* Update range state. */
3648	pFirstMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3649	pFirstMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3650	pFirstMmio->fFlags &= ~(PGMREGMMIO2RANGE_F_OVERLAPPING \| PGMREGMMIO2RANGE_F_MAPPED);
3651	}
3652	else
3653	{
3654	/*
3655	* Unlink the chunks related to the MMIO/MMIO2 region.
3656	*/
3657	for (PPGMREGMMIO2RANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3658	{
3659	pgmR3PhysUnlinkRamRange(pVM, &pCurMmio->RamRange);
3660	pCurMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3661	pCurMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3662	pCurMmio->fFlags &= ~(PGMREGMMIO2RANGE_F_OVERLAPPING \| PGMREGMMIO2RANGE_F_MAPPED);
3663	if (pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK)
3664	break;
3665	}
3666	}
3667
3668	/* Force a PGM pool flush as guest ram references have been changed. */
3669	/** @todo not entirely SMP safe; assuming for now the guest takes care
3670	* of this internally (not touch mapped mmio while changing the
3671	* mapping). */
3672	PVMCPU pVCpu = VMMGetCpu(pVM);
3673	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3674	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3675
3676	pgmPhysInvalidatePageMapTLB(pVM);
3677	pgmPhysInvalidRamRangeTlbs(pVM);
3678
3679	/*
3680	* Notify NEM while holding the lock (experimental) and REM without (like always).
3681	*/
3682	uint32_t const fNemFlags = (fOldFlags & PGMREGMMIO2RANGE_F_MMIO2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0)
3683	\| (fOldFlags & PGMREGMMIO2RANGE_F_OVERLAPPING ? NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE : 0);
3684	rc = NEMR3NotifyPhysMmioExUnmap(pVM, GCPhysRangeNotify, cbRange, fNemFlags);
3685
3686	PGM_UNLOCK(pVM);
3687	return rc;
3688	}
3689
3690
3691	/**
3692	* Reduces the mapping size of a MMIO2 region.
3693	*
3694	* This is mainly for dealing with old saved states after changing the default
3695	* size of a mapping region. See PGMDevHlpMMIOExReduce and
3696	* PDMPCIDEV::pfnRegionLoadChangeHookR3.
3697	*
3698	* The region must not currently be mapped when making this call. The VM state
3699	* must be state restore or VM construction.
3700	*
3701	* @returns VBox status code.
3702	* @param pVM The cross context VM structure.
3703	* @param pDevIns The device instance owning the region.
3704	* @param hMmio2 The handle of the region to reduce.
3705	* @param cbRegion The new mapping size.
3706	*/
3707	VMMR3_INT_DECL(int) PGMR3PhysMmio2Reduce(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, RTGCPHYS cbRegion)
3708	{
3709	/*
3710	* Validate input
3711	*/
3712	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3713	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3714	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3715	AssertReturn(cbRegion >= X86_PAGE_SIZE, VERR_INVALID_PARAMETER);
3716	AssertReturn(!(cbRegion & X86_PAGE_OFFSET_MASK), VERR_UNSUPPORTED_ALIGNMENT);
3717	VMSTATE enmVmState = VMR3GetState(pVM);
3718	AssertLogRelMsgReturn( enmVmState == VMSTATE_CREATING
3719	\|\| enmVmState == VMSTATE_LOADING,
3720	("enmVmState=%d (%s)\n", enmVmState, VMR3GetStateName(enmVmState)),
3721	VERR_VM_INVALID_VM_STATE);
3722
3723	int rc = PGM_LOCK(pVM);
3724	AssertRCReturn(rc, rc);
3725
3726	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3727	if (pFirstMmio)
3728	{
3729	Assert(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK);
3730	if (!(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED))
3731	{
3732	/*
3733	* NOTE! Current implementation does not support multiple ranges.
3734	* Implement when there is a real world need and thus a testcase.
3735	*/
3736	AssertLogRelMsgStmt(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK,
3737	("%s: %#x\n", pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3738	rc = VERR_NOT_SUPPORTED);
3739	if (RT_SUCCESS(rc))
3740	{
3741	/*
3742	* Make the change.
3743	*/
3744	Log(("PGMR3PhysMmio2Reduce: %s changes from %RGp bytes (%RGp) to %RGp bytes.\n",
3745	pFirstMmio->RamRange.pszDesc, pFirstMmio->RamRange.cb, pFirstMmio->cbReal, cbRegion));
3746
3747	AssertLogRelMsgStmt(cbRegion <= pFirstMmio->cbReal,
3748	("%s: cbRegion=%#RGp cbReal=%#RGp\n", pFirstMmio->RamRange.pszDesc, cbRegion, pFirstMmio->cbReal),
3749	rc = VERR_OUT_OF_RANGE);
3750	if (RT_SUCCESS(rc))
3751	{
3752	pFirstMmio->RamRange.cb = cbRegion;
3753	}
3754	}
3755	}
3756	else
3757	rc = VERR_WRONG_ORDER;
3758	}
3759	else
3760	rc = VERR_NOT_FOUND;
3761
3762	PGM_UNLOCK(pVM);
3763	return rc;
3764	}
3765
3766
3767	/**
3768	* Validates @a hMmio2, making sure it belongs to @a pDevIns.
3769	*
3770	* @returns VBox status code.
3771	* @param pVM The cross context VM structure.
3772	* @param pDevIns The device which allegedly owns @a hMmio2.
3773	* @param hMmio2 The handle to validate.
3774	*/
3775	VMMR3_INT_DECL(int) PGMR3PhysMmio2ValidateHandle(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3776	{
3777	/*
3778	* Validate input
3779	*/
3780	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3781	AssertPtrReturn(pDevIns, VERR_INVALID_POINTER);
3782
3783	/*
3784	* Just do this the simple way. No need for locking as this is only taken at
3785	*/
3786	PGM_LOCK_VOID(pVM);
3787	PPGMREGMMIO2RANGE pFirstMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3788	PGM_UNLOCK(pVM);
3789	AssertReturn(pFirstMmio, VERR_INVALID_HANDLE);
3790	AssertReturn(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_MMIO2, VERR_INVALID_HANDLE);
3791	AssertReturn(pFirstMmio->fFlags & PGMREGMMIO2RANGE_F_FIRST_CHUNK, VERR_INVALID_HANDLE);
3792	return VINF_SUCCESS;
3793	}
3794
3795
3796	#ifndef PGM_WITHOUT_MAPPINGS
3797	/**
3798	* Gets the HC physical address of a page in the MMIO2 region.
3799	*
3800	* This is API is intended for MMHyper and shouldn't be called
3801	* by anyone else...
3802	*
3803	* @returns VBox status code.
3804	* @param pVM The cross context VM structure.
3805	* @param pDevIns The owner of the memory, optional.
3806	* @param iSubDev Sub-device number.
3807	* @param iRegion The region.
3808	* @param off The page expressed an offset into the MMIO2 region.
3809	* @param pHCPhys Where to store the result.
3810	*/
3811	VMMR3_INT_DECL(int) PGMR3PhysMMIO2GetHCPhys(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion,
3812	RTGCPHYS off, PRTHCPHYS pHCPhys)
3813	{
3814	/*
3815	* Validate input
3816	*/
3817	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3818	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3819	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3820	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3821
3822	PGM_LOCK_VOID(pVM);
3823	PPGMREGMMIO2RANGE pCurMmio = pgmR3PhysMmio2Find(pVM, pDevIns, iSubDev, iRegion, NIL_PGMMMIO2HANDLE);
3824	AssertReturn(pCurMmio, VERR_NOT_FOUND);
3825	AssertReturn(pCurMmio->fFlags & (PGMREGMMIO2RANGE_F_MMIO2 \| PGMREGMMIO2RANGE_F_FIRST_CHUNK), VERR_WRONG_TYPE);
3826
3827	while ( off >= pCurMmio->RamRange.cb
3828	&& !(pCurMmio->fFlags & PGMREGMMIO2RANGE_F_LAST_CHUNK))
3829	{
3830	off -= pCurMmio->RamRange.cb;
3831	pCurMmio = pCurMmio->pNextR3;
3832	}
3833	AssertReturn(off < pCurMmio->RamRange.cb, VERR_INVALID_PARAMETER);
3834
3835	PCPGMPAGE pPage = &pCurMmio->RamRange.aPages[off >> PAGE_SHIFT];
3836	*pHCPhys = PGM_PAGE_GET_HCPHYS(pPage);
3837	PGM_UNLOCK(pVM);
3838	return VINF_SUCCESS;
3839	}
3840	#endif /* !PGM_WITHOUT_MAPPINGS */
3841
3842
3843	/**
3844	* Gets the mapping address of an MMIO2 region.
3845	*
3846	* @returns Mapping address, NIL_RTGCPHYS if not mapped or invalid handle.
3847	*
3848	* @param pVM The cross context VM structure.
3849	* @param pDevIns The device owning the MMIO2 handle.
3850	* @param hMmio2 The region handle.
3851	*/
3852	VMMR3_INT_DECL(RTGCPHYS) PGMR3PhysMmio2GetMappingAddress(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2)
3853	{
3854	AssertPtrReturn(pDevIns, NIL_RTGCPHYS);
3855
3856	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3857	AssertReturn(pFirstRegMmio, NIL_RTGCPHYS);
3858
3859	if (pFirstRegMmio->fFlags & PGMREGMMIO2RANGE_F_MAPPED)
3860	return pFirstRegMmio->RamRange.GCPhys;
3861	return NIL_RTGCPHYS;
3862	}
3863
3864	/**
3865	* Changes the region number of an MMIO2 region.
3866	*
3867	* This is only for dealing with save state issues, nothing else.
3868	*
3869	* @return VBox status code.
3870	*
3871	* @param pVM The cross context VM structure.
3872	* @param pDevIns The device owning the MMIO2 memory.
3873	* @param hMmio2 The handle of the region.
3874	* @param iNewRegion The new region index.
3875	*
3876	* @thread EMT(0)
3877	* @sa @bugref{9359}
3878	*/
3879	VMMR3_INT_DECL(int) PGMR3PhysMmio2ChangeRegionNo(PVM pVM, PPDMDEVINS pDevIns, PGMMMIO2HANDLE hMmio2, uint32_t iNewRegion)
3880	{
3881	/*
3882	* Validate input.
3883	*/
3884	VM_ASSERT_EMT0_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3885	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_LOADING, VERR_VM_INVALID_VM_STATE);
3886	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3887	AssertReturn(hMmio2 != NIL_PGMMMIO2HANDLE, VERR_INVALID_HANDLE);
3888	AssertReturn(iNewRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3889
3890	AssertReturn(pVM->enmVMState == VMSTATE_LOADING, VERR_INVALID_STATE);
3891
3892	int rc = PGM_LOCK(pVM);
3893	AssertRCReturn(rc, rc);
3894
3895	PPGMREGMMIO2RANGE pFirstRegMmio = pgmR3PhysMmio2Find(pVM, pDevIns, UINT32_MAX, UINT32_MAX, hMmio2);
3896	AssertReturnStmt(pFirstRegMmio, PGM_UNLOCK(pVM), VERR_NOT_FOUND);
3897	AssertReturnStmt(pgmR3PhysMmio2Find(pVM, pDevIns, pFirstRegMmio->iSubDev, iNewRegion, NIL_PGMMMIO2HANDLE) == NULL,
3898	PGM_UNLOCK(pVM), VERR_RESOURCE_IN_USE);
3899
3900	/*
3901	* Make the change.
3902	*/
3903	pFirstRegMmio->iRegion = (uint8_t)iNewRegion;
3904
3905	PGM_UNLOCK(pVM);
3906	return VINF_SUCCESS;
3907	}
3908
3909
3910	/**
3911	* Worker for PGMR3PhysRomRegister.
3912	*
3913	* This is here to simplify lock management, i.e. the caller does all the
3914	* locking and we can simply return without needing to remember to unlock
3915	* anything first.
3916	*
3917	* @returns VBox status code.
3918	* @param pVM The cross context VM structure.
3919	* @param pDevIns The device instance owning the ROM.
3920	* @param GCPhys First physical address in the range.
3921	* Must be page aligned!
3922	* @param cb The size of the range (in bytes).
3923	* Must be page aligned!
3924	* @param pvBinary Pointer to the binary data backing the ROM image.
3925	* @param cbBinary The size of the binary data pvBinary points to.
3926	* This must be less or equal to @a cb.
3927	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
3928	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
3929	* @param pszDesc Pointer to description string. This must not be freed.
3930	*/
3931	static int pgmR3PhysRomRegisterLocked(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
3932	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
3933	{
3934	/*
3935	* Validate input.
3936	*/
3937	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3938	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
3939	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
3940	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
3941	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3942	AssertPtrReturn(pvBinary, VERR_INVALID_PARAMETER);
3943	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
3944	AssertReturn(!(fFlags & ~PGMPHYS_ROM_FLAGS_VALID_MASK), VERR_INVALID_PARAMETER);
3945	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_CREATING, VERR_VM_INVALID_VM_STATE);
3946
3947	const uint32_t cPages = cb >> PAGE_SHIFT;
3948
3949	/*
3950	* Find the ROM location in the ROM list first.
3951	*/
3952	PPGMROMRANGE pRomPrev = NULL;
3953	PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3;
3954	while (pRom && GCPhysLast >= pRom->GCPhys)
3955	{
3956	if ( GCPhys <= pRom->GCPhysLast
3957	&& GCPhysLast >= pRom->GCPhys)
3958	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
3959	GCPhys, GCPhysLast, pszDesc,
3960	pRom->GCPhys, pRom->GCPhysLast, pRom->pszDesc),
3961	VERR_PGM_RAM_CONFLICT);
3962	/* next */
3963	pRomPrev = pRom;
3964	pRom = pRom->pNextR3;
3965	}
3966
3967	/*
3968	* Find the RAM location and check for conflicts.
3969	*
3970	* Conflict detection is a bit different than for RAM
3971	* registration since a ROM can be located within a RAM
3972	* range. So, what we have to check for is other memory
3973	* types (other than RAM that is) and that we don't span
3974	* more than one RAM range (layz).
3975	*/
3976	bool fRamExists = false;
3977	PPGMRAMRANGE pRamPrev = NULL;
3978	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3979	while (pRam && GCPhysLast >= pRam->GCPhys)
3980	{
3981	if ( GCPhys <= pRam->GCPhysLast
3982	&& GCPhysLast >= pRam->GCPhys)
3983	{
3984	/* completely within? */
3985	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
3986	&& GCPhysLast <= pRam->GCPhysLast,
3987	("%RGp-%RGp (%s) falls partly outside %RGp-%RGp (%s)\n",
3988	GCPhys, GCPhysLast, pszDesc,
3989	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3990	VERR_PGM_RAM_CONFLICT);
3991	fRamExists = true;
3992	break;
3993	}
3994
3995	/* next */
3996	pRamPrev = pRam;
3997	pRam = pRam->pNextR3;
3998	}
3999	if (fRamExists)
4000	{
4001	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
4002	uint32_t cPagesLeft = cPages;
4003	while (cPagesLeft-- > 0)
4004	{
4005	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
4006	("%RGp (%R[pgmpage]) isn't a RAM page - registering %RGp-%RGp (%s).\n",
4007	pRam->GCPhys + ((RTGCPHYS)(uintptr_t)(pPage - &pRam->aPages[0]) << PAGE_SHIFT),
4008	pPage, GCPhys, GCPhysLast, pszDesc), VERR_PGM_RAM_CONFLICT);
4009	Assert(PGM_PAGE_IS_ZERO(pPage));
4010	pPage++;
4011	}
4012	}
4013
4014	/*
4015	* Update the base memory reservation if necessary.
4016	*/
4017	uint32_t cExtraBaseCost = fRamExists ? 0 : cPages;
4018	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4019	cExtraBaseCost += cPages;
4020	if (cExtraBaseCost)
4021	{
4022	int rc = MMR3IncreaseBaseReservation(pVM, cExtraBaseCost);
4023	if (RT_FAILURE(rc))
4024	return rc;
4025	}
4026
4027	/*
4028	* Allocate memory for the virgin copy of the RAM.
4029	*/
4030	PGMMALLOCATEPAGESREQ pReq;
4031	int rc = GMMR3AllocatePagesPrepare(pVM, &pReq, cPages, GMMACCOUNT_BASE);
4032	AssertRCReturn(rc, rc);
4033
4034	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4035	{
4036	pReq->aPages[iPage].HCPhysGCPhys = GCPhys + (iPage << PAGE_SHIFT);
4037	pReq->aPages[iPage].idPage = NIL_GMM_PAGEID;
4038	pReq->aPages[iPage].idSharedPage = NIL_GMM_PAGEID;
4039	}
4040
4041	rc = GMMR3AllocatePagesPerform(pVM, pReq);
4042	if (RT_FAILURE(rc))
4043	{
4044	GMMR3AllocatePagesCleanup(pReq);
4045	return rc;
4046	}
4047
4048	/*
4049	* Allocate the new ROM range and RAM range (if necessary).
4050	*/
4051	PPGMROMRANGE pRomNew;
4052	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMROMRANGE, aPages[cPages]), 0, MM_TAG_PGM_PHYS, (void **)&pRomNew);
4053	if (RT_SUCCESS(rc))
4054	{
4055	PPGMRAMRANGE pRamNew = NULL;
4056	if (!fRamExists)
4057	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), sizeof(PGMPAGE), MM_TAG_PGM_PHYS, (void **)&pRamNew);
4058	if (RT_SUCCESS(rc))
4059	{
4060	/*
4061	* Initialize and insert the RAM range (if required).
4062	*/
4063	PPGMROMPAGE pRomPage = &pRomNew->aPages[0];
4064	if (!fRamExists)
4065	{
4066	pRamNew->pSelfR0 = MMHyperCCToR0(pVM, pRamNew);
4067	pRamNew->GCPhys = GCPhys;
4068	pRamNew->GCPhysLast = GCPhysLast;
4069	pRamNew->cb = cb;
4070	pRamNew->pszDesc = pszDesc;
4071	pRamNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_ROM;
4072	pRamNew->pvR3 = NULL;
4073	pRamNew->paLSPages = NULL;
4074
4075	PPGMPAGE pPage = &pRamNew->aPages[0];
4076	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
4077	{
4078	PGM_PAGE_INIT(pPage,
4079	pReq->aPages[iPage].HCPhysGCPhys,
4080	pReq->aPages[iPage].idPage,
4081	PGMPAGETYPE_ROM,
4082	PGM_PAGE_STATE_ALLOCATED);
4083
4084	pRomPage->Virgin = *pPage;
4085	}
4086
4087	pVM->pgm.s.cAllPages += cPages;
4088	pgmR3PhysLinkRamRange(pVM, pRamNew, pRamPrev);
4089	}
4090	else
4091	{
4092	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
4093	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
4094	{
4095	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_ROM);
4096	PGM_PAGE_SET_HCPHYS(pVM, pPage, pReq->aPages[iPage].HCPhysGCPhys);
4097	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
4098	PGM_PAGE_SET_PAGEID(pVM, pPage, pReq->aPages[iPage].idPage);
4099	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4100	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
4101	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
4102
4103	pRomPage->Virgin = *pPage;
4104	}
4105
4106	pRamNew = pRam;
4107
4108	pVM->pgm.s.cZeroPages -= cPages;
4109	}
4110	pVM->pgm.s.cPrivatePages += cPages;
4111
4112	/* Flush physical page map TLB. */
4113	pgmPhysInvalidatePageMapTLB(pVM);
4114
4115
4116	/* Notify NEM before we register handlers. */
4117	uint32_t const fNemNotify = (fRamExists ? NEM_NOTIFY_PHYS_ROM_F_REPLACE : 0)
4118	\| (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED ? NEM_NOTIFY_PHYS_ROM_F_SHADOW : 0);
4119	rc = NEMR3NotifyPhysRomRegisterEarly(pVM, GCPhys, cb, fNemNotify);
4120
4121	/* Register the ROM access handler. */
4122	if (RT_SUCCESS(rc))
4123	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, pVM->pgm.s.hRomPhysHandlerType,
4124	pRomNew, MMHyperCCToR0(pVM, pRomNew), MMHyperCCToRC(pVM, pRomNew),
4125	pszDesc);
4126	if (RT_SUCCESS(rc))
4127	{
4128	/*
4129	* Copy the image over to the virgin pages.
4130	* This must be done after linking in the RAM range.
4131	*/
4132	size_t cbBinaryLeft = cbBinary;
4133	PPGMPAGE pRamPage = &pRamNew->aPages[(GCPhys - pRamNew->GCPhys) >> PAGE_SHIFT];
4134	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
4135	{
4136	void *pvDstPage;
4137	rc = pgmPhysPageMap(pVM, pRamPage, GCPhys + (iPage << PAGE_SHIFT), &pvDstPage);
4138	if (RT_FAILURE(rc))
4139	{
4140	VMSetError(pVM, rc, RT_SRC_POS, "Failed to map virgin ROM page at %RGp", GCPhys);
4141	break;
4142	}
4143	if (cbBinaryLeft >= PAGE_SIZE)
4144	{
4145	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), PAGE_SIZE);
4146	cbBinaryLeft -= PAGE_SIZE;
4147	}
4148	else
4149	{
4150	ASMMemZeroPage(pvDstPage); /* (shouldn't be necessary, but can't hurt either) */
4151	if (cbBinaryLeft > 0)
4152	{
4153	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), cbBinaryLeft);
4154	cbBinaryLeft = 0;
4155	}
4156	}
4157	}
4158	if (RT_SUCCESS(rc))
4159	{
4160	/*
4161	* Initialize the ROM range.
4162	* Note that the Virgin member of the pages has already been initialized above.
4163	*/
4164	pRomNew->GCPhys = GCPhys;
4165	pRomNew->GCPhysLast = GCPhysLast;
4166	pRomNew->cb = cb;
4167	pRomNew->fFlags = fFlags;
4168	pRomNew->idSavedState = UINT8_MAX;
4169	pRomNew->cbOriginal = cbBinary;
4170	pRomNew->pszDesc = pszDesc;
4171	pRomNew->pvOriginal = fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY
4172	? pvBinary : RTMemDup(pvBinary, cbBinary);
4173	if (pRomNew->pvOriginal)
4174	{
4175	for (unsigned iPage = 0; iPage < cPages; iPage++)
4176	{
4177	PPGMROMPAGE pPage = &pRomNew->aPages[iPage];
4178	pPage->enmProt = PGMROMPROT_READ_ROM_WRITE_IGNORE;
4179	PGM_PAGE_INIT_ZERO(&pPage->Shadow, pVM, PGMPAGETYPE_ROM_SHADOW);
4180	}
4181
4182	/* update the page count stats for the shadow pages. */
4183	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4184	{
4185	pVM->pgm.s.cZeroPages += cPages;
4186	pVM->pgm.s.cAllPages += cPages;
4187	}
4188
4189	/*
4190	* Insert the ROM range, tell REM and return successfully.
4191	*/
4192	pRomNew->pNextR3 = pRom;
4193	pRomNew->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4194
4195	if (pRomPrev)
4196	{
4197	pRomPrev->pNextR3 = pRomNew;
4198	pRomPrev->pNextR0 = MMHyperCCToR0(pVM, pRomNew);
4199	}
4200	else
4201	{
4202	pVM->pgm.s.pRomRangesR3 = pRomNew;
4203	pVM->pgm.s.pRomRangesR0 = MMHyperCCToR0(pVM, pRomNew);
4204	}
4205
4206	pgmPhysInvalidatePageMapTLB(pVM);
4207	GMMR3AllocatePagesCleanup(pReq);
4208
4209	/* Notify NEM again. */
4210	return NEMR3NotifyPhysRomRegisterLate(pVM, GCPhys, cb, fNemNotify);
4211	}
4212
4213	/* bail out */
4214	rc = VERR_NO_MEMORY;
4215	}
4216
4217	int rc2 = PGMHandlerPhysicalDeregister(pVM, GCPhys);
4218	AssertRC(rc2);
4219	}
4220
4221	if (!fRamExists)
4222	{
4223	pgmR3PhysUnlinkRamRange2(pVM, pRamNew, pRamPrev);
4224	MMHyperFree(pVM, pRamNew);
4225	}
4226	}
4227	MMHyperFree(pVM, pRomNew);
4228	}
4229
4230	/** @todo Purge the mapping cache or something... */
4231	GMMR3FreeAllocatedPages(pVM, pReq);
4232	GMMR3AllocatePagesCleanup(pReq);
4233	return rc;
4234	}
4235
4236
4237	/**
4238	* Registers a ROM image.
4239	*
4240	* Shadowed ROM images requires double the amount of backing memory, so,
4241	* don't use that unless you have to. Shadowing of ROM images is process
4242	* where we can select where the reads go and where the writes go. On real
4243	* hardware the chipset provides means to configure this. We provide
4244	* PGMR3PhysProtectROM() for this purpose.
4245	*
4246	* A read-only copy of the ROM image will always be kept around while we
4247	* will allocate RAM pages for the changes on demand (unless all memory
4248	* is configured to be preallocated).
4249	*
4250	* @returns VBox status code.
4251	* @param pVM The cross context VM structure.
4252	* @param pDevIns The device instance owning the ROM.
4253	* @param GCPhys First physical address in the range.
4254	* Must be page aligned!
4255	* @param cb The size of the range (in bytes).
4256	* Must be page aligned!
4257	* @param pvBinary Pointer to the binary data backing the ROM image.
4258	* @param cbBinary The size of the binary data pvBinary points to.
4259	* This must be less or equal to @a cb.
4260	* @param fFlags Mask of flags, PGMPHYS_ROM_FLAGS_XXX.
4261	* @param pszDesc Pointer to description string. This must not be freed.
4262	*
4263	* @remark There is no way to remove the rom, automatically on device cleanup or
4264	* manually from the device yet. This isn't difficult in any way, it's
4265	* just not something we expect to be necessary for a while.
4266	*/
4267	VMMR3DECL(int) PGMR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
4268	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
4269	{
4270	Log(("PGMR3PhysRomRegister: pDevIns=%p GCPhys=%RGp(-%RGp) cb=%RGp pvBinary=%p cbBinary=%#x fFlags=%#x pszDesc=%s\n",
4271	pDevIns, GCPhys, GCPhys + cb, cb, pvBinary, cbBinary, fFlags, pszDesc));
4272	PGM_LOCK_VOID(pVM);
4273	int rc = pgmR3PhysRomRegisterLocked(pVM, pDevIns, GCPhys, cb, pvBinary, cbBinary, fFlags, pszDesc);
4274	PGM_UNLOCK(pVM);
4275	return rc;
4276	}
4277
4278
4279	/**
4280	* Called by PGMR3MemSetup to reset the shadow, switch to the virgin, and verify
4281	* that the virgin part is untouched.
4282	*
4283	* This is done after the normal memory has been cleared.
4284	*
4285	* ASSUMES that the caller owns the PGM lock.
4286	*
4287	* @param pVM The cross context VM structure.
4288	*/
4289	int pgmR3PhysRomReset(PVM pVM)
4290	{
4291	PGM_LOCK_ASSERT_OWNER(pVM);
4292	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4293	{
4294	const uint32_t cPages = pRom->cb >> PAGE_SHIFT;
4295
4296	if (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4297	{
4298	/*
4299	* Reset the physical handler.
4300	*/
4301	int rc = PGMR3PhysRomProtect(pVM, pRom->GCPhys, pRom->cb, PGMROMPROT_READ_ROM_WRITE_IGNORE);
4302	AssertRCReturn(rc, rc);
4303
4304	/*
4305	* What we do with the shadow pages depends on the memory
4306	* preallocation option. If not enabled, we'll just throw
4307	* out all the dirty pages and replace them by the zero page.
4308	*/
4309	if (!pVM->pgm.s.fRamPreAlloc)
4310	{
4311	/* Free the dirty pages. */
4312	uint32_t cPendingPages = 0;
4313	PGMMFREEPAGESREQ pReq;
4314	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
4315	AssertRCReturn(rc, rc);
4316
4317	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4318	if ( !PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow)
4319	&& !PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow))
4320	{
4321	Assert(PGM_PAGE_GET_STATE(&pRom->aPages[iPage].Shadow) == PGM_PAGE_STATE_ALLOCATED);
4322	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, &pRom->aPages[iPage].Shadow,
4323	pRom->GCPhys + (iPage << PAGE_SHIFT),
4324	(PGMPAGETYPE)PGM_PAGE_GET_TYPE(&pRom->aPages[iPage].Shadow));
4325	AssertLogRelRCReturn(rc, rc);
4326	}
4327
4328	if (cPendingPages)
4329	{
4330	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
4331	AssertLogRelRCReturn(rc, rc);
4332	}
4333	GMMR3FreePagesCleanup(pReq);
4334	}
4335	else
4336	{
4337	/* clear all the shadow pages. */
4338	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4339	{
4340	if (PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow))
4341	continue;
4342	Assert(!PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow));
4343	void *pvDstPage;
4344	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4345	rc = pgmPhysPageMakeWritableAndMap(pVM, &pRom->aPages[iPage].Shadow, GCPhys, &pvDstPage);
4346	if (RT_FAILURE(rc))
4347	break;
4348	ASMMemZeroPage(pvDstPage);
4349	}
4350	AssertRCReturn(rc, rc);
4351	}
4352	}
4353
4354	/*
4355	* Restore the original ROM pages after a saved state load.
4356	* Also, in strict builds check that ROM pages remain unmodified.
4357	*/
4358	#ifndef VBOX_STRICT
4359	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4360	#endif
4361	{
4362	size_t cbSrcLeft = pRom->cbOriginal;
4363	uint8_t const pbSrcPage = (uint8_t const )pRom->pvOriginal;
4364	uint32_t cRestored = 0;
4365	for (uint32_t iPage = 0; iPage < cPages && cbSrcLeft > 0; iPage++, pbSrcPage += PAGE_SIZE)
4366	{
4367	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4368	void const *pvDstPage;
4369	int rc = pgmPhysPageMapReadOnly(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPage);
4370	if (RT_FAILURE(rc))
4371	break;
4372
4373	if (memcmp(pvDstPage, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE)))
4374	{
4375	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4376	{
4377	void *pvDstPageW;
4378	rc = pgmPhysPageMap(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPageW);
4379	AssertLogRelRCReturn(rc, rc);
4380	memcpy(pvDstPageW, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE));
4381	cRestored++;
4382	}
4383	else
4384	LogRel(("pgmR3PhysRomReset: %RGp: ROM page changed (%s)\n", GCPhys, pRom->pszDesc));
4385	}
4386	cbSrcLeft -= RT_MIN(cbSrcLeft, PAGE_SIZE);
4387	}
4388	if (cRestored > 0)
4389	LogRel(("PGM: ROM \"%s\": Reloaded %u of %u pages.\n", pRom->pszDesc, cRestored, cPages));
4390	}
4391	}
4392
4393	/* Clear the ROM restore flag now as we only need to do this once after
4394	loading saved state. */
4395	pVM->pgm.s.fRestoreRomPagesOnReset = false;
4396
4397	return VINF_SUCCESS;
4398	}
4399
4400
4401	/**
4402	* Called by PGMR3Term to free resources.
4403	*
4404	* ASSUMES that the caller owns the PGM lock.
4405	*
4406	* @param pVM The cross context VM structure.
4407	*/
4408	void pgmR3PhysRomTerm(PVM pVM)
4409	{
4410	/*
4411	* Free the heap copy of the original bits.
4412	*/
4413	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4414	{
4415	if ( pRom->pvOriginal
4416	&& !(pRom->fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY))
4417	{
4418	RTMemFree((void *)pRom->pvOriginal);
4419	pRom->pvOriginal = NULL;
4420	}
4421	}
4422	}
4423
4424
4425	/**
4426	* Change the shadowing of a range of ROM pages.
4427	*
4428	* This is intended for implementing chipset specific memory registers
4429	* and will not be very strict about the input. It will silently ignore
4430	* any pages that are not the part of a shadowed ROM.
4431	*
4432	* @returns VBox status code.
4433	* @retval VINF_PGM_SYNC_CR3
4434	*
4435	* @param pVM The cross context VM structure.
4436	* @param GCPhys Where to start. Page aligned.
4437	* @param cb How much to change. Page aligned.
4438	* @param enmProt The new ROM protection.
4439	*/
4440	VMMR3DECL(int) PGMR3PhysRomProtect(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMROMPROT enmProt)
4441	{
4442	/*
4443	* Check input
4444	*/
4445	if (!cb)
4446	return VINF_SUCCESS;
4447	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4448	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4449	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
4450	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
4451	AssertReturn(enmProt >= PGMROMPROT_INVALID && enmProt <= PGMROMPROT_END, VERR_INVALID_PARAMETER);
4452
4453	/*
4454	* Process the request.
4455	*/
4456	PGM_LOCK_VOID(pVM);
4457	int rc = VINF_SUCCESS;
4458	bool fFlushTLB = false;
4459	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4460	{
4461	if ( GCPhys <= pRom->GCPhysLast
4462	&& GCPhysLast >= pRom->GCPhys
4463	&& (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED))
4464	{
4465	/*
4466	* Iterate the relevant pages and make necessary the changes.
4467	*/
4468	bool fChanges = false;
4469	uint32_t const cPages = pRom->GCPhysLast <= GCPhysLast
4470	? pRom->cb >> PAGE_SHIFT
4471	: (GCPhysLast - pRom->GCPhys + 1) >> PAGE_SHIFT;
4472	for (uint32_t iPage = (GCPhys - pRom->GCPhys) >> PAGE_SHIFT;
4473	iPage < cPages;
4474	iPage++)
4475	{
4476	PPGMROMPAGE pRomPage = &pRom->aPages[iPage];
4477	if (PGMROMPROT_IS_ROM(pRomPage->enmProt) != PGMROMPROT_IS_ROM(enmProt))
4478	{
4479	fChanges = true;
4480
4481	/* flush references to the page. */
4482	PPGMPAGE pRamPage = pgmPhysGetPage(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT));
4483	int rc2 = pgmPoolTrackUpdateGCPhys(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT), pRamPage,
4484	true /fFlushPTEs/, &fFlushTLB);
4485	if (rc2 != VINF_SUCCESS && (rc == VINF_SUCCESS \|\| RT_FAILURE(rc2)))
4486	rc = rc2;
4487	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pRamPage);
4488
4489	PPGMPAGE pOld = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Virgin : &pRomPage->Shadow;
4490	PPGMPAGE pNew = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Shadow : &pRomPage->Virgin;
4491
4492	pOld = pRamPage;
4493	pRamPage = pNew;
4494	/** @todo preserve the volatile flags (handlers) when these have been moved out of HCPhys! */
4495
4496	/* Tell NEM about the backing and protection change. */
4497	if (VM_IS_NEM_ENABLED(pVM))
4498	{
4499	PGMPAGETYPE enmType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pNew);
4500	NEMHCNotifyPhysPageChanged(pVM, GCPhys, PGM_PAGE_GET_HCPHYS(pOld), PGM_PAGE_GET_HCPHYS(pNew),
4501	pgmPhysPageCalcNemProtection(pRamPage, enmType), enmType, &u2State);
4502	PGM_PAGE_SET_NEM_STATE(pRamPage, u2State);
4503	}
4504	}
4505	pRomPage->enmProt = enmProt;
4506	}
4507
4508	/*
4509	* Reset the access handler if we made changes, no need
4510	* to optimize this.
4511	*/
4512	if (fChanges)
4513	{
4514	int rc2 = PGMHandlerPhysicalReset(pVM, pRom->GCPhys);
4515	if (RT_FAILURE(rc2))
4516	{
4517	PGM_UNLOCK(pVM);
4518	AssertRC(rc);
4519	return rc2;
4520	}
4521	}
4522
4523	/* Advance - cb isn't updated. */
4524	GCPhys = pRom->GCPhys + (cPages << PAGE_SHIFT);
4525	}
4526	}
4527	PGM_UNLOCK(pVM);
4528	if (fFlushTLB)
4529	PGM_INVL_ALL_VCPU_TLBS(pVM);
4530
4531	return rc;
4532	}
4533
4534
4535	/**
4536	* Sets the Address Gate 20 state.
4537	*
4538	* @param pVCpu The cross context virtual CPU structure.
4539	* @param fEnable True if the gate should be enabled.
4540	* False if the gate should be disabled.
4541	*/
4542	VMMDECL(void) PGMR3PhysSetA20(PVMCPU pVCpu, bool fEnable)
4543	{
4544	LogFlow(("PGMR3PhysSetA20 %d (was %d)\n", fEnable, pVCpu->pgm.s.fA20Enabled));
4545	if (pVCpu->pgm.s.fA20Enabled != fEnable)
4546	{
4547	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX
4548	PCCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
4549	if ( CPUMIsGuestInVmxRootMode(pCtx)
4550	&& !fEnable)
4551	{
4552	Log(("Cannot enter A20M mode while in VMX root mode\n"));
4553	return;
4554	}
4555	#endif
4556	pVCpu->pgm.s.fA20Enabled = fEnable;
4557	pVCpu->pgm.s.GCPhysA20Mask = ~((RTGCPHYS)!fEnable << 20);
4558	NEMR3NotifySetA20(pVCpu, fEnable);
4559	#ifdef PGM_WITH_A20
4560	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
4561	pgmR3RefreshShadowModeAfterA20Change(pVCpu);
4562	HMFlushTlb(pVCpu);
4563	#endif
4564	IEMTlbInvalidateAllPhysical(pVCpu);
4565	STAM_REL_COUNTER_INC(&pVCpu->pgm.s.cA20Changes);
4566	}
4567	}
4568
4569
4570	/**
4571	* Tree enumeration callback for dealing with age rollover.
4572	* It will perform a simple compression of the current age.
4573	*/
4574	static DECLCALLBACK(int) pgmR3PhysChunkAgeingRolloverCallback(PAVLU32NODECORE pNode, void *pvUser)
4575	{
4576	/* Age compression - ASSUMES iNow == 4. */
4577	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
4578	if (pChunk->iLastUsed >= UINT32_C(0xffffff00))
4579	pChunk->iLastUsed = 3;
4580	else if (pChunk->iLastUsed >= UINT32_C(0xfffff000))
4581	pChunk->iLastUsed = 2;
4582	else if (pChunk->iLastUsed)
4583	pChunk->iLastUsed = 1;
4584	else /* iLastUsed = 0 */
4585	pChunk->iLastUsed = 4;
4586
4587	NOREF(pvUser);
4588	return 0;
4589	}
4590
4591
4592	/**
4593	* The structure passed in the pvUser argument of pgmR3PhysChunkUnmapCandidateCallback().
4594	*/
4595	typedef struct PGMR3PHYSCHUNKUNMAPCB
4596	{
4597	PVM pVM; /*< Pointer to the VM. /
4598	PPGMCHUNKR3MAP pChunk; /*< The chunk to unmap. /
4599	} PGMR3PHYSCHUNKUNMAPCB, *PPGMR3PHYSCHUNKUNMAPCB;
4600
4601
4602	/**
4603	* Callback used to find the mapping that's been unused for
4604	* the longest time.
4605	*/
4606	static DECLCALLBACK(int) pgmR3PhysChunkUnmapCandidateCallback(PAVLU32NODECORE pNode, void *pvUser)
4607	{
4608	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
4609	PPGMR3PHYSCHUNKUNMAPCB pArg = (PPGMR3PHYSCHUNKUNMAPCB)pvUser;
4610
4611	/*
4612	* Check for locks and compare when last used.
4613	*/
4614	if (pChunk->cRefs)
4615	return 0;
4616	if (pChunk->cPermRefs)
4617	return 0;
4618	if ( pArg->pChunk
4619	&& pChunk->iLastUsed >= pArg->pChunk->iLastUsed)
4620	return 0;
4621
4622	/*
4623	* Check that it's not in any of the TLBs.
4624	*/
4625	PVM pVM = pArg->pVM;
4626	if ( pVM->pgm.s.ChunkR3Map.Tlb.aEntries[PGM_CHUNKR3MAPTLB_IDX(pChunk->Core.Key)].idChunk
4627	== pChunk->Core.Key)
4628	{
4629	pChunk = NULL;
4630	return 0;
4631	}
4632	#ifdef VBOX_STRICT
4633	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
4634	{
4635	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk != pChunk);
4636	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk != pChunk->Core.Key);
4637	}
4638	#endif
4639
4640	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.PhysTlbR3.aEntries); i++)
4641	if (pVM->pgm.s.PhysTlbR3.aEntries[i].pMap == pChunk)
4642	return 0;
4643
4644	pArg->pChunk = pChunk;
4645	return 0;
4646	}
4647
4648
4649	/**
4650	* Finds a good candidate for unmapping when the ring-3 mapping cache is full.
4651	*
4652	* The candidate will not be part of any TLBs, so no need to flush
4653	* anything afterwards.
4654	*
4655	* @returns Chunk id.
4656	* @param pVM The cross context VM structure.
4657	*/
4658	static int32_t pgmR3PhysChunkFindUnmapCandidate(PVM pVM)
4659	{
4660	PGM_LOCK_ASSERT_OWNER(pVM);
4661
4662	/*
4663	* Enumerate the age tree starting with the left most node.
4664	*/
4665	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
4666	PGMR3PHYSCHUNKUNMAPCB Args;
4667	Args.pVM = pVM;
4668	Args.pChunk = NULL;
4669	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkUnmapCandidateCallback, &Args);
4670	Assert(Args.pChunk);
4671	if (Args.pChunk)
4672	{
4673	Assert(Args.pChunk->cRefs == 0);
4674	Assert(Args.pChunk->cPermRefs == 0);
4675	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
4676	return Args.pChunk->Core.Key;
4677	}
4678
4679	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkFindCandidate, a);
4680	return INT32_MAX;
4681	}
4682
4683
4684	/**
4685	* Rendezvous callback used by pgmR3PhysUnmapChunk that unmaps a chunk
4686	*
4687	* This is only called on one of the EMTs while the other ones are waiting for
4688	* it to complete this function.
4689	*
4690	* @returns VINF_SUCCESS (VBox strict status code).
4691	* @param pVM The cross context VM structure.
4692	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
4693	* @param pvUser User pointer. Unused
4694	*
4695	*/
4696	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysUnmapChunkRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
4697	{
4698	int rc = VINF_SUCCESS;
4699	PGM_LOCK_VOID(pVM);
4700	NOREF(pVCpu); NOREF(pvUser);
4701
4702	if (pVM->pgm.s.ChunkR3Map.c >= pVM->pgm.s.ChunkR3Map.cMax)
4703	{
4704	/* Flush the pgm pool cache; call the internal rendezvous handler as we're already in a rendezvous handler here. */
4705	/** @todo also not really efficient to unmap a chunk that contains PD
4706	* or PT pages. */
4707	pgmR3PoolClearAllRendezvous(pVM, pVM->apCpusR3[0], NULL /* no need to flush the REM TLB as we already did that above */);
4708
4709	/*
4710	* Request the ring-0 part to unmap a chunk to make space in the mapping cache.
4711	*/
4712	GMMMAPUNMAPCHUNKREQ Req;
4713	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
4714	Req.Hdr.cbReq = sizeof(Req);
4715	Req.pvR3 = NULL;
4716	Req.idChunkMap = NIL_GMM_CHUNKID;
4717	Req.idChunkUnmap = pgmR3PhysChunkFindUnmapCandidate(pVM);
4718	if (Req.idChunkUnmap != INT32_MAX)
4719	{
4720	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkUnmap, a);
4721	rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
4722	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkUnmap, a);
4723	if (RT_SUCCESS(rc))
4724	{
4725	/*
4726	* Remove the unmapped one.
4727	*/
4728	PPGMCHUNKR3MAP pUnmappedChunk = (PPGMCHUNKR3MAP)RTAvlU32Remove(&pVM->pgm.s.ChunkR3Map.pTree, Req.idChunkUnmap);
4729	AssertRelease(pUnmappedChunk);
4730	AssertRelease(!pUnmappedChunk->cRefs);
4731	AssertRelease(!pUnmappedChunk->cPermRefs);
4732	pUnmappedChunk->pv = NULL;
4733	pUnmappedChunk->Core.Key = UINT32_MAX;
4734	MMR3HeapFree(pUnmappedChunk);
4735	pVM->pgm.s.ChunkR3Map.c--;
4736	pVM->pgm.s.cUnmappedChunks++;
4737
4738	/*
4739	* Flush dangling PGM pointers (R3 & R0 ptrs to GC physical addresses).
4740	*/
4741	/** @todo We should not flush chunks which include cr3 mappings. */
4742	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
4743	{
4744	PPGMCPU pPGM = &pVM->apCpusR3[idCpu]->pgm.s;
4745
4746	pPGM->pGst32BitPdR3 = NULL;
4747	pPGM->pGstPaePdptR3 = NULL;
4748	pPGM->pGstAmd64Pml4R3 = NULL;
4749	pPGM->pGst32BitPdR0 = NIL_RTR0PTR;
4750	pPGM->pGstPaePdptR0 = NIL_RTR0PTR;
4751	pPGM->pGstAmd64Pml4R0 = NIL_RTR0PTR;
4752	for (unsigned i = 0; i < RT_ELEMENTS(pPGM->apGstPaePDsR3); i++)
4753	{
4754	pPGM->apGstPaePDsR3[i] = NULL;
4755	pPGM->apGstPaePDsR0[i] = NIL_RTR0PTR;
4756	}
4757
4758	/* Flush REM TLBs. */
4759	CPUMSetChangedFlags(pVM->apCpusR3[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
4760	}
4761	}
4762	}
4763	}
4764	PGM_UNLOCK(pVM);
4765	return rc;
4766	}
4767
4768	/**
4769	* Unmap a chunk to free up virtual address space (request packet handler for pgmR3PhysChunkMap)
4770	*
4771	* @returns VBox status code.
4772	* @param pVM The cross context VM structure.
4773	*/
4774	static DECLCALLBACK(void) pgmR3PhysUnmapChunk(PVM pVM)
4775	{
4776	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysUnmapChunkRendezvous, NULL);
4777	AssertRC(rc);
4778	}
4779
4780
4781	/**
4782	* Maps the given chunk into the ring-3 mapping cache.
4783	*
4784	* This will call ring-0.
4785	*
4786	* @returns VBox status code.
4787	* @param pVM The cross context VM structure.
4788	* @param idChunk The chunk in question.
4789	* @param ppChunk Where to store the chunk tracking structure.
4790	*
4791	* @remarks Called from within the PGM critical section.
4792	* @remarks Can be called from any thread!
4793	*/
4794	int pgmR3PhysChunkMap(PVM pVM, uint32_t idChunk, PPPGMCHUNKR3MAP ppChunk)
4795	{
4796	int rc;
4797
4798	PGM_LOCK_ASSERT_OWNER(pVM);
4799
4800	/*
4801	* Move the chunk time forward.
4802	*/
4803	pVM->pgm.s.ChunkR3Map.iNow++;
4804	if (pVM->pgm.s.ChunkR3Map.iNow == 0)
4805	{
4806	pVM->pgm.s.ChunkR3Map.iNow = 4;
4807	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkAgeingRolloverCallback, NULL);
4808	}
4809
4810	/*
4811	* Allocate a new tracking structure first.
4812	*/
4813	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3HeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk));
4814	AssertReturn(pChunk, VERR_NO_MEMORY);
4815	pChunk->Core.Key = idChunk;
4816	pChunk->iLastUsed = pVM->pgm.s.ChunkR3Map.iNow;
4817
4818	/*
4819	* Request the ring-0 part to map the chunk in question.
4820	*/
4821	GMMMAPUNMAPCHUNKREQ Req;
4822	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
4823	Req.Hdr.cbReq = sizeof(Req);
4824	Req.pvR3 = NULL;
4825	Req.idChunkMap = idChunk;
4826	Req.idChunkUnmap = NIL_GMM_CHUNKID;
4827
4828	/* Must be callable from any thread, so can't use VMMR3CallR0. */
4829	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatChunkMap, a);
4830	rc = SUPR3CallVMMR0Ex(VMCC_GET_VMR0_FOR_CALL(pVM), NIL_VMCPUID, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
4831	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatChunkMap, a);
4832	if (RT_SUCCESS(rc))
4833	{
4834	pChunk->pv = Req.pvR3;
4835
4836	/*
4837	* If we're running out of virtual address space, then we should
4838	* unmap another chunk.
4839	*
4840	* Currently, an unmap operation requires that all other virtual CPUs
4841	* are idling and not by chance making use of the memory we're
4842	* unmapping. So, we create an async unmap operation here.
4843	*
4844	* Now, when creating or restoring a saved state this wont work very
4845	* well since we may want to restore all guest RAM + a little something.
4846	* So, we have to do the unmap synchronously. Fortunately for us
4847	* though, during these operations the other virtual CPUs are inactive
4848	* and it should be safe to do this.
4849	*/
4850	/** @todo Eventually we should lock all memory when used and do
4851	* map+unmap as one kernel call without any rendezvous or
4852	* other precautions. */
4853	if (pVM->pgm.s.ChunkR3Map.c + 1 >= pVM->pgm.s.ChunkR3Map.cMax)
4854	{
4855	switch (VMR3GetState(pVM))
4856	{
4857	case VMSTATE_LOADING:
4858	case VMSTATE_SAVING:
4859	{
4860	PVMCPU pVCpu = VMMGetCpu(pVM);
4861	if ( pVCpu
4862	&& pVM->pgm.s.cDeprecatedPageLocks == 0)
4863	{
4864	pgmR3PhysUnmapChunkRendezvous(pVM, pVCpu, NULL);
4865	break;
4866	}
4867	}
4868	RT_FALL_THRU();
4869	default:
4870	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysUnmapChunk, 1, pVM);
4871	AssertRC(rc);
4872	break;
4873	}
4874	}
4875
4876	/*
4877	* Update the tree. We must do this after any unmapping to make sure
4878	* the chunk we're going to return isn't unmapped by accident.
4879	*/
4880	AssertPtr(Req.pvR3);
4881	bool fRc = RTAvlU32Insert(&pVM->pgm.s.ChunkR3Map.pTree, &pChunk->Core);
4882	AssertRelease(fRc);
4883	pVM->pgm.s.ChunkR3Map.c++;
4884	pVM->pgm.s.cMappedChunks++;
4885	}
4886	else
4887	{
4888	/** @todo this may fail because of /proc/sys/vm/max_map_count, so we
4889	* should probably restrict ourselves on linux. */
4890	AssertRC(rc);
4891	MMR3HeapFree(pChunk);
4892	pChunk = NULL;
4893	}
4894
4895	*ppChunk = pChunk;
4896	return rc;
4897	}
4898
4899
4900	/**
4901	* Invalidates the TLB for the ring-3 mapping cache.
4902	*
4903	* @param pVM The cross context VM structure.
4904	*/
4905	VMMR3DECL(void) PGMR3PhysChunkInvalidateTLB(PVM pVM)
4906	{
4907	PGM_LOCK_VOID(pVM);
4908	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
4909	{
4910	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk = NIL_GMM_CHUNKID;
4911	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk = NULL;
4912	}
4913	/* The page map TLB references chunks, so invalidate that one too. */
4914	pgmPhysInvalidatePageMapTLB(pVM);
4915	PGM_UNLOCK(pVM);
4916	}
4917
4918
4919	/**
4920	* Response to VMMCALLRING3_PGM_ALLOCATE_LARGE_HANDY_PAGE to allocate a large
4921	* (2MB) page for use with a nested paging PDE.
4922	*
4923	* @returns The following VBox status codes.
4924	* @retval VINF_SUCCESS on success.
4925	* @retval VINF_EM_NO_MEMORY if we're out of memory.
4926	*
4927	* @param pVM The cross context VM structure.
4928	* @param GCPhys GC physical start address of the 2 MB range
4929	*/
4930	VMMR3_INT_DECL(int) PGMR3PhysAllocateLargePage(PVM pVM, RTGCPHYS GCPhys)
4931	{
4932	#ifdef PGM_WITH_LARGE_PAGES
4933	PGM_LOCK_VOID(pVM);
4934
4935	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatAllocLargePage, a);
4936	uint64_t const msAllocStart = RTTimeMilliTS();
4937	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_LARGE_HANDY_PAGE, 0, NULL);
4938	uint64_t const cMsElapsed = RTTimeMilliTS() - msAllocStart;
4939	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatAllocLargePage, a);
4940	if (RT_SUCCESS(rc))
4941	{
4942	Assert(pVM->pgm.s.cLargeHandyPages == 1);
4943
4944	uint32_t idPage = pVM->pgm.s.aLargeHandyPage[0].idPage;
4945	RTHCPHYS HCPhys = pVM->pgm.s.aLargeHandyPage[0].HCPhysGCPhys;
4946
4947	void *pv;
4948
4949	/* Map the large page into our address space.
4950	*
4951	* Note: assuming that within the 2 MB range:
4952	* - GCPhys + PAGE_SIZE = HCPhys + PAGE_SIZE (whole point of this exercise)
4953	* - user space mapping is continuous as well
4954	* - page id (GCPhys) + 1 = page id (GCPhys + PAGE_SIZE)
4955	*/
4956	rc = pgmPhysPageMapByPageID(pVM, idPage, HCPhys, &pv);
4957	AssertLogRelMsg(RT_SUCCESS(rc), ("idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n", idPage, HCPhys, rc));
4958
4959	if (RT_SUCCESS(rc))
4960	{
4961	/*
4962	* Clear the pages.
4963	*/
4964	STAM_PROFILE_START(&pVM->pgm.s.Stats.StatClearLargePage, b);
4965	for (unsigned i = 0; i < _2M/PAGE_SIZE; i++)
4966	{
4967	ASMMemZeroPage(pv);
4968
4969	PPGMPAGE pPage;
4970	rc = pgmPhysGetPageEx(pVM, GCPhys, &pPage);
4971	AssertRC(rc);
4972
4973	Assert(PGM_PAGE_IS_ZERO(pPage));
4974	STAM_COUNTER_INC(&pVM->pgm.s.Stats.StatRZPageReplaceZero);
4975	pVM->pgm.s.cZeroPages--;
4976
4977	/*
4978	* Do the PGMPAGE modifications.
4979	*/
4980	pVM->pgm.s.cPrivatePages++;
4981	PGM_PAGE_SET_HCPHYS(pVM, pPage, HCPhys);
4982	PGM_PAGE_SET_PAGEID(pVM, pPage, idPage);
4983	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
4984	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_PDE);
4985	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
4986	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
4987
4988	/* Somewhat dirty assumption that page ids are increasing. */
4989	idPage++;
4990
4991	HCPhys += PAGE_SIZE;
4992	GCPhys += PAGE_SIZE;
4993
4994	pv = (void *)((uintptr_t)pv + PAGE_SIZE);
4995
4996	Log3(("PGMR3PhysAllocateLargePage: idPage=%#x HCPhys=%RGp\n", idPage, HCPhys));
4997	}
4998	STAM_PROFILE_STOP(&pVM->pgm.s.Stats.StatClearLargePage, b);
4999
5000	/* Flush all TLBs. */
5001	PGM_INVL_ALL_VCPU_TLBS(pVM);
5002	pgmPhysInvalidatePageMapTLB(pVM);
5003	}
5004	pVM->pgm.s.cLargeHandyPages = 0;
5005	}
5006
5007	if (RT_SUCCESS(rc))
5008	{
5009	static uint32_t cTimeOut = 0;
5010	if (cMsElapsed > 100)
5011	{
5012	STAM_COUNTER_INC(&pVM->pgm.s.Stats.StatLargePageOverflow);
5013	if ( ++cTimeOut > 10
5014	\|\| cMsElapsed > 1000 /* more than one second forces an early retirement from allocating large pages. */)
5015	{
5016	/* If repeated attempts to allocate a large page takes more than 100 ms, then we fall back to normal 4k pages.
5017	* E.g. Vista 64 tries to move memory around, which takes a huge amount of time.
5018	*/
5019	LogRel(("PGMR3PhysAllocateLargePage: allocating large pages takes too long (last attempt %RU64 ms; nr of timeouts %d); DISABLE\n", cMsElapsed, cTimeOut));
5020	PGMSetLargePageUsage(pVM, false);
5021	}
5022	}
5023	else if (cTimeOut > 0)
5024	cTimeOut--;
5025	}
5026
5027	PGM_UNLOCK(pVM);
5028	return rc;
5029	#else
5030	RT_NOREF(pVM, GCPhys);
5031	return VERR_NOT_IMPLEMENTED;
5032	#endif /* PGM_WITH_LARGE_PAGES */
5033	}
5034
5035
5036	/**
5037	* Response to VM_FF_PGM_NEED_HANDY_PAGES and VMMCALLRING3_PGM_ALLOCATE_HANDY_PAGES.
5038	*
5039	* This function will also work the VM_FF_PGM_NO_MEMORY force action flag, to
5040	* signal and clear the out of memory condition. When contracted, this API is
5041	* used to try clear the condition when the user wants to resume.
5042	*
5043	* @returns The following VBox status codes.
5044	* @retval VINF_SUCCESS on success. FFs cleared.
5045	* @retval VINF_EM_NO_MEMORY if we're out of memory. The FF is not cleared in
5046	* this case and it gets accompanied by VM_FF_PGM_NO_MEMORY.
5047	*
5048	* @param pVM The cross context VM structure.
5049	*
5050	* @remarks The VINF_EM_NO_MEMORY status is for the benefit of the FF processing
5051	* in EM.cpp and shouldn't be propagated outside TRPM, HM, EM and
5052	* pgmPhysEnsureHandyPage. There is one exception to this in the \#PF
5053	* handler.
5054	*/
5055	VMMR3DECL(int) PGMR3PhysAllocateHandyPages(PVM pVM)
5056	{
5057	PGM_LOCK_VOID(pVM);
5058
5059	/*
5060	* Allocate more pages, noting down the index of the first new page.
5061	*/
5062	uint32_t iClear = pVM->pgm.s.cHandyPages;
5063	AssertMsgReturn(iClear <= RT_ELEMENTS(pVM->pgm.s.aHandyPages), ("%d", iClear), VERR_PGM_HANDY_PAGE_IPE);
5064	Log(("PGMR3PhysAllocateHandyPages: %d -> %d\n", iClear, RT_ELEMENTS(pVM->pgm.s.aHandyPages)));
5065	int rcAlloc = VINF_SUCCESS;
5066	int rcSeed = VINF_SUCCESS;
5067	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
5068	while (rc == VERR_GMM_SEED_ME)
5069	{
5070	void *pvChunk;
5071	rcAlloc = rc = SUPR3PageAlloc(GMM_CHUNK_SIZE >> PAGE_SHIFT, &pvChunk);
5072	if (RT_SUCCESS(rc))
5073	{
5074	rcSeed = rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_SEED_CHUNK, (uintptr_t)pvChunk, NULL);
5075	if (RT_FAILURE(rc))
5076	SUPR3PageFree(pvChunk, GMM_CHUNK_SIZE >> PAGE_SHIFT);
5077	}
5078	if (RT_SUCCESS(rc))
5079	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
5080	}
5081
5082	/** @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) */
5083	if ( rc == VERR_GMM_HIT_VM_ACCOUNT_LIMIT
5084	&& pVM->pgm.s.cHandyPages > 0)
5085	{
5086	/* Still handy pages left, so don't panic. */
5087	rc = VINF_SUCCESS;
5088	}
5089
5090	if (RT_SUCCESS(rc))
5091	{
5092	AssertMsg(rc == VINF_SUCCESS, ("%Rrc\n", rc));
5093	Assert(pVM->pgm.s.cHandyPages > 0);
5094	VM_FF_CLEAR(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
5095	VM_FF_CLEAR(pVM, VM_FF_PGM_NO_MEMORY);
5096
5097	#ifdef VBOX_STRICT
5098	uint32_t i;
5099	for (i = iClear; i < pVM->pgm.s.cHandyPages; i++)
5100	if ( pVM->pgm.s.aHandyPages[i].idPage == NIL_GMM_PAGEID
5101	\|\| pVM->pgm.s.aHandyPages[i].idSharedPage != NIL_GMM_PAGEID
5102	\|\| (pVM->pgm.s.aHandyPages[i].HCPhysGCPhys & PAGE_OFFSET_MASK))
5103	break;
5104	if (i != pVM->pgm.s.cHandyPages)
5105	{
5106	RTAssertMsg1Weak(NULL, __LINE__, __FILE__, __FUNCTION__);
5107	RTAssertMsg2Weak("i=%d iClear=%d cHandyPages=%d\n", i, iClear, pVM->pgm.s.cHandyPages);
5108	for (uint32_t j = iClear; j < pVM->pgm.s.cHandyPages; j++)
5109	RTAssertMsg2Add("%03d: idPage=%d HCPhysGCPhys=%RHp idSharedPage=%d%\n", j,
5110	pVM->pgm.s.aHandyPages[j].idPage,
5111	pVM->pgm.s.aHandyPages[j].HCPhysGCPhys,
5112	pVM->pgm.s.aHandyPages[j].idSharedPage,
5113	j == i ? " <---" : "");
5114	RTAssertPanic();
5115	}
5116	#endif
5117	/*
5118	* Clear the pages.
5119	*/
5120	while (iClear < pVM->pgm.s.cHandyPages)
5121	{
5122	PGMMPAGEDESC pPage = &pVM->pgm.s.aHandyPages[iClear];
5123	void *pv;
5124	rc = pgmPhysPageMapByPageID(pVM, pPage->idPage, pPage->HCPhysGCPhys, &pv);
5125	AssertLogRelMsgBreak(RT_SUCCESS(rc),
5126	("%u/%u: idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n",
5127	iClear, pVM->pgm.s.cHandyPages, pPage->idPage, pPage->HCPhysGCPhys, rc));
5128	ASMMemZeroPage(pv);
5129	iClear++;
5130	Log3(("PGMR3PhysAllocateHandyPages: idPage=%#x HCPhys=%RGp\n", pPage->idPage, pPage->HCPhysGCPhys));
5131	}
5132	}
5133	else
5134	{
5135	uint64_t cAllocPages, cMaxPages, cBalloonPages;
5136
5137	/*
5138	* We should never get here unless there is a genuine shortage of
5139	* memory (or some internal error). Flag the error so the VM can be
5140	* suspended ASAP and the user informed. If we're totally out of
5141	* handy pages we will return failure.
5142	*/
5143	/* Report the failure. */
5144	LogRel(("PGM: Failed to procure handy pages; rc=%Rrc rcAlloc=%Rrc rcSeed=%Rrc cHandyPages=%#x\n"
5145	" cAllPages=%#x cPrivatePages=%#x cSharedPages=%#x cZeroPages=%#x\n",
5146	rc, rcAlloc, rcSeed,
5147	pVM->pgm.s.cHandyPages,
5148	pVM->pgm.s.cAllPages,
5149	pVM->pgm.s.cPrivatePages,
5150	pVM->pgm.s.cSharedPages,
5151	pVM->pgm.s.cZeroPages));
5152
5153	if (GMMR3QueryMemoryStats(pVM, &cAllocPages, &cMaxPages, &cBalloonPages) == VINF_SUCCESS)
5154	{
5155	LogRel(("GMM: Statistics:\n"
5156	" Allocated pages: %RX64\n"
5157	" Maximum pages: %RX64\n"
5158	" Ballooned pages: %RX64\n", cAllocPages, cMaxPages, cBalloonPages));
5159	}
5160
5161	if ( rc != VERR_NO_MEMORY
5162	&& rc != VERR_NO_PHYS_MEMORY
5163	&& rc != VERR_LOCK_FAILED)
5164	{
5165	for (uint32_t i = 0; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
5166	{
5167	LogRel(("PGM: aHandyPages[#%#04x] = {.HCPhysGCPhys=%RHp, .idPage=%#08x, .idSharedPage=%#08x}\n",
5168	i, pVM->pgm.s.aHandyPages[i].HCPhysGCPhys, pVM->pgm.s.aHandyPages[i].idPage,
5169	pVM->pgm.s.aHandyPages[i].idSharedPage));
5170	uint32_t const idPage = pVM->pgm.s.aHandyPages[i].idPage;
5171	if (idPage != NIL_GMM_PAGEID)
5172	{
5173	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
5174	pRam;
5175	pRam = pRam->pNextR3)
5176	{
5177	uint32_t const cPages = pRam->cb >> PAGE_SHIFT;
5178	for (uint32_t iPage = 0; iPage < cPages; iPage++)
5179	if (PGM_PAGE_GET_PAGEID(&pRam->aPages[iPage]) == idPage)
5180	LogRel(("PGM: Used by %RGp %R[pgmpage] (%s)\n",
5181	pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pRam->aPages[iPage], pRam->pszDesc));
5182	}
5183	}
5184	}
5185	}
5186
5187	if (rc == VERR_NO_MEMORY)
5188	{
5189	uint64_t cbHostRamAvail = 0;
5190	int rc2 = RTSystemQueryAvailableRam(&cbHostRamAvail);
5191	if (RT_SUCCESS(rc2))
5192	LogRel(("Host RAM: %RU64MB available\n", cbHostRamAvail / _1M));
5193	else
5194	LogRel(("Cannot determine the amount of available host memory\n"));
5195	}
5196
5197	/* Set the FFs and adjust rc. */
5198	VM_FF_SET(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
5199	VM_FF_SET(pVM, VM_FF_PGM_NO_MEMORY);
5200	if ( rc == VERR_NO_MEMORY
5201	\|\| rc == VERR_NO_PHYS_MEMORY
5202	\|\| rc == VERR_LOCK_FAILED)
5203	rc = VINF_EM_NO_MEMORY;
5204	}
5205
5206	PGM_UNLOCK(pVM);
5207	return rc;
5208	}
5209
5210
5211	/**
5212	* Frees the specified RAM page and replaces it with the ZERO page.
5213	*
5214	* This is used by ballooning, remapping MMIO2, RAM reset and state loading.
5215	*
5216	* @param pVM The cross context VM structure.
5217	* @param pReq Pointer to the request.
5218	* @param pcPendingPages Where the number of pages waiting to be freed are
5219	* kept. This will normally be incremented.
5220	* @param pPage Pointer to the page structure.
5221	* @param GCPhys The guest physical address of the page, if applicable.
5222	* @param enmNewType New page type for NEM notification, since several
5223	* callers will change the type upon successful return.
5224	*
5225	* @remarks The caller must own the PGM lock.
5226	*/
5227	int pgmPhysFreePage(PVM pVM, PGMMFREEPAGESREQ pReq, uint32_t *pcPendingPages, PPGMPAGE pPage, RTGCPHYS GCPhys,
5228	PGMPAGETYPE enmNewType)
5229	{
5230	/*
5231	* Assert sanity.
5232	*/
5233	PGM_LOCK_ASSERT_OWNER(pVM);
5234	if (RT_UNLIKELY( PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM
5235	&& PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_ROM_SHADOW))
5236	{
5237	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
5238	return VMSetError(pVM, VERR_PGM_PHYS_NOT_RAM, RT_SRC_POS, "GCPhys=%RGp type=%d", GCPhys, PGM_PAGE_GET_TYPE(pPage));
5239	}
5240
5241	/** @todo What about ballooning of large pages??! */
5242	Assert( PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE
5243	&& PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE_DISABLED);
5244
5245	if ( PGM_PAGE_IS_ZERO(pPage)
5246	\|\| PGM_PAGE_IS_BALLOONED(pPage))
5247	return VINF_SUCCESS;
5248
5249	const uint32_t idPage = PGM_PAGE_GET_PAGEID(pPage);
5250	Log3(("pgmPhysFreePage: idPage=%#x GCPhys=%RGp pPage=%R[pgmpage]\n", idPage, GCPhys, pPage));
5251	if (RT_UNLIKELY( idPage == NIL_GMM_PAGEID
5252	\|\| idPage > GMM_PAGEID_LAST
5253	\|\| PGM_PAGE_GET_CHUNKID(pPage) == NIL_GMM_CHUNKID))
5254	{
5255	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
5256	return VMSetError(pVM, VERR_PGM_PHYS_INVALID_PAGE_ID, RT_SRC_POS, "GCPhys=%RGp idPage=%#x", GCPhys, pPage);
5257	}
5258	const RTHCPHYS HCPhysPrev = PGM_PAGE_GET_HCPHYS(pPage);
5259
5260	/* update page count stats. */
5261	if (PGM_PAGE_IS_SHARED(pPage))
5262	pVM->pgm.s.cSharedPages--;
5263	else
5264	pVM->pgm.s.cPrivatePages--;
5265	pVM->pgm.s.cZeroPages++;
5266
5267	/* Deal with write monitored pages. */
5268	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
5269	{
5270	PGM_PAGE_SET_WRITTEN_TO(pVM, pPage);
5271	pVM->pgm.s.cWrittenToPages++;
5272	}
5273
5274	/*
5275	* pPage = ZERO page.
5276	*/
5277	PGM_PAGE_SET_HCPHYS(pVM, pPage, pVM->pgm.s.HCPhysZeroPg);
5278	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
5279	PGM_PAGE_SET_PAGEID(pVM, pPage, NIL_GMM_PAGEID);
5280	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
5281	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
5282	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
5283
5284	/* Flush physical page map TLB entry. */
5285	pgmPhysInvalidatePageMapTLBEntry(pVM, GCPhys);
5286
5287	/* Notify NEM. */
5288	/** @todo consider doing batch NEM notifications. */
5289	if (VM_IS_NEM_ENABLED(pVM))
5290	{
5291	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pPage);
5292	NEMHCNotifyPhysPageChanged(pVM, GCPhys, HCPhysPrev, pVM->pgm.s.HCPhysZeroPg,
5293	pgmPhysPageCalcNemProtection(pPage, enmNewType), enmNewType, &u2State);
5294	PGM_PAGE_SET_NEM_STATE(pPage, u2State);
5295	}
5296
5297	/*
5298	* Make sure it's not in the handy page array.
5299	*/
5300	for (uint32_t i = pVM->pgm.s.cHandyPages; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
5301	{
5302	if (pVM->pgm.s.aHandyPages[i].idPage == idPage)
5303	{
5304	pVM->pgm.s.aHandyPages[i].idPage = NIL_GMM_PAGEID;
5305	break;
5306	}
5307	if (pVM->pgm.s.aHandyPages[i].idSharedPage == idPage)
5308	{
5309	pVM->pgm.s.aHandyPages[i].idSharedPage = NIL_GMM_PAGEID;
5310	break;
5311	}
5312	}
5313
5314	/*
5315	* Push it onto the page array.
5316	*/
5317	uint32_t iPage = *pcPendingPages;
5318	Assert(iPage < PGMPHYS_FREE_PAGE_BATCH_SIZE);
5319	*pcPendingPages += 1;
5320
5321	pReq->aPages[iPage].idPage = idPage;
5322
5323	if (iPage + 1 < PGMPHYS_FREE_PAGE_BATCH_SIZE)
5324	return VINF_SUCCESS;
5325
5326	/*
5327	* Flush the pages.
5328	*/
5329	int rc = GMMR3FreePagesPerform(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE);
5330	if (RT_SUCCESS(rc))
5331	{
5332	GMMR3FreePagesRePrep(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
5333	*pcPendingPages = 0;
5334	}
5335	return rc;
5336	}
5337
5338
5339	/**
5340	* Converts a GC physical address to a HC ring-3 pointer, with some
5341	* additional checks.
5342	*
5343	* @returns VBox status code.
5344	* @retval VINF_SUCCESS on success.
5345	* @retval VINF_PGM_PHYS_TLB_CATCH_WRITE and *ppv set if the page has a write
5346	* access handler of some kind.
5347	* @retval VERR_PGM_PHYS_TLB_CATCH_ALL if the page has a handler catching all
5348	* accesses or is odd in any way.
5349	* @retval VERR_PGM_PHYS_TLB_UNASSIGNED if the page doesn't exist.
5350	*
5351	* @param pVM The cross context VM structure.
5352	* @param GCPhys The GC physical address to convert. Since this is only
5353	* used for filling the REM TLB, the A20 mask must be
5354	* applied before calling this API.
5355	* @param fWritable Whether write access is required.
5356	* @param ppv Where to store the pointer corresponding to GCPhys on
5357	* success.
5358	*/
5359	VMMR3DECL(int) PGMR3PhysTlbGCPhys2Ptr(PVM pVM, RTGCPHYS GCPhys, bool fWritable, void **ppv)
5360	{
5361	PGM_LOCK_VOID(pVM);
5362	PGM_A20_ASSERT_MASKED(VMMGetCpu(pVM), GCPhys);
5363
5364	PPGMRAMRANGE pRam;
5365	PPGMPAGE pPage;
5366	int rc = pgmPhysGetPageAndRangeEx(pVM, GCPhys, &pPage, &pRam);
5367	if (RT_SUCCESS(rc))
5368	{
5369	if (PGM_PAGE_IS_BALLOONED(pPage))
5370	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5371	else if (!PGM_PAGE_HAS_ANY_HANDLERS(pPage))
5372	rc = VINF_SUCCESS;
5373	else
5374	{
5375	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)) /* catches MMIO */
5376	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
5377	else if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
5378	{
5379	/** @todo Handle TLB loads of virtual handlers so ./test.sh can be made to work
5380	* in -norawr0 mode. */
5381	if (fWritable)
5382	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5383	}
5384	else
5385	{
5386	/* Temporarily disabled physical handler(s), since the recompiler
5387	doesn't get notified when it's reset we'll have to pretend it's
5388	operating normally. */
5389	if (pgmHandlerPhysicalIsAll(pVM, GCPhys))
5390	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
5391	else
5392	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5393	}
5394	}
5395	if (RT_SUCCESS(rc))
5396	{
5397	int rc2;
5398
5399	/* Make sure what we return is writable. */
5400	if (fWritable)
5401	switch (PGM_PAGE_GET_STATE(pPage))
5402	{
5403	case PGM_PAGE_STATE_ALLOCATED:
5404	break;
5405	case PGM_PAGE_STATE_BALLOONED:
5406	AssertFailed();
5407	break;
5408	case PGM_PAGE_STATE_ZERO:
5409	case PGM_PAGE_STATE_SHARED:
5410	if (rc == VINF_PGM_PHYS_TLB_CATCH_WRITE)
5411	break;
5412	RT_FALL_THRU();
5413	case PGM_PAGE_STATE_WRITE_MONITORED:
5414	rc2 = pgmPhysPageMakeWritable(pVM, pPage, GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK);
5415	AssertLogRelRCReturn(rc2, rc2);
5416	break;
5417	}
5418
5419	/* Get a ring-3 mapping of the address. */
5420	PPGMPAGER3MAPTLBE pTlbe;
5421	rc2 = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
5422	AssertLogRelRCReturn(rc2, rc2);
5423	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
5424	/** @todo mapping/locking hell; this isn't horribly efficient since
5425	* pgmPhysPageLoadIntoTlb will repeat the lookup we've done here. */
5426
5427	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage] ppv=%p\n", GCPhys, rc, pPage, ppv));
5428	}
5429	else
5430	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage]\n", GCPhys, rc, pPage));
5431
5432	/* else: handler catching all access, no pointer returned. */
5433	}
5434	else
5435	rc = VERR_PGM_PHYS_TLB_UNASSIGNED;
5436
5437	PGM_UNLOCK(pVM);
5438	return rc;
5439	}
5440

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PGMPhys.cpp@ 91688

Download in other formats: