PGMPhys.cpp@ 77758

Last change on this file since 77758 was 77441, checked in by vboxsync, 6 years ago
PGMR3PhysBulkGCPhys2CCPtrExternal: Fixed pgmR3PhysGCPhys2CCPtrDelegated call. bugref:9172
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 214.4 KB

Line
1	/* $Id: PGMPhys.cpp 77441 2019-02-22 20:53:01Z vboxsync $ */
2	/** @file
3	* PGM - Page Manager and Monitor, Physical Memory Addressing.
4	*/
5
6	/*
7	* Copyright (C) 2006-2019 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*********************************************************************************************************************************
20	* Header Files *
21	*********************************************************************************************************************************/
22	#define LOG_GROUP LOG_GROUP_PGM_PHYS
23	#include <VBox/vmm/pgm.h>
24	#include <VBox/vmm/iem.h>
25	#include <VBox/vmm/iom.h>
26	#include <VBox/vmm/mm.h>
27	#include <VBox/vmm/nem.h>
28	#include <VBox/vmm/stam.h>
29	#ifdef VBOX_WITH_REM
30	# include <VBox/vmm/rem.h>
31	#endif
32	#include <VBox/vmm/pdmdev.h>
33	#include "PGMInternal.h"
34	#include <VBox/vmm/vm.h>
35	#include <VBox/vmm/uvm.h>
36	#include "PGMInline.h"
37	#include <VBox/sup.h>
38	#include <VBox/param.h>
39	#include <VBox/err.h>
40	#include <VBox/log.h>
41	#include <iprt/assert.h>
42	#include <iprt/alloc.h>
43	#include <iprt/asm.h>
44	#ifdef VBOX_STRICT
45	# include <iprt/crc.h>
46	#endif
47	#include <iprt/thread.h>
48	#include <iprt/string.h>
49	#include <iprt/system.h>
50
51
52	/*********************************************************************************************************************************
53	* Defined Constants And Macros *
54	*********************************************************************************************************************************/
55	/** The number of pages to free in one batch. */
56	#define PGMPHYS_FREE_PAGE_BATCH_SIZE 128
57
58
59	/*
60	* PGMR3PhysReadU8-64
61	* PGMR3PhysWriteU8-64
62	*/
63	#define PGMPHYSFN_READNAME PGMR3PhysReadU8
64	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU8
65	#define PGMPHYS_DATASIZE 1
66	#define PGMPHYS_DATATYPE uint8_t
67	#include "PGMPhysRWTmpl.h"
68
69	#define PGMPHYSFN_READNAME PGMR3PhysReadU16
70	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU16
71	#define PGMPHYS_DATASIZE 2
72	#define PGMPHYS_DATATYPE uint16_t
73	#include "PGMPhysRWTmpl.h"
74
75	#define PGMPHYSFN_READNAME PGMR3PhysReadU32
76	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU32
77	#define PGMPHYS_DATASIZE 4
78	#define PGMPHYS_DATATYPE uint32_t
79	#include "PGMPhysRWTmpl.h"
80
81	#define PGMPHYSFN_READNAME PGMR3PhysReadU64
82	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU64
83	#define PGMPHYS_DATASIZE 8
84	#define PGMPHYS_DATATYPE uint64_t
85	#include "PGMPhysRWTmpl.h"
86
87
88	/**
89	* EMT worker for PGMR3PhysReadExternal.
90	*/
91	static DECLCALLBACK(int) pgmR3PhysReadExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, void *pvBuf, size_t cbRead,
92	PGMACCESSORIGIN enmOrigin)
93	{
94	VBOXSTRICTRC rcStrict = PGMPhysRead(pVM, *pGCPhys, pvBuf, cbRead, enmOrigin);
95	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
96	return VINF_SUCCESS;
97	}
98
99
100	/**
101	* Read from physical memory, external users.
102	*
103	* @returns VBox status code.
104	* @retval VINF_SUCCESS.
105	*
106	* @param pVM The cross context VM structure.
107	* @param GCPhys Physical address to read from.
108	* @param pvBuf Where to read into.
109	* @param cbRead How many bytes to read.
110	* @param enmOrigin Who is calling.
111	*
112	* @thread Any but EMTs.
113	*/
114	VMMR3DECL(int) PGMR3PhysReadExternal(PVM pVM, RTGCPHYS GCPhys, void *pvBuf, size_t cbRead, PGMACCESSORIGIN enmOrigin)
115	{
116	VM_ASSERT_OTHER_THREAD(pVM);
117
118	AssertMsgReturn(cbRead > 0, ("don't even think about reading zero bytes!\n"), VINF_SUCCESS);
119	LogFlow(("PGMR3PhysReadExternal: %RGp %d\n", GCPhys, cbRead));
120
121	pgmLock(pVM);
122
123	/*
124	* Copy loop on ram ranges.
125	*/
126	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
127	for (;;)
128	{
129	/* Inside range or not? */
130	if (pRam && GCPhys >= pRam->GCPhys)
131	{
132	/*
133	* Must work our way thru this page by page.
134	*/
135	RTGCPHYS off = GCPhys - pRam->GCPhys;
136	while (off < pRam->cb)
137	{
138	unsigned iPage = off >> PAGE_SHIFT;
139	PPGMPAGE pPage = &pRam->aPages[iPage];
140
141	/*
142	* If the page has an ALL access handler, we'll have to
143	* delegate the job to EMT.
144	*/
145	if ( PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
146	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
147	{
148	pgmUnlock(pVM);
149
150	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysReadExternalEMT, 5,
151	pVM, &GCPhys, pvBuf, cbRead, enmOrigin);
152	}
153	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
154
155	/*
156	* Simple stuff, go ahead.
157	*/
158	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
159	if (cb > cbRead)
160	cb = cbRead;
161	PGMPAGEMAPLOCK PgMpLck;
162	const void *pvSrc;
163	int rc = pgmPhysGCPhys2CCPtrInternalReadOnly(pVM, pPage, pRam->GCPhys + off, &pvSrc, &PgMpLck);
164	if (RT_SUCCESS(rc))
165	{
166	memcpy(pvBuf, pvSrc, cb);
167	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
168	}
169	else
170	{
171	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternalReadOnly failed on %RGp / %R[pgmpage] -> %Rrc\n",
172	pRam->GCPhys + off, pPage, rc));
173	memset(pvBuf, 0xff, cb);
174	}
175
176	/* next page */
177	if (cb >= cbRead)
178	{
179	pgmUnlock(pVM);
180	return VINF_SUCCESS;
181	}
182	cbRead -= cb;
183	off += cb;
184	GCPhys += cb;
185	pvBuf = (char *)pvBuf + cb;
186	} /* walk pages in ram range. */
187	}
188	else
189	{
190	LogFlow(("PGMPhysRead: Unassigned %RGp size=%u\n", GCPhys, cbRead));
191
192	/*
193	* Unassigned address space.
194	*/
195	size_t cb = pRam ? pRam->GCPhys - GCPhys : ~(size_t)0;
196	if (cb >= cbRead)
197	{
198	memset(pvBuf, 0xff, cbRead);
199	break;
200	}
201	memset(pvBuf, 0xff, cb);
202
203	cbRead -= cb;
204	pvBuf = (char *)pvBuf + cb;
205	GCPhys += cb;
206	}
207
208	/* Advance range if necessary. */
209	while (pRam && GCPhys > pRam->GCPhysLast)
210	pRam = pRam->CTX_SUFF(pNext);
211	} /* Ram range walk */
212
213	pgmUnlock(pVM);
214
215	return VINF_SUCCESS;
216	}
217
218
219	/**
220	* EMT worker for PGMR3PhysWriteExternal.
221	*/
222	static DECLCALLBACK(int) pgmR3PhysWriteExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, const void *pvBuf, size_t cbWrite,
223	PGMACCESSORIGIN enmOrigin)
224	{
225	/** @todo VERR_EM_NO_MEMORY */
226	VBOXSTRICTRC rcStrict = PGMPhysWrite(pVM, *pGCPhys, pvBuf, cbWrite, enmOrigin);
227	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
228	return VINF_SUCCESS;
229	}
230
231
232	/**
233	* Write to physical memory, external users.
234	*
235	* @returns VBox status code.
236	* @retval VINF_SUCCESS.
237	* @retval VERR_EM_NO_MEMORY.
238	*
239	* @param pVM The cross context VM structure.
240	* @param GCPhys Physical address to write to.
241	* @param pvBuf What to write.
242	* @param cbWrite How many bytes to write.
243	* @param enmOrigin Who is calling.
244	*
245	* @thread Any but EMTs.
246	*/
247	VMMDECL(int) PGMR3PhysWriteExternal(PVM pVM, RTGCPHYS GCPhys, const void *pvBuf, size_t cbWrite, PGMACCESSORIGIN enmOrigin)
248	{
249	VM_ASSERT_OTHER_THREAD(pVM);
250
251	AssertMsg(!pVM->pgm.s.fNoMorePhysWrites,
252	("Calling PGMR3PhysWriteExternal after pgmR3Save()! GCPhys=%RGp cbWrite=%#x enmOrigin=%d\n",
253	GCPhys, cbWrite, enmOrigin));
254	AssertMsgReturn(cbWrite > 0, ("don't even think about writing zero bytes!\n"), VINF_SUCCESS);
255	LogFlow(("PGMR3PhysWriteExternal: %RGp %d\n", GCPhys, cbWrite));
256
257	pgmLock(pVM);
258
259	/*
260	* Copy loop on ram ranges, stop when we hit something difficult.
261	*/
262	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
263	for (;;)
264	{
265	/* Inside range or not? */
266	if (pRam && GCPhys >= pRam->GCPhys)
267	{
268	/*
269	* Must work our way thru this page by page.
270	*/
271	RTGCPTR off = GCPhys - pRam->GCPhys;
272	while (off < pRam->cb)
273	{
274	RTGCPTR iPage = off >> PAGE_SHIFT;
275	PPGMPAGE pPage = &pRam->aPages[iPage];
276
277	/*
278	* Is the page problematic, we have to do the work on the EMT.
279	*
280	* Allocating writable pages and access handlers are
281	* problematic, write monitored pages are simple and can be
282	* dealt with here.
283	*/
284	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
285	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
286	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
287	{
288	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
289	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
290	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, GCPhys);
291	else
292	{
293	pgmUnlock(pVM);
294
295	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysWriteExternalEMT, 5,
296	pVM, &GCPhys, pvBuf, cbWrite, enmOrigin);
297	}
298	}
299	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
300
301	/*
302	* Simple stuff, go ahead.
303	*/
304	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
305	if (cb > cbWrite)
306	cb = cbWrite;
307	PGMPAGEMAPLOCK PgMpLck;
308	void *pvDst;
309	int rc = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, pRam->GCPhys + off, &pvDst, &PgMpLck);
310	if (RT_SUCCESS(rc))
311	{
312	memcpy(pvDst, pvBuf, cb);
313	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
314	}
315	else
316	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternal failed on %RGp / %R[pgmpage] -> %Rrc\n",
317	pRam->GCPhys + off, pPage, rc));
318
319	/* next page */
320	if (cb >= cbWrite)
321	{
322	pgmUnlock(pVM);
323	return VINF_SUCCESS;
324	}
325
326	cbWrite -= cb;
327	off += cb;
328	GCPhys += cb;
329	pvBuf = (const char *)pvBuf + cb;
330	} /* walk pages in ram range */
331	}
332	else
333	{
334	/*
335	* Unassigned address space, skip it.
336	*/
337	if (!pRam)
338	break;
339	size_t cb = pRam->GCPhys - GCPhys;
340	if (cb >= cbWrite)
341	break;
342	cbWrite -= cb;
343	pvBuf = (const char *)pvBuf + cb;
344	GCPhys += cb;
345	}
346
347	/* Advance range if necessary. */
348	while (pRam && GCPhys > pRam->GCPhysLast)
349	pRam = pRam->CTX_SUFF(pNext);
350	} /* Ram range walk */
351
352	pgmUnlock(pVM);
353	return VINF_SUCCESS;
354	}
355
356
357	/**
358	* VMR3ReqCall worker for PGMR3PhysGCPhys2CCPtrExternal to make pages writable.
359	*
360	* @returns see PGMR3PhysGCPhys2CCPtrExternal
361	* @param pVM The cross context VM structure.
362	* @param pGCPhys Pointer to the guest physical address.
363	* @param ppv Where to store the mapping address.
364	* @param pLock Where to store the lock.
365	*/
366	static DECLCALLBACK(int) pgmR3PhysGCPhys2CCPtrDelegated(PVM pVM, PRTGCPHYS pGCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
367	{
368	/*
369	* Just hand it to PGMPhysGCPhys2CCPtr and check that it's not a page with
370	* an access handler after it succeeds.
371	*/
372	int rc = pgmLock(pVM);
373	AssertRCReturn(rc, rc);
374
375	rc = PGMPhysGCPhys2CCPtr(pVM, *pGCPhys, ppv, pLock);
376	if (RT_SUCCESS(rc))
377	{
378	PPGMPAGEMAPTLBE pTlbe;
379	int rc2 = pgmPhysPageQueryTlbe(pVM, *pGCPhys, &pTlbe);
380	AssertFatalRC(rc2);
381	PPGMPAGE pPage = pTlbe->pPage;
382	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
383	{
384	PGMPhysReleasePageMappingLock(pVM, pLock);
385	rc = VERR_PGM_PHYS_PAGE_RESERVED;
386	}
387	else if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
388	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
389	\|\| pgmPoolIsDirtyPage(pVM, *pGCPhys)
390	#endif
391	)
392	{
393	/* We must flush any corresponding pgm pool page here, otherwise we'll
394	* not be informed about writes and keep bogus gst->shw mappings around.
395	*/
396	pgmPoolFlushPageByGCPhys(pVM, *pGCPhys);
397	Assert(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage));
398	/** @todo r=bird: return VERR_PGM_PHYS_PAGE_RESERVED here if it still has
399	* active handlers, see the PGMR3PhysGCPhys2CCPtrExternal docs. */
400	}
401	}
402
403	pgmUnlock(pVM);
404	return rc;
405	}
406
407
408	/**
409	* Requests the mapping of a guest page into ring-3, external threads.
410	*
411	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
412	* release it.
413	*
414	* This API will assume your intention is to write to the page, and will
415	* therefore replace shared and zero pages. If you do not intend to modify the
416	* page, use the PGMR3PhysGCPhys2CCPtrReadOnlyExternal() API.
417	*
418	* @returns VBox status code.
419	* @retval VINF_SUCCESS on success.
420	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
421	* backing or if the page has any active access handlers. The caller
422	* must fall back on using PGMR3PhysWriteExternal.
423	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
424	*
425	* @param pVM The cross context VM structure.
426	* @param GCPhys The guest physical address of the page that should be mapped.
427	* @param ppv Where to store the address corresponding to GCPhys.
428	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
429	*
430	* @remark Avoid calling this API from within critical sections (other than the
431	* PGM one) because of the deadlock risk when we have to delegating the
432	* task to an EMT.
433	* @thread Any.
434	*/
435	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrExternal(PVM pVM, RTGCPHYS GCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
436	{
437	AssertPtr(ppv);
438	AssertPtr(pLock);
439
440	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
441
442	int rc = pgmLock(pVM);
443	AssertRCReturn(rc, rc);
444
445	/*
446	* Query the Physical TLB entry for the page (may fail).
447	*/
448	PPGMPAGEMAPTLBE pTlbe;
449	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
450	if (RT_SUCCESS(rc))
451	{
452	PPGMPAGE pPage = pTlbe->pPage;
453	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
454	rc = VERR_PGM_PHYS_PAGE_RESERVED;
455	else
456	{
457	/*
458	* If the page is shared, the zero page, or being write monitored
459	* it must be converted to an page that's writable if possible.
460	* We can only deal with write monitored pages here, the rest have
461	* to be on an EMT.
462	*/
463	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
464	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
465	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
466	\|\| pgmPoolIsDirtyPage(pVM, GCPhys)
467	#endif
468	)
469	{
470	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
471	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
472	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
473	&& !pgmPoolIsDirtyPage(pVM, GCPhys) /** @todo we're very likely doing this twice. */
474	#endif
475	)
476	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, GCPhys);
477	else
478	{
479	pgmUnlock(pVM);
480
481	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
482	pVM, &GCPhys, ppv, pLock);
483	}
484	}
485
486	/*
487	* Now, just perform the locking and calculate the return address.
488	*/
489	PPGMPAGEMAP pMap = pTlbe->pMap;
490	if (pMap)
491	pMap->cRefs++;
492
493	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
494	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
495	{
496	if (cLocks == 0)
497	pVM->pgm.s.cWriteLockedPages++;
498	PGM_PAGE_INC_WRITE_LOCKS(pPage);
499	}
500	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
501	{
502	PGM_PAGE_INC_WRITE_LOCKS(pPage);
503	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", GCPhys, pPage));
504	if (pMap)
505	pMap->cRefs++; /* Extra ref to prevent it from going away. */
506	}
507
508	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
509	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
510	pLock->pvMap = pMap;
511	}
512	}
513
514	pgmUnlock(pVM);
515	return rc;
516	}
517
518
519	/**
520	* Requests the mapping of a guest page into ring-3, external threads.
521	*
522	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
523	* release it.
524	*
525	* @returns VBox status code.
526	* @retval VINF_SUCCESS on success.
527	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
528	* backing or if the page as an active ALL access handler. The caller
529	* must fall back on using PGMPhysRead.
530	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
531	*
532	* @param pVM The cross context VM structure.
533	* @param GCPhys The guest physical address of the page that should be mapped.
534	* @param ppv Where to store the address corresponding to GCPhys.
535	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
536	*
537	* @remark Avoid calling this API from within critical sections (other than
538	* the PGM one) because of the deadlock risk.
539	* @thread Any.
540	*/
541	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrReadOnlyExternal(PVM pVM, RTGCPHYS GCPhys, void const **ppv, PPGMPAGEMAPLOCK pLock)
542	{
543	int rc = pgmLock(pVM);
544	AssertRCReturn(rc, rc);
545
546	/*
547	* Query the Physical TLB entry for the page (may fail).
548	*/
549	PPGMPAGEMAPTLBE pTlbe;
550	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
551	if (RT_SUCCESS(rc))
552	{
553	PPGMPAGE pPage = pTlbe->pPage;
554	#if 1
555	/* MMIO pages doesn't have any readable backing. */
556	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
557	rc = VERR_PGM_PHYS_PAGE_RESERVED;
558	#else
559	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
560	rc = VERR_PGM_PHYS_PAGE_RESERVED;
561	#endif
562	else
563	{
564	/*
565	* Now, just perform the locking and calculate the return address.
566	*/
567	PPGMPAGEMAP pMap = pTlbe->pMap;
568	if (pMap)
569	pMap->cRefs++;
570
571	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
572	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
573	{
574	if (cLocks == 0)
575	pVM->pgm.s.cReadLockedPages++;
576	PGM_PAGE_INC_READ_LOCKS(pPage);
577	}
578	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
579	{
580	PGM_PAGE_INC_READ_LOCKS(pPage);
581	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", GCPhys, pPage));
582	if (pMap)
583	pMap->cRefs++; /* Extra ref to prevent it from going away. */
584	}
585
586	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
587	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
588	pLock->pvMap = pMap;
589	}
590	}
591
592	pgmUnlock(pVM);
593	return rc;
594	}
595
596
597	/**
598	* Requests the mapping of multiple guest page into ring-3, external threads.
599	*
600	* When you're done with the pages, call PGMPhysBulkReleasePageMappingLock()
601	* ASAP to release them.
602	*
603	* This API will assume your intention is to write to the pages, and will
604	* therefore replace shared and zero pages. If you do not intend to modify the
605	* pages, use the PGMR3PhysBulkGCPhys2CCPtrReadOnlyExternal() API.
606	*
607	* @returns VBox status code.
608	* @retval VINF_SUCCESS on success.
609	* @retval VERR_PGM_PHYS_PAGE_RESERVED if any of the pages has no physical
610	* backing or if any of the pages the page has any active access
611	* handlers. The caller must fall back on using PGMR3PhysWriteExternal.
612	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if @a paGCPhysPages contains
613	* an invalid physical address.
614	*
615	* @param pVM The cross context VM structure.
616	* @param cPages Number of pages to lock.
617	* @param paGCPhysPages The guest physical address of the pages that
618	* should be mapped (@a cPages entries).
619	* @param papvPages Where to store the ring-3 mapping addresses
620	* corresponding to @a paGCPhysPages.
621	* @param paLocks Where to store the locking information that
622	* pfnPhysBulkReleasePageMappingLock needs (@a cPages
623	* in length).
624	*
625	* @remark Avoid calling this API from within critical sections (other than the
626	* PGM one) because of the deadlock risk when we have to delegating the
627	* task to an EMT.
628	* @thread Any.
629	*/
630	VMMR3DECL(int) PGMR3PhysBulkGCPhys2CCPtrExternal(PVM pVM, uint32_t cPages, PCRTGCPHYS paGCPhysPages,
631	void **papvPages, PPGMPAGEMAPLOCK paLocks)
632	{
633	Assert(cPages > 0);
634	AssertPtr(papvPages);
635	AssertPtr(paLocks);
636
637	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
638
639	int rc = pgmLock(pVM);
640	AssertRCReturn(rc, rc);
641
642	/*
643	* Lock the pages one by one.
644	* The loop body is similar to PGMR3PhysGCPhys2CCPtrExternal.
645	*/
646	int32_t cNextYield = 128;
647	uint32_t iPage;
648	for (iPage = 0; iPage < cPages; iPage++)
649	{
650	if (--cNextYield > 0)
651	{ /* likely */ }
652	else
653	{
654	pgmUnlock(pVM);
655	ASMNopPause();
656	pgmLock(pVM);
657	cNextYield = 128;
658	}
659
660	/*
661	* Query the Physical TLB entry for the page (may fail).
662	*/
663	PPGMPAGEMAPTLBE pTlbe;
664	rc = pgmPhysPageQueryTlbe(pVM, paGCPhysPages[iPage], &pTlbe);
665	if (RT_SUCCESS(rc))
666	{ }
667	else
668	break;
669	PPGMPAGE pPage = pTlbe->pPage;
670
671	/*
672	* No MMIO or active access handlers.
673	*/
674	if ( !PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage)
675	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
676	{ }
677	else
678	{
679	rc = VERR_PGM_PHYS_PAGE_RESERVED;
680	break;
681	}
682
683	/*
684	* The page must be in the allocated state and not be a dirty pool page.
685	* We can handle converting a write monitored page to an allocated one, but
686	* anything more complicated must be delegated to an EMT.
687	*/
688	bool fDelegateToEmt = false;
689	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_ALLOCATED)
690	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
691	fDelegateToEmt = pgmPoolIsDirtyPage(pVM, paGCPhysPages[iPage]);
692	#else
693	fDelegateToEmt = false;
694	#endif
695	else if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
696	{
697	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
698	if (!pgmPoolIsDirtyPage(pVM, paGCPhysPages[iPage]))
699	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage, paGCPhysPages[iPage]);
700	else
701	fDelegateToEmt = true;
702	#endif
703	}
704	else
705	fDelegateToEmt = true;
706	if (!fDelegateToEmt)
707	{ }
708	else
709	{
710	/* We could do this delegation in bulk, but considered too much work vs gain. */
711	pgmUnlock(pVM);
712	rc = VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
713	pVM, &paGCPhysPages[iPage], &papvPages[iPage], &paLocks[iPage]);
714	pgmLock(pVM);
715	if (RT_FAILURE(rc))
716	break;
717	cNextYield = 128;
718	}
719
720	/*
721	* Now, just perform the locking and address calculation.
722	*/
723	PPGMPAGEMAP pMap = pTlbe->pMap;
724	if (pMap)
725	pMap->cRefs++;
726
727	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
728	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
729	{
730	if (cLocks == 0)
731	pVM->pgm.s.cWriteLockedPages++;
732	PGM_PAGE_INC_WRITE_LOCKS(pPage);
733	}
734	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
735	{
736	PGM_PAGE_INC_WRITE_LOCKS(pPage);
737	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", paGCPhysPages[iPage], pPage));
738	if (pMap)
739	pMap->cRefs++; /* Extra ref to prevent it from going away. */
740	}
741
742	papvPages[iPage] = (void *)((uintptr_t)pTlbe->pv \| (uintptr_t)(paGCPhysPages[iPage] & PAGE_OFFSET_MASK));
743	paLocks[iPage].uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
744	paLocks[iPage].pvMap = pMap;
745	}
746
747	pgmUnlock(pVM);
748
749	/*
750	* On failure we must unlock any pages we managed to get already.
751	*/
752	if (RT_FAILURE(rc) && iPage > 0)
753	PGMPhysBulkReleasePageMappingLocks(pVM, iPage, paLocks);
754
755	return rc;
756	}
757
758
759	/**
760	* Requests the mapping of multiple guest page into ring-3, for reading only,
761	* external threads.
762	*
763	* When you're done with the pages, call PGMPhysReleasePageMappingLock() ASAP
764	* to release them.
765	*
766	* @returns VBox status code.
767	* @retval VINF_SUCCESS on success.
768	* @retval VERR_PGM_PHYS_PAGE_RESERVED if any of the pages has no physical
769	* backing or if any of the pages the page has an active ALL access
770	* handler. The caller must fall back on using PGMR3PhysWriteExternal.
771	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if @a paGCPhysPages contains
772	* an invalid physical address.
773	*
774	* @param pVM The cross context VM structure.
775	* @param cPages Number of pages to lock.
776	* @param paGCPhysPages The guest physical address of the pages that
777	* should be mapped (@a cPages entries).
778	* @param papvPages Where to store the ring-3 mapping addresses
779	* corresponding to @a paGCPhysPages.
780	* @param paLocks Where to store the lock information that
781	* pfnPhysReleasePageMappingLock needs (@a cPages
782	* in length).
783	*
784	* @remark Avoid calling this API from within critical sections (other than
785	* the PGM one) because of the deadlock risk.
786	* @thread Any.
787	*/
788	VMMR3DECL(int) PGMR3PhysBulkGCPhys2CCPtrReadOnlyExternal(PVM pVM, uint32_t cPages, PCRTGCPHYS paGCPhysPages,
789	void const **papvPages, PPGMPAGEMAPLOCK paLocks)
790	{
791	Assert(cPages > 0);
792	AssertPtr(papvPages);
793	AssertPtr(paLocks);
794
795	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
796
797	int rc = pgmLock(pVM);
798	AssertRCReturn(rc, rc);
799
800	/*
801	* Lock the pages one by one.
802	* The loop body is similar to PGMR3PhysGCPhys2CCPtrReadOnlyExternal.
803	*/
804	int32_t cNextYield = 256;
805	uint32_t iPage;
806	for (iPage = 0; iPage < cPages; iPage++)
807	{
808	if (--cNextYield > 0)
809	{ /* likely */ }
810	else
811	{
812	pgmUnlock(pVM);
813	ASMNopPause();
814	pgmLock(pVM);
815	cNextYield = 256;
816	}
817
818	/*
819	* Query the Physical TLB entry for the page (may fail).
820	*/
821	PPGMPAGEMAPTLBE pTlbe;
822	rc = pgmPhysPageQueryTlbe(pVM, paGCPhysPages[iPage], &pTlbe);
823	if (RT_SUCCESS(rc))
824	{ }
825	else
826	break;
827	PPGMPAGE pPage = pTlbe->pPage;
828
829	/*
830	* No MMIO or active all access handlers, everything else can be accessed.
831	*/
832	if ( !PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage)
833	&& !PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
834	{ }
835	else
836	{
837	rc = VERR_PGM_PHYS_PAGE_RESERVED;
838	break;
839	}
840
841	/*
842	* Now, just perform the locking and address calculation.
843	*/
844	PPGMPAGEMAP pMap = pTlbe->pMap;
845	if (pMap)
846	pMap->cRefs++;
847
848	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
849	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
850	{
851	if (cLocks == 0)
852	pVM->pgm.s.cReadLockedPages++;
853	PGM_PAGE_INC_READ_LOCKS(pPage);
854	}
855	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
856	{
857	PGM_PAGE_INC_READ_LOCKS(pPage);
858	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", paGCPhysPages[iPage], pPage));
859	if (pMap)
860	pMap->cRefs++; /* Extra ref to prevent it from going away. */
861	}
862
863	papvPages[iPage] = (void *)((uintptr_t)pTlbe->pv \| (uintptr_t)(paGCPhysPages[iPage] & PAGE_OFFSET_MASK));
864	paLocks[iPage].uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
865	paLocks[iPage].pvMap = pMap;
866	}
867
868	pgmUnlock(pVM);
869
870	/*
871	* On failure we must unlock any pages we managed to get already.
872	*/
873	if (RT_FAILURE(rc) && iPage > 0)
874	PGMPhysBulkReleasePageMappingLocks(pVM, iPage, paLocks);
875
876	return rc;
877	}
878
879
880	#define MAKE_LEAF(a_pNode) \
881	do { \
882	(a_pNode)->pLeftR3 = NIL_RTR3PTR; \
883	(a_pNode)->pRightR3 = NIL_RTR3PTR; \
884	(a_pNode)->pLeftR0 = NIL_RTR0PTR; \
885	(a_pNode)->pRightR0 = NIL_RTR0PTR; \
886	(a_pNode)->pLeftRC = NIL_RTRCPTR; \
887	(a_pNode)->pRightRC = NIL_RTRCPTR; \
888	} while (0)
889
890	#define INSERT_LEFT(a_pParent, a_pNode) \
891	do { \
892	(a_pParent)->pLeftR3 = (a_pNode); \
893	(a_pParent)->pLeftR0 = (a_pNode)->pSelfR0; \
894	(a_pParent)->pLeftRC = (a_pNode)->pSelfRC; \
895	} while (0)
896	#define INSERT_RIGHT(a_pParent, a_pNode) \
897	do { \
898	(a_pParent)->pRightR3 = (a_pNode); \
899	(a_pParent)->pRightR0 = (a_pNode)->pSelfR0; \
900	(a_pParent)->pRightRC = (a_pNode)->pSelfRC; \
901	} while (0)
902
903
904	/**
905	* Recursive tree builder.
906	*
907	* @param ppRam Pointer to the iterator variable.
908	* @param iDepth The current depth. Inserts a leaf node if 0.
909	*/
910	static PPGMRAMRANGE pgmR3PhysRebuildRamRangeSearchTreesRecursively(PPGMRAMRANGE *ppRam, int iDepth)
911	{
912	PPGMRAMRANGE pRam;
913	if (iDepth <= 0)
914	{
915	/*
916	* Leaf node.
917	*/
918	pRam = *ppRam;
919	if (pRam)
920	{
921	*ppRam = pRam->pNextR3;
922	MAKE_LEAF(pRam);
923	}
924	}
925	else
926	{
927
928	/*
929	* Intermediate node.
930	*/
931	PPGMRAMRANGE pLeft = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
932
933	pRam = *ppRam;
934	if (!pRam)
935	return pLeft;
936	*ppRam = pRam->pNextR3;
937	MAKE_LEAF(pRam);
938	INSERT_LEFT(pRam, pLeft);
939
940	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
941	if (pRight)
942	INSERT_RIGHT(pRam, pRight);
943	}
944	return pRam;
945	}
946
947
948	/**
949	* Rebuilds the RAM range search trees.
950	*
951	* @param pVM The cross context VM structure.
952	*/
953	static void pgmR3PhysRebuildRamRangeSearchTrees(PVM pVM)
954	{
955
956	/*
957	* Create the reasonably balanced tree in a sequential fashion.
958	* For simplicity (laziness) we use standard recursion here.
959	*/
960	int iDepth = 0;
961	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
962	PPGMRAMRANGE pRoot = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, 0);
963	while (pRam)
964	{
965	PPGMRAMRANGE pLeft = pRoot;
966
967	pRoot = pRam;
968	pRam = pRam->pNextR3;
969	MAKE_LEAF(pRoot);
970	INSERT_LEFT(pRoot, pLeft);
971
972	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, iDepth);
973	if (pRight)
974	INSERT_RIGHT(pRoot, pRight);
975	/** @todo else: rotate the tree. */
976
977	iDepth++;
978	}
979
980	pVM->pgm.s.pRamRangeTreeR3 = pRoot;
981	pVM->pgm.s.pRamRangeTreeR0 = pRoot ? pRoot->pSelfR0 : NIL_RTR0PTR;
982	pVM->pgm.s.pRamRangeTreeRC = pRoot ? pRoot->pSelfRC : NIL_RTRCPTR;
983
984	#ifdef VBOX_STRICT
985	/*
986	* Verify that the above code works.
987	*/
988	unsigned cRanges = 0;
989	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
990	cRanges++;
991	Assert(cRanges > 0);
992
993	unsigned cMaxDepth = ASMBitLastSetU32(cRanges);
994	if ((1U << cMaxDepth) < cRanges)
995	cMaxDepth++;
996
997	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
998	{
999	unsigned cDepth = 0;
1000	PPGMRAMRANGE pRam2 = pVM->pgm.s.pRamRangeTreeR3;
1001	for (;;)
1002	{
1003	if (pRam == pRam2)
1004	break;
1005	Assert(pRam2);
1006	if (pRam->GCPhys < pRam2->GCPhys)
1007	pRam2 = pRam2->pLeftR3;
1008	else
1009	pRam2 = pRam2->pRightR3;
1010	}
1011	AssertMsg(cDepth <= cMaxDepth, ("cDepth=%d cMaxDepth=%d\n", cDepth, cMaxDepth));
1012	}
1013	#endif /* VBOX_STRICT */
1014	}
1015
1016	#undef MAKE_LEAF
1017	#undef INSERT_LEFT
1018	#undef INSERT_RIGHT
1019
1020	/**
1021	* Relinks the RAM ranges using the pSelfRC and pSelfR0 pointers.
1022	*
1023	* Called when anything was relocated.
1024	*
1025	* @param pVM The cross context VM structure.
1026	*/
1027	void pgmR3PhysRelinkRamRanges(PVM pVM)
1028	{
1029	PPGMRAMRANGE pCur;
1030
1031	#ifdef VBOX_STRICT
1032	for (pCur = pVM->pgm.s.pRamRangesXR3; pCur; pCur = pCur->pNextR3)
1033	{
1034	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfR0 == MMHyperCCToR0(pVM, pCur));
1035	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfRC == MMHyperCCToRC(pVM, pCur));
1036	Assert((pCur->GCPhys & PAGE_OFFSET_MASK) == 0);
1037	Assert((pCur->GCPhysLast & PAGE_OFFSET_MASK) == PAGE_OFFSET_MASK);
1038	Assert((pCur->cb & PAGE_OFFSET_MASK) == 0);
1039	Assert(pCur->cb == pCur->GCPhysLast - pCur->GCPhys + 1);
1040	for (PPGMRAMRANGE pCur2 = pVM->pgm.s.pRamRangesXR3; pCur2; pCur2 = pCur2->pNextR3)
1041	Assert( pCur2 == pCur
1042	\|\| strcmp(pCur2->pszDesc, pCur->pszDesc)); /** @todo fix MMIO ranges!! */
1043	}
1044	#endif
1045
1046	pCur = pVM->pgm.s.pRamRangesXR3;
1047	if (pCur)
1048	{
1049	pVM->pgm.s.pRamRangesXR0 = pCur->pSelfR0;
1050	pVM->pgm.s.pRamRangesXRC = pCur->pSelfRC;
1051
1052	for (; pCur->pNextR3; pCur = pCur->pNextR3)
1053	{
1054	pCur->pNextR0 = pCur->pNextR3->pSelfR0;
1055	pCur->pNextRC = pCur->pNextR3->pSelfRC;
1056	}
1057
1058	Assert(pCur->pNextR0 == NIL_RTR0PTR);
1059	Assert(pCur->pNextRC == NIL_RTRCPTR);
1060	}
1061	else
1062	{
1063	Assert(pVM->pgm.s.pRamRangesXR0 == NIL_RTR0PTR);
1064	Assert(pVM->pgm.s.pRamRangesXRC == NIL_RTRCPTR);
1065	}
1066	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1067
1068	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1069	}
1070
1071
1072	/**
1073	* Links a new RAM range into the list.
1074	*
1075	* @param pVM The cross context VM structure.
1076	* @param pNew Pointer to the new list entry.
1077	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
1078	*/
1079	static void pgmR3PhysLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, PPGMRAMRANGE pPrev)
1080	{
1081	AssertMsg(pNew->pszDesc, ("%RGp-%RGp\n", pNew->GCPhys, pNew->GCPhysLast));
1082	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfR0 == MMHyperCCToR0(pVM, pNew));
1083	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfRC == MMHyperCCToRC(pVM, pNew));
1084
1085	pgmLock(pVM);
1086
1087	PPGMRAMRANGE pRam = pPrev ? pPrev->pNextR3 : pVM->pgm.s.pRamRangesXR3;
1088	pNew->pNextR3 = pRam;
1089	pNew->pNextR0 = pRam ? pRam->pSelfR0 : NIL_RTR0PTR;
1090	pNew->pNextRC = pRam ? pRam->pSelfRC : NIL_RTRCPTR;
1091
1092	if (pPrev)
1093	{
1094	pPrev->pNextR3 = pNew;
1095	pPrev->pNextR0 = pNew->pSelfR0;
1096	pPrev->pNextRC = pNew->pSelfRC;
1097	}
1098	else
1099	{
1100	pVM->pgm.s.pRamRangesXR3 = pNew;
1101	pVM->pgm.s.pRamRangesXR0 = pNew->pSelfR0;
1102	pVM->pgm.s.pRamRangesXRC = pNew->pSelfRC;
1103	}
1104	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1105
1106	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1107	pgmUnlock(pVM);
1108	}
1109
1110
1111	/**
1112	* Unlink an existing RAM range from the list.
1113	*
1114	* @param pVM The cross context VM structure.
1115	* @param pRam Pointer to the new list entry.
1116	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
1117	*/
1118	static void pgmR3PhysUnlinkRamRange2(PVM pVM, PPGMRAMRANGE pRam, PPGMRAMRANGE pPrev)
1119	{
1120	Assert(pPrev ? pPrev->pNextR3 == pRam : pVM->pgm.s.pRamRangesXR3 == pRam);
1121	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfR0 == MMHyperCCToR0(pVM, pRam));
1122	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfRC == MMHyperCCToRC(pVM, pRam));
1123
1124	pgmLock(pVM);
1125
1126	PPGMRAMRANGE pNext = pRam->pNextR3;
1127	if (pPrev)
1128	{
1129	pPrev->pNextR3 = pNext;
1130	pPrev->pNextR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
1131	pPrev->pNextRC = pNext ? pNext->pSelfRC : NIL_RTRCPTR;
1132	}
1133	else
1134	{
1135	Assert(pVM->pgm.s.pRamRangesXR3 == pRam);
1136	pVM->pgm.s.pRamRangesXR3 = pNext;
1137	pVM->pgm.s.pRamRangesXR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
1138	pVM->pgm.s.pRamRangesXRC = pNext ? pNext->pSelfRC : NIL_RTRCPTR;
1139	}
1140	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
1141
1142	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
1143	pgmUnlock(pVM);
1144	}
1145
1146
1147	/**
1148	* Unlink an existing RAM range from the list.
1149	*
1150	* @param pVM The cross context VM structure.
1151	* @param pRam Pointer to the new list entry.
1152	*/
1153	static void pgmR3PhysUnlinkRamRange(PVM pVM, PPGMRAMRANGE pRam)
1154	{
1155	pgmLock(pVM);
1156
1157	/* find prev. */
1158	PPGMRAMRANGE pPrev = NULL;
1159	PPGMRAMRANGE pCur = pVM->pgm.s.pRamRangesXR3;
1160	while (pCur != pRam)
1161	{
1162	pPrev = pCur;
1163	pCur = pCur->pNextR3;
1164	}
1165	AssertFatal(pCur);
1166
1167	pgmR3PhysUnlinkRamRange2(pVM, pRam, pPrev);
1168	pgmUnlock(pVM);
1169	}
1170
1171
1172	/**
1173	* Frees a range of pages, replacing them with ZERO pages of the specified type.
1174	*
1175	* @returns VBox status code.
1176	* @param pVM The cross context VM structure.
1177	* @param pRam The RAM range in which the pages resides.
1178	* @param GCPhys The address of the first page.
1179	* @param GCPhysLast The address of the last page.
1180	* @param enmType The page type to replace then with.
1181	*/
1182	static int pgmR3PhysFreePageRange(PVM pVM, PPGMRAMRANGE pRam, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast, PGMPAGETYPE enmType)
1183	{
1184	PGM_LOCK_ASSERT_OWNER(pVM);
1185	uint32_t cPendingPages = 0;
1186	PGMMFREEPAGESREQ pReq;
1187	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1188	AssertLogRelRCReturn(rc, rc);
1189
1190	/* Iterate the pages. */
1191	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
1192	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
1193	while (cPagesLeft-- > 0)
1194	{
1195	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys, enmType);
1196	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
1197
1198	PGM_PAGE_SET_TYPE(pVM, pPageDst, enmType);
1199
1200	GCPhys += PAGE_SIZE;
1201	pPageDst++;
1202	}
1203
1204	if (cPendingPages)
1205	{
1206	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1207	AssertLogRelRCReturn(rc, rc);
1208	}
1209	GMMR3FreePagesCleanup(pReq);
1210
1211	return rc;
1212	}
1213
1214	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
1215
1216	/**
1217	* Rendezvous callback used by PGMR3ChangeMemBalloon that changes the memory balloon size
1218	*
1219	* This is only called on one of the EMTs while the other ones are waiting for
1220	* it to complete this function.
1221	*
1222	* @returns VINF_SUCCESS (VBox strict status code).
1223	* @param pVM The cross context VM structure.
1224	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
1225	* @param pvUser User parameter
1226	*/
1227	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysChangeMemBalloonRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
1228	{
1229	uintptr_t paUser = (uintptr_t )pvUser;
1230	bool fInflate = !!paUser[0];
1231	unsigned cPages = paUser[1];
1232	RTGCPHYS paPhysPage = (RTGCPHYS )paUser[2];
1233	uint32_t cPendingPages = 0;
1234	PGMMFREEPAGESREQ pReq;
1235	int rc;
1236
1237	Log(("pgmR3PhysChangeMemBalloonRendezvous: %s %x pages\n", (fInflate) ? "inflate" : "deflate", cPages));
1238	pgmLock(pVM);
1239
1240	if (fInflate)
1241	{
1242	/* Flush the PGM pool cache as we might have stale references to pages that we just freed. */
1243	pgmR3PoolClearAllRendezvous(pVM, pVCpu, NULL);
1244
1245	/* Replace pages with ZERO pages. */
1246	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1247	if (RT_FAILURE(rc))
1248	{
1249	pgmUnlock(pVM);
1250	AssertLogRelRC(rc);
1251	return rc;
1252	}
1253
1254	/* Iterate the pages. */
1255	for (unsigned i = 0; i < cPages; i++)
1256	{
1257	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
1258	if ( pPage == NULL
1259	\|\| PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM)
1260	{
1261	Log(("pgmR3PhysChangeMemBalloonRendezvous: invalid physical page %RGp pPage->u3Type=%d\n", paPhysPage[i], pPage ? PGM_PAGE_GET_TYPE(pPage) : 0));
1262	break;
1263	}
1264
1265	LogFlow(("balloon page: %RGp\n", paPhysPage[i]));
1266
1267	/* Flush the shadow PT if this page was previously used as a guest page table. */
1268	pgmPoolFlushPageByGCPhys(pVM, paPhysPage[i]);
1269
1270	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, paPhysPage[i], (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage));
1271	if (RT_FAILURE(rc))
1272	{
1273	pgmUnlock(pVM);
1274	AssertLogRelRC(rc);
1275	return rc;
1276	}
1277	Assert(PGM_PAGE_IS_ZERO(pPage));
1278	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_BALLOONED);
1279	}
1280
1281	if (cPendingPages)
1282	{
1283	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1284	if (RT_FAILURE(rc))
1285	{
1286	pgmUnlock(pVM);
1287	AssertLogRelRC(rc);
1288	return rc;
1289	}
1290	}
1291	GMMR3FreePagesCleanup(pReq);
1292	}
1293	else
1294	{
1295	/* Iterate the pages. */
1296	for (unsigned i = 0; i < cPages; i++)
1297	{
1298	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
1299	AssertBreak(pPage && PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM);
1300
1301	LogFlow(("Free ballooned page: %RGp\n", paPhysPage[i]));
1302
1303	Assert(PGM_PAGE_IS_BALLOONED(pPage));
1304
1305	/* Change back to zero page. (NEM does not need to be informed.) */
1306	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
1307	}
1308
1309	/* Note that we currently do not map any ballooned pages in our shadow page tables, so no need to flush the pgm pool. */
1310	}
1311
1312	/* Notify GMM about the balloon change. */
1313	rc = GMMR3BalloonedPages(pVM, (fInflate) ? GMMBALLOONACTION_INFLATE : GMMBALLOONACTION_DEFLATE, cPages);
1314	if (RT_SUCCESS(rc))
1315	{
1316	if (!fInflate)
1317	{
1318	Assert(pVM->pgm.s.cBalloonedPages >= cPages);
1319	pVM->pgm.s.cBalloonedPages -= cPages;
1320	}
1321	else
1322	pVM->pgm.s.cBalloonedPages += cPages;
1323	}
1324
1325	pgmUnlock(pVM);
1326
1327	/* Flush the recompiler's TLB as well. */
1328	for (VMCPUID i = 0; i < pVM->cCpus; i++)
1329	CPUMSetChangedFlags(&pVM->aCpus[i], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1330
1331	AssertLogRelRC(rc);
1332	return rc;
1333	}
1334
1335
1336	/**
1337	* Frees a range of ram pages, replacing them with ZERO pages; helper for PGMR3PhysFreeRamPages
1338	*
1339	* @returns VBox status code.
1340	* @param pVM The cross context VM structure.
1341	* @param fInflate Inflate or deflate memory balloon
1342	* @param cPages Number of pages to free
1343	* @param paPhysPage Array of guest physical addresses
1344	*/
1345	static DECLCALLBACK(void) pgmR3PhysChangeMemBalloonHelper(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1346	{
1347	uintptr_t paUser[3];
1348
1349	paUser[0] = fInflate;
1350	paUser[1] = cPages;
1351	paUser[2] = (uintptr_t)paPhysPage;
1352	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1353	AssertRC(rc);
1354
1355	/* Made a copy in PGMR3PhysFreeRamPages; free it here. */
1356	RTMemFree(paPhysPage);
1357	}
1358
1359	#endif /* 64-bit host && (Windows \|\| Solaris \|\| Linux \|\| FreeBSD) */
1360
1361	/**
1362	* Inflate or deflate a memory balloon
1363	*
1364	* @returns VBox status code.
1365	* @param pVM The cross context VM structure.
1366	* @param fInflate Inflate or deflate memory balloon
1367	* @param cPages Number of pages to free
1368	* @param paPhysPage Array of guest physical addresses
1369	*/
1370	VMMR3DECL(int) PGMR3PhysChangeMemBalloon(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1371	{
1372	/* This must match GMMR0Init; currently we only support memory ballooning on all 64-bit hosts except Mac OS X */
1373	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
1374	int rc;
1375
1376	/* Older additions (ancient non-functioning balloon code) pass wrong physical addresses. */
1377	AssertReturn(!(paPhysPage[0] & 0xfff), VERR_INVALID_PARAMETER);
1378
1379	/* We own the IOM lock here and could cause a deadlock by waiting for another VCPU that is blocking on the IOM lock.
1380	* In the SMP case we post a request packet to postpone the job.
1381	*/
1382	if (pVM->cCpus > 1)
1383	{
1384	unsigned cbPhysPage = cPages * sizeof(paPhysPage[0]);
1385	RTGCPHYS paPhysPageCopy = (RTGCPHYS )RTMemAlloc(cbPhysPage);
1386	AssertReturn(paPhysPageCopy, VERR_NO_MEMORY);
1387
1388	memcpy(paPhysPageCopy, paPhysPage, cbPhysPage);
1389
1390	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysChangeMemBalloonHelper, 4, pVM, fInflate, cPages, paPhysPageCopy);
1391	AssertRC(rc);
1392	}
1393	else
1394	{
1395	uintptr_t paUser[3];
1396
1397	paUser[0] = fInflate;
1398	paUser[1] = cPages;
1399	paUser[2] = (uintptr_t)paPhysPage;
1400	rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1401	AssertRC(rc);
1402	}
1403	return rc;
1404
1405	#else
1406	NOREF(pVM); NOREF(fInflate); NOREF(cPages); NOREF(paPhysPage);
1407	return VERR_NOT_IMPLEMENTED;
1408	#endif
1409	}
1410
1411
1412	/**
1413	* Rendezvous callback used by PGMR3WriteProtectRAM that write protects all
1414	* physical RAM.
1415	*
1416	* This is only called on one of the EMTs while the other ones are waiting for
1417	* it to complete this function.
1418	*
1419	* @returns VINF_SUCCESS (VBox strict status code).
1420	* @param pVM The cross context VM structure.
1421	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
1422	* @param pvUser User parameter, unused.
1423	*/
1424	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysWriteProtectRAMRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
1425	{
1426	int rc = VINF_SUCCESS;
1427	NOREF(pvUser); NOREF(pVCpu);
1428
1429	pgmLock(pVM);
1430	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
1431	pgmPoolResetDirtyPages(pVM);
1432	#endif
1433
1434	/** @todo pointless to write protect the physical page pointed to by RSP. */
1435
1436	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
1437	pRam;
1438	pRam = pRam->CTX_SUFF(pNext))
1439	{
1440	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
1441	for (uint32_t iPage = 0; iPage < cPages; iPage++)
1442	{
1443	PPGMPAGE pPage = &pRam->aPages[iPage];
1444	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
1445
1446	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
1447	\|\| enmPageType == PGMPAGETYPE_MMIO2)
1448	{
1449	/*
1450	* A RAM page.
1451	*/
1452	switch (PGM_PAGE_GET_STATE(pPage))
1453	{
1454	case PGM_PAGE_STATE_ALLOCATED:
1455	/** @todo Optimize this: Don't always re-enable write
1456	* monitoring if the page is known to be very busy. */
1457	if (PGM_PAGE_IS_WRITTEN_TO(pPage))
1458	{
1459	PGM_PAGE_CLEAR_WRITTEN_TO(pVM, pPage);
1460	/* Remember this dirty page for the next (memory) sync. */
1461	PGM_PAGE_SET_FT_DIRTY(pPage);
1462	}
1463
1464	pgmPhysPageWriteMonitor(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
1465	break;
1466
1467	case PGM_PAGE_STATE_SHARED:
1468	AssertFailed();
1469	break;
1470
1471	case PGM_PAGE_STATE_WRITE_MONITORED: /* nothing to change. */
1472	default:
1473	break;
1474	}
1475	}
1476	}
1477	}
1478	pgmR3PoolWriteProtectPages(pVM);
1479	PGM_INVL_ALL_VCPU_TLBS(pVM);
1480	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
1481	CPUMSetChangedFlags(&pVM->aCpus[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1482
1483	pgmUnlock(pVM);
1484	return rc;
1485	}
1486
1487	/**
1488	* Protect all physical RAM to monitor writes
1489	*
1490	* @returns VBox status code.
1491	* @param pVM The cross context VM structure.
1492	*/
1493	VMMR3DECL(int) PGMR3PhysWriteProtectRAM(PVM pVM)
1494	{
1495	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1496
1497	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysWriteProtectRAMRendezvous, NULL);
1498	AssertRC(rc);
1499	return rc;
1500	}
1501
1502	/**
1503	* Enumerate all dirty FT pages.
1504	*
1505	* @returns VBox status code.
1506	* @param pVM The cross context VM structure.
1507	* @param pfnEnum Enumerate callback handler.
1508	* @param pvUser Enumerate callback handler parameter.
1509	*/
1510	VMMR3DECL(int) PGMR3PhysEnumDirtyFTPages(PVM pVM, PFNPGMENUMDIRTYFTPAGES pfnEnum, void *pvUser)
1511	{
1512	int rc = VINF_SUCCESS;
1513
1514	pgmLock(pVM);
1515	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
1516	pRam;
1517	pRam = pRam->CTX_SUFF(pNext))
1518	{
1519	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
1520	for (uint32_t iPage = 0; iPage < cPages; iPage++)
1521	{
1522	PPGMPAGE pPage = &pRam->aPages[iPage];
1523	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
1524
1525	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
1526	\|\| enmPageType == PGMPAGETYPE_MMIO2)
1527	{
1528	/*
1529	* A RAM page.
1530	*/
1531	switch (PGM_PAGE_GET_STATE(pPage))
1532	{
1533	case PGM_PAGE_STATE_ALLOCATED:
1534	case PGM_PAGE_STATE_WRITE_MONITORED:
1535	if ( !PGM_PAGE_IS_WRITTEN_TO(pPage) /* not very recently updated? */
1536	&& PGM_PAGE_IS_FT_DIRTY(pPage))
1537	{
1538	uint32_t cbPageRange = PAGE_SIZE;
1539	uint32_t iPageClean = iPage + 1;
1540	RTGCPHYS GCPhysPage = pRam->GCPhys + iPage * PAGE_SIZE;
1541	uint8_t *pu8Page = NULL;
1542	PGMPAGEMAPLOCK Lock;
1543
1544	/* Find the next clean page, so we can merge adjacent dirty pages. */
1545	for (; iPageClean < cPages; iPageClean++)
1546	{
1547	PPGMPAGE pPageNext = &pRam->aPages[iPageClean];
1548	if ( RT_UNLIKELY(PGM_PAGE_GET_TYPE(pPageNext) != PGMPAGETYPE_RAM)
1549	\|\| PGM_PAGE_GET_STATE(pPageNext) != PGM_PAGE_STATE_ALLOCATED
1550	\|\| PGM_PAGE_IS_WRITTEN_TO(pPageNext)
1551	\|\| !PGM_PAGE_IS_FT_DIRTY(pPageNext)
1552	/* Crossing a chunk boundary? */
1553	\|\| (GCPhysPage & GMM_PAGEID_IDX_MASK) != ((GCPhysPage + cbPageRange) & GMM_PAGEID_IDX_MASK)
1554	)
1555	break;
1556
1557	cbPageRange += PAGE_SIZE;
1558	}
1559
1560	rc = PGMPhysGCPhys2CCPtrReadOnly(pVM, GCPhysPage, (const void **)&pu8Page, &Lock);
1561	if (RT_SUCCESS(rc))
1562	{
1563	/** @todo this is risky; the range might be changed, but little choice as the sync
1564	* costs a lot of time. */
1565	pgmUnlock(pVM);
1566	pfnEnum(pVM, GCPhysPage, pu8Page, cbPageRange, pvUser);
1567	pgmLock(pVM);
1568	PGMPhysReleasePageMappingLock(pVM, &Lock);
1569	}
1570
1571	for (uint32_t iTmp = iPage; iTmp < iPageClean; iTmp++)
1572	PGM_PAGE_CLEAR_FT_DIRTY(&pRam->aPages[iTmp]);
1573	}
1574	break;
1575	}
1576	}
1577	}
1578	}
1579	pgmUnlock(pVM);
1580	return rc;
1581	}
1582
1583
1584	/**
1585	* Gets the number of ram ranges.
1586	*
1587	* @returns Number of ram ranges. Returns UINT32_MAX if @a pVM is invalid.
1588	* @param pVM The cross context VM structure.
1589	*/
1590	VMMR3DECL(uint32_t) PGMR3PhysGetRamRangeCount(PVM pVM)
1591	{
1592	VM_ASSERT_VALID_EXT_RETURN(pVM, UINT32_MAX);
1593
1594	pgmLock(pVM);
1595	uint32_t cRamRanges = 0;
1596	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext))
1597	cRamRanges++;
1598	pgmUnlock(pVM);
1599	return cRamRanges;
1600	}
1601
1602
1603	/**
1604	* Get information about a range.
1605	*
1606	* @returns VINF_SUCCESS or VERR_OUT_OF_RANGE.
1607	* @param pVM The cross context VM structure.
1608	* @param iRange The ordinal of the range.
1609	* @param pGCPhysStart Where to return the start of the range. Optional.
1610	* @param pGCPhysLast Where to return the address of the last byte in the
1611	* range. Optional.
1612	* @param ppszDesc Where to return the range description. Optional.
1613	* @param pfIsMmio Where to indicate that this is a pure MMIO range.
1614	* Optional.
1615	*/
1616	VMMR3DECL(int) PGMR3PhysGetRange(PVM pVM, uint32_t iRange, PRTGCPHYS pGCPhysStart, PRTGCPHYS pGCPhysLast,
1617	const char *ppszDesc, bool pfIsMmio)
1618	{
1619	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1620
1621	pgmLock(pVM);
1622	uint32_t iCurRange = 0;
1623	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext), iCurRange++)
1624	if (iCurRange == iRange)
1625	{
1626	if (pGCPhysStart)
1627	*pGCPhysStart = pCur->GCPhys;
1628	if (pGCPhysLast)
1629	*pGCPhysLast = pCur->GCPhysLast;
1630	if (ppszDesc)
1631	*ppszDesc = pCur->pszDesc;
1632	if (pfIsMmio)
1633	*pfIsMmio = !!(pCur->fFlags & PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO);
1634
1635	pgmUnlock(pVM);
1636	return VINF_SUCCESS;
1637	}
1638	pgmUnlock(pVM);
1639	return VERR_OUT_OF_RANGE;
1640	}
1641
1642
1643	/**
1644	* Query the amount of free memory inside VMMR0
1645	*
1646	* @returns VBox status code.
1647	* @param pUVM The user mode VM handle.
1648	* @param pcbAllocMem Where to return the amount of memory allocated
1649	* by VMs.
1650	* @param pcbFreeMem Where to return the amount of memory that is
1651	* allocated from the host but not currently used
1652	* by any VMs.
1653	* @param pcbBallonedMem Where to return the sum of memory that is
1654	* currently ballooned by the VMs.
1655	* @param pcbSharedMem Where to return the amount of memory that is
1656	* currently shared.
1657	*/
1658	VMMR3DECL(int) PGMR3QueryGlobalMemoryStats(PUVM pUVM, uint64_t pcbAllocMem, uint64_t pcbFreeMem,
1659	uint64_t pcbBallonedMem, uint64_t pcbSharedMem)
1660	{
1661	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1662	VM_ASSERT_VALID_EXT_RETURN(pUVM->pVM, VERR_INVALID_VM_HANDLE);
1663
1664	uint64_t cAllocPages = 0;
1665	uint64_t cFreePages = 0;
1666	uint64_t cBalloonPages = 0;
1667	uint64_t cSharedPages = 0;
1668	int rc = GMMR3QueryHypervisorMemoryStats(pUVM->pVM, &cAllocPages, &cFreePages, &cBalloonPages, &cSharedPages);
1669	AssertRCReturn(rc, rc);
1670
1671	if (pcbAllocMem)
1672	pcbAllocMem = cAllocPages _4K;
1673
1674	if (pcbFreeMem)
1675	pcbFreeMem = cFreePages _4K;
1676
1677	if (pcbBallonedMem)
1678	pcbBallonedMem = cBalloonPages _4K;
1679
1680	if (pcbSharedMem)
1681	pcbSharedMem = cSharedPages _4K;
1682
1683	Log(("PGMR3QueryVMMMemoryStats: all=%llx free=%llx ballooned=%llx shared=%llx\n",
1684	cAllocPages, cFreePages, cBalloonPages, cSharedPages));
1685	return VINF_SUCCESS;
1686	}
1687
1688
1689	/**
1690	* Query memory stats for the VM.
1691	*
1692	* @returns VBox status code.
1693	* @param pUVM The user mode VM handle.
1694	* @param pcbTotalMem Where to return total amount memory the VM may
1695	* possibly use.
1696	* @param pcbPrivateMem Where to return the amount of private memory
1697	* currently allocated.
1698	* @param pcbSharedMem Where to return the amount of actually shared
1699	* memory currently used by the VM.
1700	* @param pcbZeroMem Where to return the amount of memory backed by
1701	* zero pages.
1702	*
1703	* @remarks The total mem is normally larger than the sum of the three
1704	* components. There are two reasons for this, first the amount of
1705	* shared memory is what we're sure is shared instead of what could
1706	* possibly be shared with someone. Secondly, because the total may
1707	* include some pure MMIO pages that doesn't go into any of the three
1708	* sub-counts.
1709	*
1710	* @todo Why do we return reused shared pages instead of anything that could
1711	* potentially be shared? Doesn't this mean the first VM gets a much
1712	* lower number of shared pages?
1713	*/
1714	VMMR3DECL(int) PGMR3QueryMemoryStats(PUVM pUVM, uint64_t pcbTotalMem, uint64_t pcbPrivateMem,
1715	uint64_t pcbSharedMem, uint64_t pcbZeroMem)
1716	{
1717	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1718	PVM pVM = pUVM->pVM;
1719	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1720
1721	if (pcbTotalMem)
1722	pcbTotalMem = (uint64_t)pVM->pgm.s.cAllPages PAGE_SIZE;
1723
1724	if (pcbPrivateMem)
1725	pcbPrivateMem = (uint64_t)pVM->pgm.s.cPrivatePages PAGE_SIZE;
1726
1727	if (pcbSharedMem)
1728	pcbSharedMem = (uint64_t)pVM->pgm.s.cReusedSharedPages PAGE_SIZE;
1729
1730	if (pcbZeroMem)
1731	pcbZeroMem = (uint64_t)pVM->pgm.s.cZeroPages PAGE_SIZE;
1732
1733	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));
1734	return VINF_SUCCESS;
1735	}
1736
1737
1738	/**
1739	* PGMR3PhysRegisterRam worker that initializes and links a RAM range.
1740	*
1741	* @param pVM The cross context VM structure.
1742	* @param pNew The new RAM range.
1743	* @param GCPhys The address of the RAM range.
1744	* @param GCPhysLast The last address of the RAM range.
1745	* @param RCPtrNew The RC address if the range is floating. NIL_RTRCPTR
1746	* if in HMA.
1747	* @param R0PtrNew Ditto for R0.
1748	* @param pszDesc The description.
1749	* @param pPrev The previous RAM range (for linking).
1750	*/
1751	static void pgmR3PhysInitAndLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast,
1752	RTRCPTR RCPtrNew, RTR0PTR R0PtrNew, const char *pszDesc, PPGMRAMRANGE pPrev)
1753	{
1754	/*
1755	* Initialize the range.
1756	*/
1757	pNew->pSelfR0 = R0PtrNew != NIL_RTR0PTR ? R0PtrNew : MMHyperCCToR0(pVM, pNew);
1758	pNew->pSelfRC = RCPtrNew != NIL_RTRCPTR ? RCPtrNew : MMHyperCCToRC(pVM, pNew);
1759	pNew->GCPhys = GCPhys;
1760	pNew->GCPhysLast = GCPhysLast;
1761	pNew->cb = GCPhysLast - GCPhys + 1;
1762	pNew->pszDesc = pszDesc;
1763	pNew->fFlags = RCPtrNew != NIL_RTRCPTR ? PGM_RAM_RANGE_FLAGS_FLOATING : 0;
1764	pNew->pvR3 = NULL;
1765	pNew->paLSPages = NULL;
1766
1767	uint32_t const cPages = pNew->cb >> PAGE_SHIFT;
1768	RTGCPHYS iPage = cPages;
1769	while (iPage-- > 0)
1770	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_RAM);
1771
1772	/* Update the page count stats. */
1773	pVM->pgm.s.cZeroPages += cPages;
1774	pVM->pgm.s.cAllPages += cPages;
1775
1776	/*
1777	* Link it.
1778	*/
1779	pgmR3PhysLinkRamRange(pVM, pNew, pPrev);
1780	}
1781
1782
1783	/**
1784	* @callback_method_impl{FNPGMRELOCATE, Relocate a floating RAM range.}
1785	* @sa pgmR3PhysMMIO2ExRangeRelocate
1786	*/
1787	static DECLCALLBACK(bool) pgmR3PhysRamRangeRelocate(PVM pVM, RTGCPTR GCPtrOld, RTGCPTR GCPtrNew,
1788	PGMRELOCATECALL enmMode, void *pvUser)
1789	{
1790	PPGMRAMRANGE pRam = (PPGMRAMRANGE)pvUser;
1791	Assert(pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING);
1792	Assert(pRam->pSelfRC == GCPtrOld + PAGE_SIZE); RT_NOREF_PV(GCPtrOld);
1793
1794	switch (enmMode)
1795	{
1796	case PGMRELOCATECALL_SUGGEST:
1797	return true;
1798
1799	case PGMRELOCATECALL_RELOCATE:
1800	{
1801	/*
1802	* Update myself, then relink all the ranges and flush the RC TLB.
1803	*/
1804	pgmLock(pVM);
1805
1806	pRam->pSelfRC = (RTRCPTR)(GCPtrNew + PAGE_SIZE);
1807
1808	pgmR3PhysRelinkRamRanges(pVM);
1809	for (unsigned i = 0; i < PGM_RAMRANGE_TLB_ENTRIES; i++)
1810	pVM->pgm.s.apRamRangesTlbRC[i] = NIL_RTRCPTR;
1811
1812	pgmUnlock(pVM);
1813	return true;
1814	}
1815
1816	default:
1817	AssertFailedReturn(false);
1818	}
1819	}
1820
1821
1822	/**
1823	* PGMR3PhysRegisterRam worker that registers a high chunk.
1824	*
1825	* @returns VBox status code.
1826	* @param pVM The cross context VM structure.
1827	* @param GCPhys The address of the RAM.
1828	* @param cRamPages The number of RAM pages to register.
1829	* @param cbChunk The size of the PGMRAMRANGE guest mapping.
1830	* @param iChunk The chunk number.
1831	* @param pszDesc The RAM range description.
1832	* @param ppPrev Previous RAM range pointer. In/Out.
1833	*/
1834	static int pgmR3PhysRegisterHighRamChunk(PVM pVM, RTGCPHYS GCPhys, uint32_t cRamPages,
1835	uint32_t cbChunk, uint32_t iChunk, const char *pszDesc,
1836	PPGMRAMRANGE *ppPrev)
1837	{
1838	const char *pszDescChunk = iChunk == 0
1839	? pszDesc
1840	: MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s (#%u)", pszDesc, iChunk + 1);
1841	AssertReturn(pszDescChunk, VERR_NO_MEMORY);
1842
1843	/*
1844	* Allocate memory for the new chunk.
1845	*/
1846	size_t const cChunkPages = RT_ALIGN_Z(RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cRamPages]), PAGE_SIZE) >> PAGE_SHIFT;
1847	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
1848	AssertReturn(paChunkPages, VERR_NO_TMP_MEMORY);
1849	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
1850	void *pvChunk = NULL;
1851	int rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk,
1852	#if defined(VBOX_WITH_MORE_RING0_MEM_MAPPINGS)
1853	&R0PtrChunk,
1854	#elif defined(VBOX_WITH_2X_4GB_ADDR_SPACE)
1855	VM_IS_HM_OR_NEM_ENABLED(pVM) ? &R0PtrChunk : NULL,
1856	#else
1857	NULL,
1858	#endif
1859	paChunkPages);
1860	if (RT_SUCCESS(rc))
1861	{
1862	#if defined(VBOX_WITH_MORE_RING0_MEM_MAPPINGS)
1863	Assert(R0PtrChunk != NIL_RTR0PTR);
1864	#elif defined(VBOX_WITH_2X_4GB_ADDR_SPACE)
1865	if (!VM_IS_HM_OR_NEM_ENABLED(pVM))
1866	R0PtrChunk = NIL_RTR0PTR;
1867	#else
1868	R0PtrChunk = (uintptr_t)pvChunk;
1869	#endif
1870	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
1871
1872	PPGMRAMRANGE pNew = (PPGMRAMRANGE)pvChunk;
1873
1874	/*
1875	* Create a mapping and map the pages into it.
1876	* We push these in below the HMA.
1877	*/
1878	RTGCPTR GCPtrChunkMap = pVM->pgm.s.GCPtrPrevRamRangeMapping - cbChunk;
1879	rc = PGMR3MapPT(pVM, GCPtrChunkMap, cbChunk, 0 /fFlags/, pgmR3PhysRamRangeRelocate, pNew, pszDescChunk);
1880	if (RT_SUCCESS(rc))
1881	{
1882	pVM->pgm.s.GCPtrPrevRamRangeMapping = GCPtrChunkMap;
1883
1884	RTGCPTR const GCPtrChunk = GCPtrChunkMap + PAGE_SIZE;
1885	RTGCPTR GCPtrPage = GCPtrChunk;
1886	for (uint32_t iPage = 0; iPage < cChunkPages && RT_SUCCESS(rc); iPage++, GCPtrPage += PAGE_SIZE)
1887	rc = PGMMap(pVM, GCPtrPage, paChunkPages[iPage].Phys, PAGE_SIZE, 0);
1888	if (RT_SUCCESS(rc))
1889	{
1890	/*
1891	* Ok, init and link the range.
1892	*/
1893	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhys + ((RTGCPHYS)cRamPages << PAGE_SHIFT) - 1,
1894	(RTRCPTR)GCPtrChunk, R0PtrChunk, pszDescChunk, *ppPrev);
1895	*ppPrev = pNew;
1896	}
1897	}
1898
1899	if (RT_FAILURE(rc))
1900	SUPR3PageFreeEx(pvChunk, cChunkPages);
1901	}
1902
1903	RTMemTmpFree(paChunkPages);
1904	return rc;
1905	}
1906
1907
1908	/**
1909	* Sets up a range RAM.
1910	*
1911	* This will check for conflicting registrations, make a resource
1912	* reservation for the memory (with GMM), and setup the per-page
1913	* tracking structures (PGMPAGE).
1914	*
1915	* @returns VBox status code.
1916	* @param pVM The cross context VM structure.
1917	* @param GCPhys The physical address of the RAM.
1918	* @param cb The size of the RAM.
1919	* @param pszDesc The description - not copied, so, don't free or change it.
1920	*/
1921	VMMR3DECL(int) PGMR3PhysRegisterRam(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, const char *pszDesc)
1922	{
1923	/*
1924	* Validate input.
1925	*/
1926	Log(("PGMR3PhysRegisterRam: GCPhys=%RGp cb=%RGp pszDesc=%s\n", GCPhys, cb, pszDesc));
1927	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
1928	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
1929	AssertReturn(cb > 0, VERR_INVALID_PARAMETER);
1930	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
1931	AssertMsgReturn(GCPhysLast > GCPhys, ("The range wraps! GCPhys=%RGp cb=%RGp\n", GCPhys, cb), VERR_INVALID_PARAMETER);
1932	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
1933	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1934
1935	pgmLock(pVM);
1936
1937	/*
1938	* Find range location and check for conflicts.
1939	* (We don't lock here because the locking by EMT is only required on update.)
1940	*/
1941	PPGMRAMRANGE pPrev = NULL;
1942	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
1943	while (pRam && GCPhysLast >= pRam->GCPhys)
1944	{
1945	if ( GCPhysLast >= pRam->GCPhys
1946	&& GCPhys <= pRam->GCPhysLast)
1947	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
1948	GCPhys, GCPhysLast, pszDesc,
1949	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
1950	VERR_PGM_RAM_CONFLICT);
1951
1952	/* next */
1953	pPrev = pRam;
1954	pRam = pRam->pNextR3;
1955	}
1956
1957	/*
1958	* Register it with GMM (the API bitches).
1959	*/
1960	const RTGCPHYS cPages = cb >> PAGE_SHIFT;
1961	int rc = MMR3IncreaseBaseReservation(pVM, cPages);
1962	if (RT_FAILURE(rc))
1963	{
1964	pgmUnlock(pVM);
1965	return rc;
1966	}
1967
1968	if ( GCPhys >= _4G
1969	&& cPages > 256)
1970	{
1971	/*
1972	* The PGMRAMRANGE structures for the high memory can get very big.
1973	* In order to avoid SUPR3PageAllocEx allocation failures due to the
1974	* allocation size limit there and also to avoid being unable to find
1975	* guest mapping space for them, we split this memory up into 4MB in
1976	* (potential) raw-mode configs and 16MB chunks in forced AMD-V/VT-x
1977	* mode.
1978	*
1979	* The first and last page of each mapping are guard pages and marked
1980	* not-present. So, we've got 4186112 and 16769024 bytes available for
1981	* the PGMRAMRANGE structure.
1982	*
1983	* Note! The sizes used here will influence the saved state.
1984	*/
1985	uint32_t cbChunk;
1986	uint32_t cPagesPerChunk;
1987	if (!VM_IS_RAW_MODE_ENABLED(pVM))
1988	{
1989	cbChunk = 16U*_1M;
1990	cPagesPerChunk = 1048048; /* max ~1048059 */
1991	AssertCompile(sizeof(PGMRAMRANGE) + sizeof(PGMPAGE) * 1048048 < 16U_1M - PAGE_SIZE 2);
1992	}
1993	else
1994	{
1995	cbChunk = 4U*_1M;
1996	cPagesPerChunk = 261616; /* max ~261627 */
1997	AssertCompile(sizeof(PGMRAMRANGE) + sizeof(PGMPAGE) * 261616 < 4U_1M - PAGE_SIZE 2);
1998	}
1999	AssertRelease(RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPagesPerChunk]) + PAGE_SIZE * 2 <= cbChunk);
2000
2001	RTGCPHYS cPagesLeft = cPages;
2002	RTGCPHYS GCPhysChunk = GCPhys;
2003	uint32_t iChunk = 0;
2004	while (cPagesLeft > 0)
2005	{
2006	uint32_t cPagesInChunk = cPagesLeft;
2007	if (cPagesInChunk > cPagesPerChunk)
2008	cPagesInChunk = cPagesPerChunk;
2009
2010	rc = pgmR3PhysRegisterHighRamChunk(pVM, GCPhysChunk, cPagesInChunk, cbChunk, iChunk, pszDesc, &pPrev);
2011	AssertRCReturn(rc, rc);
2012
2013	/* advance */
2014	GCPhysChunk += (RTGCPHYS)cPagesInChunk << PAGE_SHIFT;
2015	cPagesLeft -= cPagesInChunk;
2016	iChunk++;
2017	}
2018	}
2019	else
2020	{
2021	/*
2022	* Allocate, initialize and link the new RAM range.
2023	*/
2024	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
2025	PPGMRAMRANGE pNew;
2026	rc = MMR3HyperAllocOnceNoRel(pVM, cbRamRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
2027	AssertLogRelMsgRCReturn(rc, ("cbRamRange=%zu\n", cbRamRange), rc);
2028
2029	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhysLast, NIL_RTRCPTR, NIL_RTR0PTR, pszDesc, pPrev);
2030	}
2031	pgmPhysInvalidatePageMapTLB(pVM);
2032
2033	/*
2034	* Notify NEM while holding the lock (experimental) and REM without (like always).
2035	*/
2036	rc = NEMR3NotifyPhysRamRegister(pVM, GCPhys, cb);
2037	pgmUnlock(pVM);
2038	#ifdef VBOX_WITH_REM
2039	REMR3NotifyPhysRamRegister(pVM, GCPhys, cb, REM_NOTIFY_PHYS_RAM_FLAGS_RAM);
2040	#endif
2041	return rc;
2042	}
2043
2044
2045	/**
2046	* Worker called by PGMR3InitFinalize if we're configured to pre-allocate RAM.
2047	*
2048	* We do this late in the init process so that all the ROM and MMIO ranges have
2049	* been registered already and we don't go wasting memory on them.
2050	*
2051	* @returns VBox status code.
2052	*
2053	* @param pVM The cross context VM structure.
2054	*/
2055	int pgmR3PhysRamPreAllocate(PVM pVM)
2056	{
2057	Assert(pVM->pgm.s.fRamPreAlloc);
2058	Log(("pgmR3PhysRamPreAllocate: enter\n"));
2059
2060	/*
2061	* Walk the RAM ranges and allocate all RAM pages, halt at
2062	* the first allocation error.
2063	*/
2064	uint64_t cPages = 0;
2065	uint64_t NanoTS = RTTimeNanoTS();
2066	pgmLock(pVM);
2067	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2068	{
2069	PPGMPAGE pPage = &pRam->aPages[0];
2070	RTGCPHYS GCPhys = pRam->GCPhys;
2071	uint32_t cLeft = pRam->cb >> PAGE_SHIFT;
2072	while (cLeft-- > 0)
2073	{
2074	if (PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
2075	{
2076	switch (PGM_PAGE_GET_STATE(pPage))
2077	{
2078	case PGM_PAGE_STATE_ZERO:
2079	{
2080	int rc = pgmPhysAllocPage(pVM, pPage, GCPhys);
2081	if (RT_FAILURE(rc))
2082	{
2083	LogRel(("PGM: RAM Pre-allocation failed at %RGp (in %s) with rc=%Rrc\n", GCPhys, pRam->pszDesc, rc));
2084	pgmUnlock(pVM);
2085	return rc;
2086	}
2087	cPages++;
2088	break;
2089	}
2090
2091	case PGM_PAGE_STATE_BALLOONED:
2092	case PGM_PAGE_STATE_ALLOCATED:
2093	case PGM_PAGE_STATE_WRITE_MONITORED:
2094	case PGM_PAGE_STATE_SHARED:
2095	/* nothing to do here. */
2096	break;
2097	}
2098	}
2099
2100	/* next */
2101	pPage++;
2102	GCPhys += PAGE_SIZE;
2103	}
2104	}
2105	pgmUnlock(pVM);
2106	NanoTS = RTTimeNanoTS() - NanoTS;
2107
2108	LogRel(("PGM: Pre-allocated %llu pages in %llu ms\n", cPages, NanoTS / 1000000));
2109	Log(("pgmR3PhysRamPreAllocate: returns VINF_SUCCESS\n"));
2110	return VINF_SUCCESS;
2111	}
2112
2113
2114	/**
2115	* Checks shared page checksums.
2116	*
2117	* @param pVM The cross context VM structure.
2118	*/
2119	void pgmR3PhysAssertSharedPageChecksums(PVM pVM)
2120	{
2121	#ifdef VBOX_STRICT
2122	pgmLock(pVM);
2123
2124	if (pVM->pgm.s.cSharedPages > 0)
2125	{
2126	/*
2127	* Walk the ram ranges.
2128	*/
2129	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2130	{
2131	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2132	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2133
2134	while (iPage-- > 0)
2135	{
2136	PPGMPAGE pPage = &pRam->aPages[iPage];
2137	if (PGM_PAGE_IS_SHARED(pPage))
2138	{
2139	uint32_t u32Checksum = pPage->s.u2Unused0/* \| ((uint32_t)pPage->s.u2Unused1 << 8)*/;
2140	if (!u32Checksum)
2141	{
2142	RTGCPHYS GCPhysPage = pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT);
2143	void const *pvPage;
2144	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhysPage, &pvPage);
2145	if (RT_SUCCESS(rc))
2146	{
2147	uint32_t u32Checksum2 = RTCrc32(pvPage, PAGE_SIZE);
2148	# if 0
2149	AssertMsg((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum, ("GCPhysPage=%RGp\n", GCPhysPage));
2150	# else
2151	if ((u32Checksum2 & /UINT32_C(0x00000303)/ 0x3) == u32Checksum)
2152	LogFlow(("shpg %#x @ %RGp %#x [OK]\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
2153	else
2154	AssertMsgFailed(("shpg %#x @ %RGp %#x\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
2155	# endif
2156	}
2157	else
2158	AssertRC(rc);
2159	}
2160	}
2161
2162	} /* for each page */
2163
2164	} /* for each ram range */
2165	}
2166
2167	pgmUnlock(pVM);
2168	#endif /* VBOX_STRICT */
2169	NOREF(pVM);
2170	}
2171
2172
2173	/**
2174	* Resets the physical memory state.
2175	*
2176	* ASSUMES that the caller owns the PGM lock.
2177	*
2178	* @returns VBox status code.
2179	* @param pVM The cross context VM structure.
2180	*/
2181	int pgmR3PhysRamReset(PVM pVM)
2182	{
2183	PGM_LOCK_ASSERT_OWNER(pVM);
2184
2185	/* Reset the memory balloon. */
2186	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2187	AssertRC(rc);
2188
2189	#ifdef VBOX_WITH_PAGE_SHARING
2190	/* Clear all registered shared modules. */
2191	pgmR3PhysAssertSharedPageChecksums(pVM);
2192	rc = GMMR3ResetSharedModules(pVM);
2193	AssertRC(rc);
2194	#endif
2195	/* Reset counters. */
2196	pVM->pgm.s.cReusedSharedPages = 0;
2197	pVM->pgm.s.cBalloonedPages = 0;
2198
2199	return VINF_SUCCESS;
2200	}
2201
2202
2203	/**
2204	* Resets (zeros) the RAM after all devices and components have been reset.
2205	*
2206	* ASSUMES that the caller owns the PGM lock.
2207	*
2208	* @returns VBox status code.
2209	* @param pVM The cross context VM structure.
2210	*/
2211	int pgmR3PhysRamZeroAll(PVM pVM)
2212	{
2213	PGM_LOCK_ASSERT_OWNER(pVM);
2214
2215	/*
2216	* We batch up pages that should be freed instead of calling GMM for
2217	* each and every one of them.
2218	*/
2219	uint32_t cPendingPages = 0;
2220	PGMMFREEPAGESREQ pReq;
2221	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2222	AssertLogRelRCReturn(rc, rc);
2223
2224	/*
2225	* Walk the ram ranges.
2226	*/
2227	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2228	{
2229	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2230	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2231
2232	if ( !pVM->pgm.s.fRamPreAlloc
2233	&& pVM->pgm.s.fZeroRamPagesOnReset)
2234	{
2235	/* Replace all RAM pages by ZERO pages. */
2236	while (iPage-- > 0)
2237	{
2238	PPGMPAGE pPage = &pRam->aPages[iPage];
2239	switch (PGM_PAGE_GET_TYPE(pPage))
2240	{
2241	case PGMPAGETYPE_RAM:
2242	/* Do not replace pages part of a 2 MB continuous range
2243	with zero pages, but zero them instead. */
2244	if ( PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE
2245	\|\| PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE_DISABLED)
2246	{
2247	void *pvPage;
2248	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2249	AssertLogRelRCReturn(rc, rc);
2250	ASMMemZeroPage(pvPage);
2251	}
2252	else if (PGM_PAGE_IS_BALLOONED(pPage))
2253	{
2254	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2255	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2256	}
2257	else if (!PGM_PAGE_IS_ZERO(pPage))
2258	{
2259	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2260	PGMPAGETYPE_RAM);
2261	AssertLogRelRCReturn(rc, rc);
2262	}
2263	break;
2264
2265	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2266	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2267	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2268	true /fDoAccounting/);
2269	break;
2270
2271	case PGMPAGETYPE_MMIO2:
2272	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2273	case PGMPAGETYPE_ROM:
2274	case PGMPAGETYPE_MMIO:
2275	break;
2276	default:
2277	AssertFailed();
2278	}
2279	} /* for each page */
2280	}
2281	else
2282	{
2283	/* Zero the memory. */
2284	while (iPage-- > 0)
2285	{
2286	PPGMPAGE pPage = &pRam->aPages[iPage];
2287	switch (PGM_PAGE_GET_TYPE(pPage))
2288	{
2289	case PGMPAGETYPE_RAM:
2290	switch (PGM_PAGE_GET_STATE(pPage))
2291	{
2292	case PGM_PAGE_STATE_ZERO:
2293	break;
2294
2295	case PGM_PAGE_STATE_BALLOONED:
2296	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2297	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2298	break;
2299
2300	case PGM_PAGE_STATE_SHARED:
2301	case PGM_PAGE_STATE_WRITE_MONITORED:
2302	rc = pgmPhysPageMakeWritable(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2303	AssertLogRelRCReturn(rc, rc);
2304	RT_FALL_THRU();
2305
2306	case PGM_PAGE_STATE_ALLOCATED:
2307	if (pVM->pgm.s.fZeroRamPagesOnReset)
2308	{
2309	void *pvPage;
2310	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2311	AssertLogRelRCReturn(rc, rc);
2312	ASMMemZeroPage(pvPage);
2313	}
2314	break;
2315	}
2316	break;
2317
2318	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2319	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2320	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2321	true /fDoAccounting/);
2322	break;
2323
2324	case PGMPAGETYPE_MMIO2:
2325	case PGMPAGETYPE_ROM_SHADOW:
2326	case PGMPAGETYPE_ROM:
2327	case PGMPAGETYPE_MMIO:
2328	break;
2329	default:
2330	AssertFailed();
2331
2332	}
2333	} /* for each page */
2334	}
2335
2336	}
2337
2338	/*
2339	* Finish off any pages pending freeing.
2340	*/
2341	if (cPendingPages)
2342	{
2343	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2344	AssertLogRelRCReturn(rc, rc);
2345	}
2346	GMMR3FreePagesCleanup(pReq);
2347	return VINF_SUCCESS;
2348	}
2349
2350
2351	/**
2352	* Frees all RAM during VM termination
2353	*
2354	* ASSUMES that the caller owns the PGM lock.
2355	*
2356	* @returns VBox status code.
2357	* @param pVM The cross context VM structure.
2358	*/
2359	int pgmR3PhysRamTerm(PVM pVM)
2360	{
2361	PGM_LOCK_ASSERT_OWNER(pVM);
2362
2363	/* Reset the memory balloon. */
2364	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2365	AssertRC(rc);
2366
2367	#ifdef VBOX_WITH_PAGE_SHARING
2368	/*
2369	* Clear all registered shared modules.
2370	*/
2371	pgmR3PhysAssertSharedPageChecksums(pVM);
2372	rc = GMMR3ResetSharedModules(pVM);
2373	AssertRC(rc);
2374
2375	/*
2376	* Flush the handy pages updates to make sure no shared pages are hiding
2377	* in there. (No unlikely if the VM shuts down, apparently.)
2378	*/
2379	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_FLUSH_HANDY_PAGES, 0, NULL);
2380	#endif
2381
2382	/*
2383	* We batch up pages that should be freed instead of calling GMM for
2384	* each and every one of them.
2385	*/
2386	uint32_t cPendingPages = 0;
2387	PGMMFREEPAGESREQ pReq;
2388	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2389	AssertLogRelRCReturn(rc, rc);
2390
2391	/*
2392	* Walk the ram ranges.
2393	*/
2394	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2395	{
2396	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2397	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2398
2399	while (iPage-- > 0)
2400	{
2401	PPGMPAGE pPage = &pRam->aPages[iPage];
2402	switch (PGM_PAGE_GET_TYPE(pPage))
2403	{
2404	case PGMPAGETYPE_RAM:
2405	/* Free all shared pages. Private pages are automatically freed during GMM VM cleanup. */
2406	/** @todo change this to explicitly free private pages here. */
2407	if (PGM_PAGE_IS_SHARED(pPage))
2408	{
2409	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2410	PGMPAGETYPE_RAM);
2411	AssertLogRelRCReturn(rc, rc);
2412	}
2413	break;
2414
2415	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2416	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO:
2417	case PGMPAGETYPE_MMIO2:
2418	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2419	case PGMPAGETYPE_ROM:
2420	case PGMPAGETYPE_MMIO:
2421	break;
2422	default:
2423	AssertFailed();
2424	}
2425	} /* for each page */
2426	}
2427
2428	/*
2429	* Finish off any pages pending freeing.
2430	*/
2431	if (cPendingPages)
2432	{
2433	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2434	AssertLogRelRCReturn(rc, rc);
2435	}
2436	GMMR3FreePagesCleanup(pReq);
2437	return VINF_SUCCESS;
2438	}
2439
2440
2441	/**
2442	* This is the interface IOM is using to register an MMIO region.
2443	*
2444	* It will check for conflicts and ensure that a RAM range structure
2445	* is present before calling the PGMR3HandlerPhysicalRegister API to
2446	* register the callbacks.
2447	*
2448	* @returns VBox status code.
2449	*
2450	* @param pVM The cross context VM structure.
2451	* @param GCPhys The start of the MMIO region.
2452	* @param cb The size of the MMIO region.
2453	* @param hType The physical access handler type registration.
2454	* @param pvUserR3 The user argument for R3.
2455	* @param pvUserR0 The user argument for R0.
2456	* @param pvUserRC The user argument for RC.
2457	* @param pszDesc The description of the MMIO region.
2458	*/
2459	VMMR3DECL(int) PGMR3PhysMMIORegister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMPHYSHANDLERTYPE hType,
2460	RTR3PTR pvUserR3, RTR0PTR pvUserR0, RTRCPTR pvUserRC, const char *pszDesc)
2461	{
2462	/*
2463	* Assert on some assumption.
2464	*/
2465	VM_ASSERT_EMT(pVM);
2466	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2467	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2468	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2469	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2470	Assert(((PPGMPHYSHANDLERTYPEINT)MMHyperHeapOffsetToPtr(pVM, hType))->enmKind == PGMPHYSHANDLERKIND_MMIO);
2471
2472	int rc = pgmLock(pVM);
2473	AssertRCReturn(rc, rc);
2474
2475	/*
2476	* Make sure there's a RAM range structure for the region.
2477	*/
2478	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2479	bool fRamExists = false;
2480	PPGMRAMRANGE pRamPrev = NULL;
2481	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2482	while (pRam && GCPhysLast >= pRam->GCPhys)
2483	{
2484	if ( GCPhysLast >= pRam->GCPhys
2485	&& GCPhys <= pRam->GCPhysLast)
2486	{
2487	/* Simplification: all within the same range. */
2488	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
2489	&& GCPhysLast <= pRam->GCPhysLast,
2490	("%RGp-%RGp (MMIO/%s) falls partly outside %RGp-%RGp (%s)\n",
2491	GCPhys, GCPhysLast, pszDesc,
2492	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2493	pgmUnlock(pVM),
2494	VERR_PGM_RAM_CONFLICT);
2495
2496	/* Check that it's all RAM or MMIO pages. */
2497	PCPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2498	uint32_t cLeft = cb >> PAGE_SHIFT;
2499	while (cLeft-- > 0)
2500	{
2501	AssertLogRelMsgReturnStmt( PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM
2502	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO,
2503	("%RGp-%RGp (MMIO/%s): %RGp is not a RAM or MMIO page - type=%d desc=%s\n",
2504	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, PGM_PAGE_GET_TYPE(pPage), pRam->pszDesc),
2505	pgmUnlock(pVM),
2506	VERR_PGM_RAM_CONFLICT);
2507	pPage++;
2508	}
2509
2510	/* Looks good. */
2511	fRamExists = true;
2512	break;
2513	}
2514
2515	/* next */
2516	pRamPrev = pRam;
2517	pRam = pRam->pNextR3;
2518	}
2519	PPGMRAMRANGE pNew;
2520	if (fRamExists)
2521	{
2522	pNew = NULL;
2523
2524	/*
2525	* Make all the pages in the range MMIO/ZERO pages, freeing any
2526	* RAM pages currently mapped here. This might not be 100% correct
2527	* for PCI memory, but we're doing the same thing for MMIO2 pages.
2528	*/
2529	rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, PGMPAGETYPE_MMIO);
2530	AssertRCReturnStmt(rc, pgmUnlock(pVM), rc);
2531
2532	/* Force a PGM pool flush as guest ram references have been changed. */
2533	/** @todo not entirely SMP safe; assuming for now the guest takes
2534	* care of this internally (not touch mapped mmio while changing the
2535	* mapping). */
2536	PVMCPU pVCpu = VMMGetCpu(pVM);
2537	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2538	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2539	}
2540	else
2541	{
2542
2543	/*
2544	* No RAM range, insert an ad hoc one.
2545	*
2546	* Note that we don't have to tell REM about this range because
2547	* PGMHandlerPhysicalRegisterEx will do that for us.
2548	*/
2549	Log(("PGMR3PhysMMIORegister: Adding ad hoc MMIO range for %RGp-%RGp %s\n", GCPhys, GCPhysLast, pszDesc));
2550
2551	const uint32_t cPages = cb >> PAGE_SHIFT;
2552	const size_t cbRamRange = RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]);
2553	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), 16, MM_TAG_PGM_PHYS, (void **)&pNew);
2554	AssertLogRelMsgRCReturnStmt(rc, ("cbRamRange=%zu\n", cbRamRange), pgmUnlock(pVM), rc);
2555
2556	/* Initialize the range. */
2557	pNew->pSelfR0 = MMHyperCCToR0(pVM, pNew);
2558	pNew->pSelfRC = MMHyperCCToRC(pVM, pNew);
2559	pNew->GCPhys = GCPhys;
2560	pNew->GCPhysLast = GCPhysLast;
2561	pNew->cb = cb;
2562	pNew->pszDesc = pszDesc;
2563	pNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO;
2564	pNew->pvR3 = NULL;
2565	pNew->paLSPages = NULL;
2566
2567	uint32_t iPage = cPages;
2568	while (iPage-- > 0)
2569	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_MMIO);
2570	Assert(PGM_PAGE_GET_TYPE(&pNew->aPages[0]) == PGMPAGETYPE_MMIO);
2571
2572	/* update the page count stats. */
2573	pVM->pgm.s.cPureMmioPages += cPages;
2574	pVM->pgm.s.cAllPages += cPages;
2575
2576	/* link it */
2577	pgmR3PhysLinkRamRange(pVM, pNew, pRamPrev);
2578	}
2579
2580	/*
2581	* Register the access handler.
2582	*/
2583	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, hType, pvUserR3, pvUserR0, pvUserRC, pszDesc);
2584	if ( RT_FAILURE(rc)
2585	&& !fRamExists)
2586	{
2587	pVM->pgm.s.cPureMmioPages -= cb >> PAGE_SHIFT;
2588	pVM->pgm.s.cAllPages -= cb >> PAGE_SHIFT;
2589
2590	/* remove the ad hoc range. */
2591	pgmR3PhysUnlinkRamRange2(pVM, pNew, pRamPrev);
2592	pNew->cb = pNew->GCPhys = pNew->GCPhysLast = NIL_RTGCPHYS;
2593	MMHyperFree(pVM, pRam);
2594	}
2595	pgmPhysInvalidatePageMapTLB(pVM);
2596
2597	pgmUnlock(pVM);
2598	return rc;
2599	}
2600
2601
2602	/**
2603	* This is the interface IOM is using to register an MMIO region.
2604	*
2605	* It will take care of calling PGMHandlerPhysicalDeregister and clean up
2606	* any ad hoc PGMRAMRANGE left behind.
2607	*
2608	* @returns VBox status code.
2609	* @param pVM The cross context VM structure.
2610	* @param GCPhys The start of the MMIO region.
2611	* @param cb The size of the MMIO region.
2612	*/
2613	VMMR3DECL(int) PGMR3PhysMMIODeregister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb)
2614	{
2615	VM_ASSERT_EMT(pVM);
2616
2617	int rc = pgmLock(pVM);
2618	AssertRCReturn(rc, rc);
2619
2620	/*
2621	* First deregister the handler, then check if we should remove the ram range.
2622	*/
2623	rc = PGMHandlerPhysicalDeregister(pVM, GCPhys);
2624	if (RT_SUCCESS(rc))
2625	{
2626	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2627	PPGMRAMRANGE pRamPrev = NULL;
2628	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2629	while (pRam && GCPhysLast >= pRam->GCPhys)
2630	{
2631	/** @todo We're being a bit too careful here. rewrite. */
2632	if ( GCPhysLast == pRam->GCPhysLast
2633	&& GCPhys == pRam->GCPhys)
2634	{
2635	Assert(pRam->cb == cb);
2636
2637	/*
2638	* See if all the pages are dead MMIO pages.
2639	*/
2640	uint32_t const cPages = cb >> PAGE_SHIFT;
2641	bool fAllMMIO = true;
2642	uint32_t iPage = 0;
2643	uint32_t cLeft = cPages;
2644	while (cLeft-- > 0)
2645	{
2646	PPGMPAGE pPage = &pRam->aPages[iPage];
2647	if ( !PGM_PAGE_IS_MMIO_OR_ALIAS(pPage)
2648	/\|\| not-out-of-action later /)
2649	{
2650	fAllMMIO = false;
2651	AssertMsgFailed(("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2652	break;
2653	}
2654	Assert( PGM_PAGE_IS_ZERO(pPage)
2655	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2656	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO);
2657	pPage++;
2658	}
2659	if (fAllMMIO)
2660	{
2661	/*
2662	* Ad-hoc range, unlink and free it.
2663	*/
2664	Log(("PGMR3PhysMMIODeregister: Freeing ad hoc MMIO range for %RGp-%RGp %s\n",
2665	GCPhys, GCPhysLast, pRam->pszDesc));
2666
2667	pVM->pgm.s.cAllPages -= cPages;
2668	pVM->pgm.s.cPureMmioPages -= cPages;
2669
2670	pgmR3PhysUnlinkRamRange2(pVM, pRam, pRamPrev);
2671	pRam->cb = pRam->GCPhys = pRam->GCPhysLast = NIL_RTGCPHYS;
2672	MMHyperFree(pVM, pRam);
2673	break;
2674	}
2675	}
2676
2677	/*
2678	* Range match? It will all be within one range (see PGMAllHandler.cpp).
2679	*/
2680	if ( GCPhysLast >= pRam->GCPhys
2681	&& GCPhys <= pRam->GCPhysLast)
2682	{
2683	Assert(GCPhys >= pRam->GCPhys);
2684	Assert(GCPhysLast <= pRam->GCPhysLast);
2685
2686	/*
2687	* Turn the pages back into RAM pages.
2688	*/
2689	uint32_t iPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2690	uint32_t cLeft = cb >> PAGE_SHIFT;
2691	while (cLeft--)
2692	{
2693	PPGMPAGE pPage = &pRam->aPages[iPage];
2694	AssertMsg( (PGM_PAGE_IS_MMIO(pPage) && PGM_PAGE_IS_ZERO(pPage))
2695	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2696	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO,
2697	("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2698	if (PGM_PAGE_IS_MMIO_OR_ALIAS(pPage))
2699	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_RAM);
2700	}
2701	break;
2702	}
2703
2704	/* next */
2705	pRamPrev = pRam;
2706	pRam = pRam->pNextR3;
2707	}
2708	}
2709
2710	/* Force a PGM pool flush as guest ram references have been changed. */
2711	/** @todo Not entirely SMP safe; assuming for now the guest takes care of
2712	* this internally (not touch mapped mmio while changing the mapping). */
2713	PVMCPU pVCpu = VMMGetCpu(pVM);
2714	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2715	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2716
2717	pgmPhysInvalidatePageMapTLB(pVM);
2718	pgmPhysInvalidRamRangeTlbs(pVM);
2719	pgmUnlock(pVM);
2720	return rc;
2721	}
2722
2723
2724	/**
2725	* Locate a MMIO2 range.
2726	*
2727	* @returns Pointer to the MMIO2 range.
2728	* @param pVM The cross context VM structure.
2729	* @param pDevIns The device instance owning the region.
2730	* @param iSubDev The sub-device number.
2731	* @param iRegion The region.
2732	*/
2733	DECLINLINE(PPGMREGMMIORANGE) pgmR3PhysMMIOExFind(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion)
2734	{
2735	/*
2736	* Search the list. There shouldn't be many entries.
2737	*/
2738	/** @todo Optimize this lookup! There may now be many entries and it'll
2739	* become really slow when doing MMR3HyperMapMMIO2 and similar. */
2740	for (PPGMREGMMIORANGE pCur = pVM->pgm.s.pRegMmioRangesR3; pCur; pCur = pCur->pNextR3)
2741	if ( pCur->pDevInsR3 == pDevIns
2742	&& pCur->iRegion == iRegion
2743	&& pCur->iSubDev == iSubDev)
2744	return pCur;
2745	return NULL;
2746	}
2747
2748
2749	/**
2750	* @callback_method_impl{FNPGMRELOCATE, Relocate a floating MMIO/MMIO2 range.}
2751	* @sa pgmR3PhysRamRangeRelocate
2752	*/
2753	static DECLCALLBACK(bool) pgmR3PhysMMIOExRangeRelocate(PVM pVM, RTGCPTR GCPtrOld, RTGCPTR GCPtrNew,
2754	PGMRELOCATECALL enmMode, void *pvUser)
2755	{
2756	PPGMREGMMIORANGE pMmio = (PPGMREGMMIORANGE)pvUser;
2757	Assert(pMmio->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING);
2758	Assert(pMmio->RamRange.pSelfRC == GCPtrOld + PAGE_SIZE + RT_UOFFSETOF(PGMREGMMIORANGE, RamRange)); RT_NOREF_PV(GCPtrOld);
2759
2760	switch (enmMode)
2761	{
2762	case PGMRELOCATECALL_SUGGEST:
2763	return true;
2764
2765	case PGMRELOCATECALL_RELOCATE:
2766	{
2767	/*
2768	* Update myself, then relink all the ranges and flush the RC TLB.
2769	*/
2770	pgmLock(pVM);
2771
2772	pMmio->RamRange.pSelfRC = (RTRCPTR)(GCPtrNew + PAGE_SIZE + RT_UOFFSETOF(PGMREGMMIORANGE, RamRange));
2773
2774	pgmR3PhysRelinkRamRanges(pVM);
2775	for (unsigned i = 0; i < PGM_RAMRANGE_TLB_ENTRIES; i++)
2776	pVM->pgm.s.apRamRangesTlbRC[i] = NIL_RTRCPTR;
2777
2778	pgmUnlock(pVM);
2779	return true;
2780	}
2781
2782	default:
2783	AssertFailedReturn(false);
2784	}
2785	}
2786
2787
2788	/**
2789	* Calculates the number of chunks
2790	*
2791	* @returns Number of registration chunk needed.
2792	* @param pVM The cross context VM structure.
2793	* @param cb The size of the MMIO/MMIO2 range.
2794	* @param pcPagesPerChunk Where to return the number of pages tracked by each
2795	* chunk. Optional.
2796	* @param pcbChunk Where to return the guest mapping size for a chunk.
2797	*/
2798	static uint16_t pgmR3PhysMMIOExCalcChunkCount(PVM pVM, RTGCPHYS cb, uint32_t pcPagesPerChunk, uint32_t pcbChunk)
2799	{
2800	RT_NOREF_PV(pVM); /* without raw mode */
2801
2802	/*
2803	* This is the same calculation as PGMR3PhysRegisterRam does, except we'll be
2804	* needing a few bytes extra the PGMREGMMIORANGE structure.
2805	*
2806	* Note! In additions, we've got a 24 bit sub-page range for MMIO2 ranges, leaving
2807	* us with an absolute maximum of 16777215 pages per chunk (close to 64 GB).
2808	*/
2809	uint32_t cbChunk;
2810	uint32_t cPagesPerChunk;
2811	if (!VM_IS_RAW_MODE_ENABLED(pVM))
2812	{
2813	cbChunk = 16U*_1M;
2814	cPagesPerChunk = 1048048; /* max ~1048059 */
2815	AssertCompile(sizeof(PGMREGMMIORANGE) + sizeof(PGMPAGE) * 1048048 < 16U_1M - PAGE_SIZE 2);
2816	}
2817	else
2818	{
2819	cbChunk = 4U*_1M;
2820	cPagesPerChunk = 261616; /* max ~261627 */
2821	AssertCompile(sizeof(PGMREGMMIORANGE) + sizeof(PGMPAGE) * 261616 < 4U_1M - PAGE_SIZE 2);
2822	}
2823	AssertRelease(cPagesPerChunk <= PGM_MMIO2_MAX_PAGE_COUNT); /* See above note. */
2824	AssertRelease(RT_UOFFSETOF_DYN(PGMREGMMIORANGE, RamRange.aPages[cPagesPerChunk]) + PAGE_SIZE * 2 <= cbChunk);
2825	if (pcbChunk)
2826	*pcbChunk = cbChunk;
2827	if (pcPagesPerChunk)
2828	*pcPagesPerChunk = cPagesPerChunk;
2829
2830	/* Calc the number of chunks we need. */
2831	RTGCPHYS const cPages = cb >> X86_PAGE_SHIFT;
2832	uint16_t cChunks = (uint16_t)((cPages + cPagesPerChunk - 1) / cPagesPerChunk);
2833	AssertRelease((RTGCPHYS)cChunks * cPagesPerChunk >= cPages);
2834	return cChunks;
2835	}
2836
2837
2838	/**
2839	* Worker for PGMR3PhysMMIOExPreRegister & PGMR3PhysMMIO2Register that allocates
2840	* and the PGMREGMMIORANGE structures and does basic initialization.
2841	*
2842	* Caller must set type specfic members and initialize the PGMPAGE structures.
2843	*
2844	* @returns VBox status code.
2845	* @param pVM The cross context VM structure.
2846	* @param pDevIns The device instance owning the region.
2847	* @param iSubDev The sub-device number (internal PCI config number).
2848	* @param iRegion The region number. If the MMIO2 memory is a PCI
2849	* I/O region this number has to be the number of that
2850	* region. Otherwise it can be any number safe
2851	* UINT8_MAX.
2852	* @param cb The size of the region. Must be page aligned.
2853	* @param pszDesc The description.
2854	* @param ppHeadRet Where to return the pointer to the first
2855	* registration chunk.
2856	*
2857	* @thread EMT
2858	*/
2859	static int pgmR3PhysMMIOExCreate(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb,
2860	const char pszDesc, PPGMREGMMIORANGE ppHeadRet)
2861	{
2862	/*
2863	* Figure out how many chunks we need and of which size.
2864	*/
2865	uint32_t cPagesPerChunk;
2866	uint16_t cChunks = pgmR3PhysMMIOExCalcChunkCount(pVM, cb, &cPagesPerChunk, NULL);
2867	AssertReturn(cChunks, VERR_PGM_PHYS_MMIO_EX_IPE);
2868
2869	/*
2870	* Allocate the chunks.
2871	*/
2872	PPGMREGMMIORANGE *ppNext = ppHeadRet;
2873	*ppNext = NULL;
2874
2875	int rc = VINF_SUCCESS;
2876	uint32_t cPagesLeft = cb >> X86_PAGE_SHIFT;
2877	for (uint16_t iChunk = 0; iChunk < cChunks && RT_SUCCESS(rc); iChunk++)
2878	{
2879	/*
2880	* We currently do a single RAM range for the whole thing. This will
2881	* probably have to change once someone needs really large MMIO regions,
2882	* as we will be running into SUPR3PageAllocEx limitations and such.
2883	*/
2884	const uint32_t cPagesTrackedByChunk = RT_MIN(cPagesLeft, cPagesPerChunk);
2885	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIORANGE, RamRange.aPages[cPagesTrackedByChunk]);
2886	PPGMREGMMIORANGE pNew = NULL;
2887	if ( iChunk + 1 < cChunks
2888	\|\| cbRange >= _1M)
2889	{
2890	/*
2891	* Allocate memory for the registration structure.
2892	*/
2893	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
2894	size_t const cbChunk = (1 + cChunkPages + 1) << PAGE_SHIFT;
2895	AssertLogRelBreakStmt(cbChunk == (uint32_t)cbChunk, rc = VERR_OUT_OF_RANGE);
2896	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
2897	AssertBreakStmt(paChunkPages, rc = VERR_NO_TMP_MEMORY);
2898	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
2899	void *pvChunk = NULL;
2900	rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk,
2901	#if defined(VBOX_WITH_MORE_RING0_MEM_MAPPINGS)
2902	&R0PtrChunk,
2903	#elif defined(VBOX_WITH_2X_4GB_ADDR_SPACE)
2904	VM_IS_HM_OR_NEM_ENABLED(pVM) ? &R0PtrChunk : NULL,
2905	#else
2906	NULL,
2907	#endif
2908	paChunkPages);
2909	AssertLogRelMsgRCBreakStmt(rc, ("rc=%Rrc, cChunkPages=%#zx\n", rc, cChunkPages), RTMemTmpFree(paChunkPages));
2910
2911	#if defined(VBOX_WITH_MORE_RING0_MEM_MAPPINGS)
2912	Assert(R0PtrChunk != NIL_RTR0PTR);
2913	#elif defined(VBOX_WITH_2X_4GB_ADDR_SPACE)
2914	if (!VM_IS_HM_OR_NEM_ENABLED(pVM))
2915	R0PtrChunk = NIL_RTR0PTR;
2916	#else
2917	R0PtrChunk = (uintptr_t)pvChunk;
2918	#endif
2919	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
2920
2921	pNew = (PPGMREGMMIORANGE)pvChunk;
2922	pNew->RamRange.fFlags = PGM_RAM_RANGE_FLAGS_FLOATING;
2923	pNew->RamRange.pSelfR0 = R0PtrChunk + RT_UOFFSETOF(PGMREGMMIORANGE, RamRange);
2924
2925	/*
2926	* If we might end up in raw-mode, make a HMA mapping of the range,
2927	* just like we do for memory above 4GB.
2928	*/
2929	if (!VM_IS_RAW_MODE_ENABLED(pVM))
2930	pNew->RamRange.pSelfRC = NIL_RTRCPTR;
2931	else
2932	{
2933	RTGCPTR GCPtrChunkMap = pVM->pgm.s.GCPtrPrevRamRangeMapping - RT_ALIGN_Z(cbChunk, _4M);
2934	RTGCPTR const GCPtrChunk = GCPtrChunkMap + PAGE_SIZE;
2935	rc = PGMR3MapPT(pVM, GCPtrChunkMap, (uint32_t)cbChunk, 0 /fFlags/, pgmR3PhysMMIOExRangeRelocate, pNew, pszDesc);
2936	if (RT_SUCCESS(rc))
2937	{
2938	pVM->pgm.s.GCPtrPrevRamRangeMapping = GCPtrChunkMap;
2939
2940	RTGCPTR GCPtrPage = GCPtrChunk;
2941	for (uint32_t iPage = 0; iPage < cChunkPages && RT_SUCCESS(rc); iPage++, GCPtrPage += PAGE_SIZE)
2942	rc = PGMMap(pVM, GCPtrPage, paChunkPages[iPage].Phys, PAGE_SIZE, 0);
2943	}
2944	if (RT_FAILURE(rc))
2945	{
2946	SUPR3PageFreeEx(pvChunk, cChunkPages);
2947	break;
2948	}
2949	pNew->RamRange.pSelfRC = GCPtrChunk + RT_UOFFSETOF(PGMREGMMIORANGE, RamRange);
2950	}
2951	}
2952	/*
2953	* Not so big, do a one time hyper allocation.
2954	*/
2955	else
2956	{
2957	rc = MMR3HyperAllocOnceNoRel(pVM, cbRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
2958	AssertLogRelMsgRCBreak(rc, ("cbRange=%zu\n", cbRange));
2959
2960	/*
2961	* Initialize allocation specific items.
2962	*/
2963	//pNew->RamRange.fFlags = 0;
2964	pNew->RamRange.pSelfR0 = MMHyperCCToR0(pVM, &pNew->RamRange);
2965	pNew->RamRange.pSelfRC = MMHyperCCToRC(pVM, &pNew->RamRange);
2966	}
2967
2968	/*
2969	* Initialize the registration structure (caller does specific bits).
2970	*/
2971	pNew->pDevInsR3 = pDevIns;
2972	//pNew->pvR3 = NULL;
2973	//pNew->pNext = NULL;
2974	//pNew->fFlags = 0;
2975	if (iChunk == 0)
2976	pNew->fFlags \|= PGMREGMMIORANGE_F_FIRST_CHUNK;
2977	if (iChunk + 1 == cChunks)
2978	pNew->fFlags \|= PGMREGMMIORANGE_F_LAST_CHUNK;
2979	pNew->iSubDev = iSubDev;
2980	pNew->iRegion = iRegion;
2981	pNew->idSavedState = UINT8_MAX;
2982	pNew->idMmio2 = UINT8_MAX;
2983	//pNew->pPhysHandlerR3 = NULL;
2984	//pNew->paLSPages = NULL;
2985	pNew->RamRange.GCPhys = NIL_RTGCPHYS;
2986	pNew->RamRange.GCPhysLast = NIL_RTGCPHYS;
2987	pNew->RamRange.pszDesc = pszDesc;
2988	pNew->RamRange.cb = pNew->cbReal = (RTGCPHYS)cPagesTrackedByChunk << X86_PAGE_SHIFT;
2989	pNew->RamRange.fFlags \|= PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO_EX;
2990	//pNew->RamRange.pvR3 = NULL;
2991	//pNew->RamRange.paLSPages = NULL;
2992
2993	*ppNext = pNew;
2994	ASMCompilerBarrier();
2995	cPagesLeft -= cPagesTrackedByChunk;
2996	ppNext = &pNew->pNextR3;
2997	}
2998	Assert(cPagesLeft == 0);
2999
3000	if (RT_SUCCESS(rc))
3001	{
3002	Assert((*ppHeadRet)->fFlags & PGMREGMMIORANGE_F_FIRST_CHUNK);
3003	return VINF_SUCCESS;
3004	}
3005
3006	/*
3007	* Free floating ranges.
3008	*/
3009	while (*ppHeadRet)
3010	{
3011	PPGMREGMMIORANGE pFree = *ppHeadRet;
3012	*ppHeadRet = pFree->pNextR3;
3013
3014	if (pFree->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
3015	{
3016	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIORANGE, RamRange.aPages[pFree->RamRange.cb >> X86_PAGE_SHIFT]);
3017	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
3018	SUPR3PageFreeEx(pFree, cChunkPages);
3019	}
3020	}
3021
3022	return rc;
3023	}
3024
3025
3026	/**
3027	* Common worker PGMR3PhysMMIOExPreRegister & PGMR3PhysMMIO2Register that links
3028	* a complete registration entry into the lists and lookup tables.
3029	*
3030	* @param pVM The cross context VM structure.
3031	* @param pNew The new MMIO / MMIO2 registration to link.
3032	*/
3033	static void pgmR3PhysMMIOExLink(PVM pVM, PPGMREGMMIORANGE pNew)
3034	{
3035	/*
3036	* Link it into the list (order doesn't matter, so insert it at the head).
3037	*
3038	* Note! The range we're link may consist of multiple chunks, so we have to
3039	* find the last one.
3040	*/
3041	PPGMREGMMIORANGE pLast = pNew;
3042	for (pLast = pNew; ; pLast = pLast->pNextR3)
3043	{
3044	if (pLast->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3045	break;
3046	Assert(pLast->pNextR3);
3047	Assert(pLast->pNextR3->pDevInsR3 == pNew->pDevInsR3);
3048	Assert(pLast->pNextR3->iSubDev == pNew->iSubDev);
3049	Assert(pLast->pNextR3->iRegion == pNew->iRegion);
3050	Assert((pLast->pNextR3->fFlags & PGMREGMMIORANGE_F_MMIO2) == (pNew->fFlags & PGMREGMMIORANGE_F_MMIO2));
3051	Assert(pLast->pNextR3->idMmio2 == (pLast->fFlags & PGMREGMMIORANGE_F_MMIO2 ? pNew->idMmio2 + 1 : UINT8_MAX));
3052	}
3053
3054	pgmLock(pVM);
3055
3056	/* Link in the chain of ranges at the head of the list. */
3057	pLast->pNextR3 = pVM->pgm.s.pRegMmioRangesR3;
3058	pVM->pgm.s.pRegMmioRangesR3 = pNew;
3059
3060	/* If MMIO, insert the MMIO2 range/page IDs. */
3061	uint8_t idMmio2 = pNew->idMmio2;
3062	if (idMmio2 != UINT8_MAX)
3063	{
3064	for (;;)
3065	{
3066	Assert(pNew->fFlags & PGMREGMMIORANGE_F_MMIO2);
3067	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == NULL);
3068	Assert(pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] == NIL_RTR0PTR);
3069	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = pNew;
3070	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = pNew->RamRange.pSelfR0 - RT_UOFFSETOF(PGMREGMMIORANGE, RamRange);
3071	if (pNew->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3072	break;
3073	pNew = pNew->pNextR3;
3074	}
3075	}
3076	else
3077	Assert(!(pNew->fFlags & PGMREGMMIORANGE_F_MMIO2));
3078
3079	pgmPhysInvalidatePageMapTLB(pVM);
3080	pgmUnlock(pVM);
3081	}
3082
3083
3084	/**
3085	* Allocate and pre-register an MMIO region.
3086	*
3087	* This is currently the way to deal with large MMIO regions. It may in the
3088	* future be extended to be the way we deal with all MMIO regions, but that
3089	* means we'll have to do something about the simple list based approach we take
3090	* to tracking the registrations.
3091	*
3092	* @returns VBox status code.
3093	* @retval VINF_SUCCESS on success, *ppv pointing to the R3 mapping of the
3094	* memory.
3095	* @retval VERR_ALREADY_EXISTS if the region already exists.
3096	*
3097	* @param pVM The cross context VM structure.
3098	* @param pDevIns The device instance owning the region.
3099	* @param iSubDev The sub-device number.
3100	* @param iRegion The region number. If the MMIO2 memory is a PCI
3101	* I/O region this number has to be the number of that
3102	* region. Otherwise it can be any number safe
3103	* UINT8_MAX.
3104	* @param cbRegion The size of the region. Must be page aligned.
3105	* @param hType The physical handler callback type.
3106	* @param pvUserR3 User parameter for ring-3 context callbacks.
3107	* @param pvUserR0 User parameter for ring-0 context callbacks.
3108	* @param pvUserRC User parameter for raw-mode context callbacks.
3109	* @param pszDesc The description.
3110	*
3111	* @thread EMT
3112	*
3113	* @sa PGMR3PhysMMIORegister, PGMR3PhysMMIO2Register,
3114	* PGMR3PhysMMIOExMap, PGMR3PhysMMIOExUnmap, PGMR3PhysMMIOExDeregister.
3115	*/
3116	VMMR3DECL(int) PGMR3PhysMMIOExPreRegister(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cbRegion,
3117	PGMPHYSHANDLERTYPE hType, RTR3PTR pvUserR3, RTR0PTR pvUserR0, RTRCPTR pvUserRC,
3118	const char *pszDesc)
3119	{
3120	/*
3121	* Validate input.
3122	*/
3123	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3124	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3125	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3126	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3127	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
3128	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
3129	AssertReturn(pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iRegion) == NULL, VERR_ALREADY_EXISTS);
3130	AssertReturn(!(cbRegion & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3131	AssertReturn(cbRegion, VERR_INVALID_PARAMETER);
3132
3133	const uint32_t cPages = cbRegion >> PAGE_SHIFT;
3134	AssertLogRelReturn(((RTGCPHYS)cPages << PAGE_SHIFT) == cbRegion, VERR_INVALID_PARAMETER);
3135	AssertLogRelReturn(cPages <= (MM_MMIO_64_MAX >> X86_PAGE_SHIFT), VERR_OUT_OF_RANGE);
3136
3137	/*
3138	* For the 2nd+ instance, mangle the description string so it's unique.
3139	*/
3140	if (pDevIns->iInstance > 0) /** @todo Move to PDMDevHlp.cpp and use a real string cache. */
3141	{
3142	pszDesc = MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s [%u]", pszDesc, pDevIns->iInstance);
3143	if (!pszDesc)
3144	return VERR_NO_MEMORY;
3145	}
3146
3147	/*
3148	* Register the MMIO callbacks.
3149	*/
3150	PPGMPHYSHANDLER pPhysHandler;
3151	int rc = pgmHandlerPhysicalExCreate(pVM, hType, pvUserR3, pvUserR0, pvUserRC, pszDesc, &pPhysHandler);
3152	if (RT_SUCCESS(rc))
3153	{
3154	/*
3155	* Create the registered MMIO range record for it.
3156	*/
3157	PPGMREGMMIORANGE pNew;
3158	rc = pgmR3PhysMMIOExCreate(pVM, pDevIns, iSubDev, iRegion, cbRegion, pszDesc, &pNew);
3159	if (RT_SUCCESS(rc))
3160	{
3161	Assert(!(pNew->fFlags & PGMREGMMIORANGE_F_MMIO2));
3162
3163	/*
3164	* Intialize the page structures and set up physical handlers (one for each chunk).
3165	*/
3166	for (PPGMREGMMIORANGE pCur = pNew; pCur != NULL && RT_SUCCESS(rc); pCur = pCur->pNextR3)
3167	{
3168	if (pCur == pNew)
3169	pCur->pPhysHandlerR3 = pPhysHandler;
3170	else
3171	rc = pgmHandlerPhysicalExDup(pVM, pPhysHandler, &pCur->pPhysHandlerR3);
3172
3173	uint32_t iPage = pCur->RamRange.cb >> X86_PAGE_SHIFT;
3174	while (iPage-- > 0)
3175	PGM_PAGE_INIT_ZERO(&pCur->RamRange.aPages[iPage], pVM, PGMPAGETYPE_MMIO);
3176	}
3177	if (RT_SUCCESS(rc))
3178	{
3179	/*
3180	* Update the page count stats, link the registration and we're done.
3181	*/
3182	pVM->pgm.s.cAllPages += cPages;
3183	pVM->pgm.s.cPureMmioPages += cPages;
3184
3185	pgmR3PhysMMIOExLink(pVM, pNew);
3186	return VINF_SUCCESS;
3187	}
3188
3189	/*
3190	* Clean up in case we're out of memory for extra access handlers.
3191	*/
3192	while (pNew != NULL)
3193	{
3194	PPGMREGMMIORANGE pFree = pNew;
3195	pNew = pFree->pNextR3;
3196
3197	if (pFree->pPhysHandlerR3)
3198	{
3199	pgmHandlerPhysicalExDestroy(pVM, pFree->pPhysHandlerR3);
3200	pFree->pPhysHandlerR3 = NULL;
3201	}
3202
3203	if (pFree->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
3204	{
3205	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIORANGE, RamRange.aPages[pFree->RamRange.cb >> X86_PAGE_SHIFT]);
3206	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
3207	SUPR3PageFreeEx(pFree, cChunkPages);
3208	}
3209	}
3210	}
3211	else
3212	pgmHandlerPhysicalExDestroy(pVM, pPhysHandler);
3213	}
3214	return rc;
3215	}
3216
3217
3218	/**
3219	* Allocate and register an MMIO2 region.
3220	*
3221	* As mentioned elsewhere, MMIO2 is just RAM spelled differently. It's RAM
3222	* associated with a device. It is also non-shared memory with a permanent
3223	* ring-3 mapping and page backing (presently).
3224	*
3225	* A MMIO2 range may overlap with base memory if a lot of RAM is configured for
3226	* the VM, in which case we'll drop the base memory pages. Presently we will
3227	* make no attempt to preserve anything that happens to be present in the base
3228	* memory that is replaced, this is of course incorrect but it's too much
3229	* effort.
3230	*
3231	* @returns VBox status code.
3232	* @retval VINF_SUCCESS on success, *ppv pointing to the R3 mapping of the
3233	* memory.
3234	* @retval VERR_ALREADY_EXISTS if the region already exists.
3235	*
3236	* @param pVM The cross context VM structure.
3237	* @param pDevIns The device instance owning the region.
3238	* @param iSubDev The sub-device number.
3239	* @param iRegion The region number. If the MMIO2 memory is a PCI
3240	* I/O region this number has to be the number of that
3241	* region. Otherwise it can be any number safe
3242	* UINT8_MAX.
3243	* @param cb The size of the region. Must be page aligned.
3244	* @param fFlags Reserved for future use, must be zero.
3245	* @param ppv Where to store the pointer to the ring-3 mapping of
3246	* the memory.
3247	* @param pszDesc The description.
3248	* @thread EMT
3249	*/
3250	VMMR3DECL(int) PGMR3PhysMMIO2Register(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cb,
3251	uint32_t fFlags, void *ppv, const char pszDesc)
3252	{
3253	/*
3254	* Validate input.
3255	*/
3256	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3257	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3258	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3259	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3260	AssertPtrReturn(ppv, VERR_INVALID_POINTER);
3261	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
3262	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
3263	AssertReturn(pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iRegion) == NULL, VERR_ALREADY_EXISTS);
3264	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3265	AssertReturn(cb, VERR_INVALID_PARAMETER);
3266	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
3267
3268	const uint32_t cPages = cb >> PAGE_SHIFT;
3269	AssertLogRelReturn(((RTGCPHYS)cPages << PAGE_SHIFT) == cb, VERR_INVALID_PARAMETER);
3270	AssertLogRelReturn(cPages <= (MM_MMIO_64_MAX >> X86_PAGE_SHIFT), VERR_OUT_OF_RANGE);
3271
3272	/*
3273	* For the 2nd+ instance, mangle the description string so it's unique.
3274	*/
3275	if (pDevIns->iInstance > 0) /** @todo Move to PDMDevHlp.cpp and use a real string cache. */
3276	{
3277	pszDesc = MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s [%u]", pszDesc, pDevIns->iInstance);
3278	if (!pszDesc)
3279	return VERR_NO_MEMORY;
3280	}
3281
3282	/*
3283	* Allocate an MMIO2 range ID (not freed on failure).
3284	*
3285	* The zero ID is not used as it could be confused with NIL_GMM_PAGEID, so
3286	* the IDs goes from 1 thru PGM_MMIO2_MAX_RANGES.
3287	*/
3288	unsigned cChunks = pgmR3PhysMMIOExCalcChunkCount(pVM, cb, NULL, NULL);
3289	pgmLock(pVM);
3290	uint8_t idMmio2 = pVM->pgm.s.cMmio2Regions + 1;
3291	unsigned cNewMmio2Regions = pVM->pgm.s.cMmio2Regions + cChunks;
3292	if (cNewMmio2Regions > PGM_MMIO2_MAX_RANGES)
3293	{
3294	pgmUnlock(pVM);
3295	AssertLogRelFailedReturn(VERR_PGM_TOO_MANY_MMIO2_RANGES);
3296	}
3297	pVM->pgm.s.cMmio2Regions = cNewMmio2Regions;
3298	pgmUnlock(pVM);
3299
3300	/*
3301	* Try reserve and allocate the backing memory first as this is what is
3302	* most likely to fail.
3303	*/
3304	int rc = MMR3AdjustFixedReservation(pVM, cPages, pszDesc);
3305	if (RT_SUCCESS(rc))
3306	{
3307	PSUPPAGE paPages = (PSUPPAGE)RTMemTmpAlloc(cPages * sizeof(SUPPAGE));
3308	if (RT_SUCCESS(rc))
3309	{
3310	void *pvPages;
3311	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, NULL /pR0Ptr/, paPages);
3312	if (RT_SUCCESS(rc))
3313	{
3314	memset(pvPages, 0, cPages * PAGE_SIZE);
3315
3316	/*
3317	* Create the registered MMIO range record for it.
3318	*/
3319	PPGMREGMMIORANGE pNew;
3320	rc = pgmR3PhysMMIOExCreate(pVM, pDevIns, iSubDev, iRegion, cb, pszDesc, &pNew);
3321	if (RT_SUCCESS(rc))
3322	{
3323	uint32_t iSrcPage = 0;
3324	uint8_t pbCurPages = (uint8_t )pvPages;
3325	for (PPGMREGMMIORANGE pCur = pNew; pCur; pCur = pCur->pNextR3)
3326	{
3327	pCur->pvR3 = pbCurPages;
3328	pCur->RamRange.pvR3 = pbCurPages;
3329	pCur->idMmio2 = idMmio2;
3330	pCur->fFlags \|= PGMREGMMIORANGE_F_MMIO2;
3331
3332	uint32_t iDstPage = pCur->RamRange.cb >> X86_PAGE_SHIFT;
3333	while (iDstPage-- > 0)
3334	{
3335	PGM_PAGE_INIT(&pNew->RamRange.aPages[iDstPage],
3336	paPages[iDstPage + iSrcPage].Phys,
3337	PGM_MMIO2_PAGEID_MAKE(idMmio2, iDstPage),
3338	PGMPAGETYPE_MMIO2, PGM_PAGE_STATE_ALLOCATED);
3339	}
3340
3341	/* advance. */
3342	iSrcPage += pCur->RamRange.cb >> X86_PAGE_SHIFT;
3343	pbCurPages += pCur->RamRange.cb;
3344	idMmio2++;
3345	}
3346
3347	RTMemTmpFree(paPages);
3348
3349	/*
3350	* Update the page count stats, link the registration and we're done.
3351	*/
3352	pVM->pgm.s.cAllPages += cPages;
3353	pVM->pgm.s.cPrivatePages += cPages;
3354
3355	pgmR3PhysMMIOExLink(pVM, pNew);
3356
3357	*ppv = pvPages;
3358	return VINF_SUCCESS;
3359	}
3360
3361	SUPR3PageFreeEx(pvPages, cPages);
3362	}
3363	}
3364	RTMemTmpFree(paPages);
3365	MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pszDesc);
3366	}
3367	if (pDevIns->iInstance > 0)
3368	MMR3HeapFree((void *)pszDesc);
3369	return rc;
3370	}
3371
3372
3373	/**
3374	* Deregisters and frees an MMIO2 region or a pre-registered MMIO region
3375	*
3376	* Any physical (and virtual) access handlers registered for the region must
3377	* be deregistered before calling this function.
3378	*
3379	* @returns VBox status code.
3380	* @param pVM The cross context VM structure.
3381	* @param pDevIns The device instance owning the region.
3382	* @param iSubDev The sub-device number. Pass UINT32_MAX for wildcard
3383	* matching.
3384	* @param iRegion The region. Pass UINT32_MAX for wildcard matching.
3385	*/
3386	VMMR3DECL(int) PGMR3PhysMMIOExDeregister(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion)
3387	{
3388	/*
3389	* Validate input.
3390	*/
3391	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3392	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3393	AssertReturn(iSubDev <= UINT8_MAX \|\| iSubDev == UINT32_MAX, VERR_INVALID_PARAMETER);
3394	AssertReturn(iRegion <= UINT8_MAX \|\| iRegion == UINT32_MAX, VERR_INVALID_PARAMETER);
3395
3396	/*
3397	* The loop here scanning all registrations will make sure that multi-chunk ranges
3398	* get properly deregistered, though it's original purpose was the wildcard iRegion.
3399	*/
3400	pgmLock(pVM);
3401	int rc = VINF_SUCCESS;
3402	unsigned cFound = 0;
3403	PPGMREGMMIORANGE pPrev = NULL;
3404	PPGMREGMMIORANGE pCur = pVM->pgm.s.pRegMmioRangesR3;
3405	while (pCur)
3406	{
3407	if ( pCur->pDevInsR3 == pDevIns
3408	&& ( iRegion == UINT32_MAX
3409	\|\| pCur->iRegion == iRegion)
3410	&& ( iSubDev == UINT32_MAX
3411	\|\| pCur->iSubDev == iSubDev) )
3412	{
3413	cFound++;
3414
3415	/*
3416	* Unmap it if it's mapped.
3417	*/
3418	if (pCur->fFlags & PGMREGMMIORANGE_F_MAPPED)
3419	{
3420	int rc2 = PGMR3PhysMMIOExUnmap(pVM, pCur->pDevInsR3, pCur->iSubDev, pCur->iRegion, pCur->RamRange.GCPhys);
3421	AssertRC(rc2);
3422	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3423	rc = rc2;
3424	}
3425
3426	/*
3427	* Must tell IOM about MMIO (first one only).
3428	*/
3429	if ((pCur->fFlags & (PGMREGMMIORANGE_F_MMIO2 \| PGMREGMMIORANGE_F_FIRST_CHUNK)) == PGMREGMMIORANGE_F_MMIO2)
3430	IOMR3MmioExNotifyDeregistered(pVM, pCur->pPhysHandlerR3->pvUserR3);
3431
3432	/*
3433	* Unlink it
3434	*/
3435	PPGMREGMMIORANGE pNext = pCur->pNextR3;
3436	if (pPrev)
3437	pPrev->pNextR3 = pNext;
3438	else
3439	pVM->pgm.s.pRegMmioRangesR3 = pNext;
3440	pCur->pNextR3 = NULL;
3441
3442	uint8_t idMmio2 = pCur->idMmio2;
3443	if (idMmio2 != UINT8_MAX)
3444	{
3445	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == pCur);
3446	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = NULL;
3447	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = NIL_RTR0PTR;
3448	}
3449
3450	/*
3451	* Free the memory.
3452	*/
3453	uint32_t const cPages = pCur->cbReal >> PAGE_SHIFT;
3454	if (pCur->fFlags & PGMREGMMIORANGE_F_MMIO2)
3455	{
3456	int rc2 = SUPR3PageFreeEx(pCur->pvR3, cPages);
3457	AssertRC(rc2);
3458	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3459	rc = rc2;
3460
3461	rc2 = MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pCur->RamRange.pszDesc);
3462	AssertRC(rc2);
3463	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
3464	rc = rc2;
3465	}
3466
3467	/* we're leaking hyper memory here if done at runtime. */
3468	#ifdef VBOX_STRICT
3469	VMSTATE const enmState = VMR3GetState(pVM);
3470	AssertMsg( enmState == VMSTATE_POWERING_OFF
3471	\|\| enmState == VMSTATE_POWERING_OFF_LS
3472	\|\| enmState == VMSTATE_OFF
3473	\|\| enmState == VMSTATE_OFF_LS
3474	\|\| enmState == VMSTATE_DESTROYING
3475	\|\| enmState == VMSTATE_TERMINATED
3476	\|\| enmState == VMSTATE_CREATING
3477	, ("%s\n", VMR3GetStateName(enmState)));
3478	#endif
3479
3480	const bool fIsMmio2 = RT_BOOL(pCur->fFlags & PGMREGMMIORANGE_F_MMIO2);
3481	if (pCur->RamRange.fFlags & PGM_RAM_RANGE_FLAGS_FLOATING)
3482	{
3483	const size_t cbRange = RT_UOFFSETOF_DYN(PGMREGMMIORANGE, RamRange.aPages[cPages]);
3484	size_t const cChunkPages = RT_ALIGN_Z(cbRange, PAGE_SIZE) >> PAGE_SHIFT;
3485	SUPR3PageFreeEx(pCur, cChunkPages);
3486	}
3487	/*else
3488	{
3489	rc = MMHyperFree(pVM, pCur); - does not work, see the alloc call.
3490	AssertRCReturn(rc, rc);
3491	} */
3492
3493
3494	/* update page count stats */
3495	pVM->pgm.s.cAllPages -= cPages;
3496	if (fIsMmio2)
3497	pVM->pgm.s.cPrivatePages -= cPages;
3498	else
3499	pVM->pgm.s.cPureMmioPages -= cPages;
3500
3501	/* next */
3502	pCur = pNext;
3503	}
3504	else
3505	{
3506	pPrev = pCur;
3507	pCur = pCur->pNextR3;
3508	}
3509	}
3510	pgmPhysInvalidatePageMapTLB(pVM);
3511	pgmUnlock(pVM);
3512	return !cFound && iRegion != UINT32_MAX && iSubDev != UINT32_MAX ? VERR_NOT_FOUND : rc;
3513	}
3514
3515
3516	/**
3517	* Maps a MMIO2 region or a pre-registered MMIO region.
3518	*
3519	* This is done when a guest / the bios / state loading changes the
3520	* PCI config. The replacing of base memory has the same restrictions
3521	* as during registration, of course.
3522	*
3523	* @returns VBox status code.
3524	*
3525	* @param pVM The cross context VM structure.
3526	* @param pDevIns The device instance owning the region.
3527	* @param iSubDev The sub-device number of the registered region.
3528	* @param iRegion The index of the registered region.
3529	* @param GCPhys The guest-physical address to be remapped.
3530	*/
3531	VMMR3DECL(int) PGMR3PhysMMIOExMap(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS GCPhys)
3532	{
3533	/*
3534	* Validate input.
3535	*
3536	* Note! It's safe to walk the MMIO/MMIO2 list since registrations only
3537	* happens during VM construction.
3538	*/
3539	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3540	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3541	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3542	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3543	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
3544	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3545	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3546
3547	PPGMREGMMIORANGE pFirstMmio = pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iRegion);
3548	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3549	Assert(pFirstMmio->fFlags & PGMREGMMIORANGE_F_FIRST_CHUNK);
3550
3551	PPGMREGMMIORANGE pLastMmio = pFirstMmio;
3552	RTGCPHYS cbRange = 0;
3553	for (;;)
3554	{
3555	AssertReturn(!(pLastMmio->fFlags & PGMREGMMIORANGE_F_MAPPED), VERR_WRONG_ORDER);
3556	Assert(pLastMmio->RamRange.GCPhys == NIL_RTGCPHYS);
3557	Assert(pLastMmio->RamRange.GCPhysLast == NIL_RTGCPHYS);
3558	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3559	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3560	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3561	cbRange += pLastMmio->RamRange.cb;
3562	if (pLastMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3563	break;
3564	pLastMmio = pLastMmio->pNextR3;
3565	}
3566
3567	RTGCPHYS GCPhysLast = GCPhys + cbRange - 1;
3568	AssertLogRelReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3569
3570	/*
3571	* Find our location in the ram range list, checking for restriction
3572	* we don't bother implementing yet (partially overlapping, multiple
3573	* ram ranges).
3574	*/
3575	pgmLock(pVM);
3576
3577	AssertReturnStmt(!(pFirstMmio->fFlags & PGMREGMMIORANGE_F_MAPPED), pgmUnlock(pVM), VERR_WRONG_ORDER);
3578
3579	bool fRamExists = false;
3580	PPGMRAMRANGE pRamPrev = NULL;
3581	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3582	while (pRam && GCPhysLast >= pRam->GCPhys)
3583	{
3584	if ( GCPhys <= pRam->GCPhysLast
3585	&& GCPhysLast >= pRam->GCPhys)
3586	{
3587	/* Completely within? */
3588	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
3589	&& GCPhysLast <= pRam->GCPhysLast,
3590	("%RGp-%RGp (MMIOEx/%s) falls partly outside %RGp-%RGp (%s)\n",
3591	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc,
3592	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3593	pgmUnlock(pVM),
3594	VERR_PGM_RAM_CONFLICT);
3595
3596	/* Check that all the pages are RAM pages. */
3597	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3598	uint32_t cPagesLeft = cbRange >> PAGE_SHIFT;
3599	while (cPagesLeft-- > 0)
3600	{
3601	AssertLogRelMsgReturnStmt(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
3602	("%RGp isn't a RAM page (%d) - mapping %RGp-%RGp (MMIO2/%s).\n",
3603	GCPhys, PGM_PAGE_GET_TYPE(pPage), GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc),
3604	pgmUnlock(pVM),
3605	VERR_PGM_RAM_CONFLICT);
3606	pPage++;
3607	}
3608
3609	/* There can only be one MMIO/MMIO2 chunk matching here! */
3610	AssertLogRelMsgReturnStmt(pFirstMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK,
3611	("%RGp-%RGp (MMIOEx/%s, flags %#X) consists of multiple chunks whereas the RAM somehow doesn't!\n",
3612	GCPhys, GCPhysLast, pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
3613	pgmUnlock(pVM),
3614	VERR_PGM_PHYS_MMIO_EX_IPE);
3615
3616	fRamExists = true;
3617	break;
3618	}
3619
3620	/* next */
3621	pRamPrev = pRam;
3622	pRam = pRam->pNextR3;
3623	}
3624	Log(("PGMR3PhysMMIOExMap: %RGp-%RGp fRamExists=%RTbool %s\n", GCPhys, GCPhysLast, fRamExists, pFirstMmio->RamRange.pszDesc));
3625
3626
3627	/*
3628	* Make the changes.
3629	*/
3630	RTGCPHYS GCPhysCur = GCPhys;
3631	for (PPGMREGMMIORANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3632	{
3633	pCurMmio->RamRange.GCPhys = GCPhysCur;
3634	pCurMmio->RamRange.GCPhysLast = GCPhysCur + pCurMmio->RamRange.cb - 1;
3635	if (pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3636	{
3637	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3638	break;
3639	}
3640	GCPhysCur += pCurMmio->RamRange.cb;
3641	}
3642
3643	if (fRamExists)
3644	{
3645	/*
3646	* Make all the pages in the range MMIO/ZERO pages, freeing any
3647	* RAM pages currently mapped here. This might not be 100% correct
3648	* for PCI memory, but we're doing the same thing for MMIO2 pages.
3649	*
3650	* We replace this MMIO/ZERO pages with real pages in the MMIO2 case.
3651	*/
3652	Assert(pFirstMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK); /* Only one chunk */
3653
3654	int rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, PGMPAGETYPE_MMIO);
3655	AssertRCReturnStmt(rc, pgmUnlock(pVM), rc);
3656
3657	if (pFirstMmio->fFlags & PGMREGMMIORANGE_F_MMIO2)
3658	{
3659	/* replace the pages, freeing all present RAM pages. */
3660	PPGMPAGE pPageSrc = &pFirstMmio->RamRange.aPages[0];
3661	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3662	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3663	while (cPagesLeft-- > 0)
3664	{
3665	Assert(PGM_PAGE_IS_MMIO(pPageDst));
3666
3667	RTHCPHYS const HCPhys = PGM_PAGE_GET_HCPHYS(pPageSrc);
3668	uint32_t const idPage = PGM_PAGE_GET_PAGEID(pPageSrc);
3669	PGM_PAGE_SET_PAGEID(pVM, pPageDst, idPage);
3670	PGM_PAGE_SET_HCPHYS(pVM, pPageDst, HCPhys);
3671	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO2);
3672	PGM_PAGE_SET_STATE(pVM, pPageDst, PGM_PAGE_STATE_ALLOCATED);
3673	PGM_PAGE_SET_PDE_TYPE(pVM, pPageDst, PGM_PAGE_PDE_TYPE_DONTCARE);
3674	PGM_PAGE_SET_PTE_INDEX(pVM, pPageDst, 0);
3675	PGM_PAGE_SET_TRACKING(pVM, pPageDst, 0);
3676	/* (We tell NEM at the end of the function.) */
3677
3678	pVM->pgm.s.cZeroPages--;
3679	GCPhys += PAGE_SIZE;
3680	pPageSrc++;
3681	pPageDst++;
3682	}
3683	}
3684
3685	/* Flush physical page map TLB. */
3686	pgmPhysInvalidatePageMapTLB(pVM);
3687
3688	/* Force a PGM pool flush as guest ram references have been changed. */
3689	/** @todo not entirely SMP safe; assuming for now the guest takes care of
3690	* this internally (not touch mapped mmio while changing the mapping). */
3691	PVMCPU pVCpu = VMMGetCpu(pVM);
3692	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3693	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3694	}
3695	else
3696	{
3697	/*
3698	* No RAM range, insert the ones prepared during registration.
3699	*/
3700	for (PPGMREGMMIORANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3701	{
3702	/* Clear the tracking data of pages we're going to reactivate. */
3703	PPGMPAGE pPageSrc = &pCurMmio->RamRange.aPages[0];
3704	uint32_t cPagesLeft = pCurMmio->RamRange.cb >> PAGE_SHIFT;
3705	while (cPagesLeft-- > 0)
3706	{
3707	PGM_PAGE_SET_TRACKING(pVM, pPageSrc, 0);
3708	PGM_PAGE_SET_PTE_INDEX(pVM, pPageSrc, 0);
3709	pPageSrc++;
3710	}
3711
3712	/* link in the ram range */
3713	pgmR3PhysLinkRamRange(pVM, &pCurMmio->RamRange, pRamPrev);
3714
3715	if (pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3716	{
3717	Assert(pCurMmio->RamRange.GCPhysLast == GCPhysLast);
3718	break;
3719	}
3720	pRamPrev = &pCurMmio->RamRange;
3721	}
3722	}
3723
3724	/*
3725	* Register the access handler if plain MMIO.
3726	*
3727	* We must register access handlers for each range since the access handler
3728	* code refuses to deal with multiple ranges (and we can).
3729	*/
3730	if (!(pFirstMmio->fFlags & PGMREGMMIORANGE_F_MMIO2))
3731	{
3732	int rc = VINF_SUCCESS;
3733	for (PPGMREGMMIORANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3734	{
3735	Assert(!(pCurMmio->fFlags & PGMREGMMIORANGE_F_MAPPED));
3736	rc = pgmHandlerPhysicalExRegister(pVM, pCurMmio->pPhysHandlerR3, pCurMmio->RamRange.GCPhys,
3737	pCurMmio->RamRange.GCPhysLast);
3738	if (RT_FAILURE(rc))
3739	break;
3740	pCurMmio->fFlags \|= PGMREGMMIORANGE_F_MAPPED; /* Use this to mark that the handler is registered. */
3741	if (pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3742	{
3743	rc = IOMR3MmioExNotifyMapped(pVM, pFirstMmio->pPhysHandlerR3->pvUserR3, GCPhys);
3744	break;
3745	}
3746	}
3747	if (RT_FAILURE(rc))
3748	{
3749	/* Almost impossible, but try clean up properly and get out of here. */
3750	for (PPGMREGMMIORANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3751	{
3752	if (pCurMmio->fFlags & PGMREGMMIORANGE_F_MAPPED)
3753	{
3754	pCurMmio->fFlags &= ~PGMREGMMIORANGE_F_MAPPED;
3755	pgmHandlerPhysicalExDeregister(pVM, pCurMmio->pPhysHandlerR3, fRamExists);
3756	}
3757
3758	if (!fRamExists)
3759	pgmR3PhysUnlinkRamRange(pVM, &pCurMmio->RamRange);
3760	else
3761	{
3762	Assert(pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK); /* Only one chunk */
3763
3764	uint32_t cPagesLeft = pCurMmio->RamRange.cb >> PAGE_SHIFT;
3765	PPGMPAGE pPageDst = &pRam->aPages[(pCurMmio->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3766	while (cPagesLeft-- > 0)
3767	{
3768	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
3769	pPageDst++;
3770	}
3771	}
3772
3773	pCurMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3774	pCurMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3775	if (pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3776	break;
3777	}
3778
3779	pgmUnlock(pVM);
3780	return rc;
3781	}
3782	}
3783
3784	/*
3785	* We're good, set the flags and invalid the mapping TLB.
3786	*/
3787	for (PPGMREGMMIORANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3788	{
3789	pCurMmio->fFlags \|= PGMREGMMIORANGE_F_MAPPED;
3790	if (fRamExists)
3791	pCurMmio->fFlags \|= PGMREGMMIORANGE_F_OVERLAPPING;
3792	else
3793	pCurMmio->fFlags &= ~PGMREGMMIORANGE_F_OVERLAPPING;
3794	if (pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3795	break;
3796	}
3797	pgmPhysInvalidatePageMapTLB(pVM);
3798
3799	/*
3800	* Notify NEM while holding the lock (experimental) and REM without (like always).
3801	*/
3802	uint32_t const fNemNotify = (pFirstMmio->fFlags & PGMREGMMIORANGE_F_MMIO2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0)
3803	\| (pFirstMmio->fFlags & PGMREGMMIORANGE_F_OVERLAPPING ? NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE : 0);
3804	int rc = NEMR3NotifyPhysMmioExMap(pVM, GCPhys, cbRange, fNemNotify, pFirstMmio->pvR3);
3805
3806	pgmUnlock(pVM);
3807
3808	#ifdef VBOX_WITH_REM
3809	if (!fRamExists && (pFirstMmio->fFlags & PGMREGMMIORANGE_F_MMIO2)) /** @todo this doesn't look right. */
3810	REMR3NotifyPhysRamRegister(pVM, GCPhys, cbRange, REM_NOTIFY_PHYS_RAM_FLAGS_MMIO2);
3811	#endif
3812	return rc;
3813	}
3814
3815
3816	/**
3817	* Unmaps a MMIO2 or a pre-registered MMIO region.
3818	*
3819	* This is done when a guest / the bios / state loading changes the
3820	* PCI config. The replacing of base memory has the same restrictions
3821	* as during registration, of course.
3822	*/
3823	VMMR3DECL(int) PGMR3PhysMMIOExUnmap(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS GCPhys)
3824	{
3825	/*
3826	* Validate input
3827	*/
3828	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3829	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3830	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3831	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3832	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
3833	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
3834	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3835
3836	PPGMREGMMIORANGE pFirstMmio = pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iRegion);
3837	AssertReturn(pFirstMmio, VERR_NOT_FOUND);
3838	Assert(pFirstMmio->fFlags & PGMREGMMIORANGE_F_FIRST_CHUNK);
3839
3840	PPGMREGMMIORANGE pLastMmio = pFirstMmio;
3841	RTGCPHYS cbRange = 0;
3842	for (;;)
3843	{
3844	AssertReturn(pLastMmio->fFlags & PGMREGMMIORANGE_F_MAPPED, VERR_WRONG_ORDER);
3845	AssertReturn(pLastMmio->RamRange.GCPhys == GCPhys + cbRange, VERR_INVALID_PARAMETER);
3846	Assert(pLastMmio->pDevInsR3 == pFirstMmio->pDevInsR3);
3847	Assert(pLastMmio->iSubDev == pFirstMmio->iSubDev);
3848	Assert(pLastMmio->iRegion == pFirstMmio->iRegion);
3849	cbRange += pLastMmio->RamRange.cb;
3850	if (pLastMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3851	break;
3852	pLastMmio = pLastMmio->pNextR3;
3853	}
3854
3855	Log(("PGMR3PhysMMIOExUnmap: %RGp-%RGp %s\n",
3856	pFirstMmio->RamRange.GCPhys, pLastMmio->RamRange.GCPhysLast, pFirstMmio->RamRange.pszDesc));
3857
3858	int rc = pgmLock(pVM);
3859	AssertRCReturn(rc, rc);
3860	uint16_t const fOldFlags = pFirstMmio->fFlags;
3861	AssertReturnStmt(fOldFlags & PGMREGMMIORANGE_F_MAPPED, pgmUnlock(pVM), VERR_WRONG_ORDER);
3862
3863	/*
3864	* If plain MMIO, we must deregister the handlers first.
3865	*/
3866	if (!(fOldFlags & PGMREGMMIORANGE_F_MMIO2))
3867	{
3868	PPGMREGMMIORANGE pCurMmio = pFirstMmio;
3869	rc = pgmHandlerPhysicalExDeregister(pVM, pFirstMmio->pPhysHandlerR3, RT_BOOL(fOldFlags & PGMREGMMIORANGE_F_OVERLAPPING));
3870	AssertRCReturnStmt(rc, pgmUnlock(pVM), rc);
3871	while (!(pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK))
3872	{
3873	pCurMmio = pCurMmio->pNextR3;
3874	rc = pgmHandlerPhysicalExDeregister(pVM, pCurMmio->pPhysHandlerR3, RT_BOOL(fOldFlags & PGMREGMMIORANGE_F_OVERLAPPING));
3875	AssertRCReturnStmt(rc, pgmUnlock(pVM), VERR_PGM_PHYS_MMIO_EX_IPE);
3876	}
3877
3878	IOMR3MmioExNotifyUnmapped(pVM, pFirstMmio->pPhysHandlerR3->pvUserR3, GCPhys);
3879	}
3880
3881	/*
3882	* Unmap it.
3883	*/
3884	RTGCPHYS const GCPhysRangeNotify = pFirstMmio->RamRange.GCPhys;
3885	if (fOldFlags & PGMREGMMIORANGE_F_OVERLAPPING)
3886	{
3887	/*
3888	* We've replaced RAM, replace with zero pages.
3889	*
3890	* Note! This is where we might differ a little from a real system, because
3891	* it's likely to just show the RAM pages as they were before the
3892	* MMIO/MMIO2 region was mapped here.
3893	*/
3894	/* Only one chunk allowed when overlapping! */
3895	Assert(fOldFlags & PGMREGMMIORANGE_F_LAST_CHUNK);
3896
3897	/* Restore the RAM pages we've replaced. */
3898	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3899	while (pRam->GCPhys > pFirstMmio->RamRange.GCPhysLast)
3900	pRam = pRam->pNextR3;
3901
3902	uint32_t cPagesLeft = pFirstMmio->RamRange.cb >> PAGE_SHIFT;
3903	if (fOldFlags & PGMREGMMIORANGE_F_MMIO2)
3904	pVM->pgm.s.cZeroPages += cPagesLeft;
3905
3906	PPGMPAGE pPageDst = &pRam->aPages[(pFirstMmio->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3907	while (cPagesLeft-- > 0)
3908	{
3909	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
3910	pPageDst++;
3911	}
3912
3913	/* Flush physical page map TLB. */
3914	pgmPhysInvalidatePageMapTLB(pVM);
3915
3916	/* Update range state. */
3917	pFirstMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3918	pFirstMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3919	pFirstMmio->fFlags &= ~(PGMREGMMIORANGE_F_OVERLAPPING \| PGMREGMMIORANGE_F_MAPPED);
3920	}
3921	else
3922	{
3923	/*
3924	* Unlink the chunks related to the MMIO/MMIO2 region.
3925	*/
3926	for (PPGMREGMMIORANGE pCurMmio = pFirstMmio; ; pCurMmio = pCurMmio->pNextR3)
3927	{
3928	pgmR3PhysUnlinkRamRange(pVM, &pCurMmio->RamRange);
3929	pCurMmio->RamRange.GCPhys = NIL_RTGCPHYS;
3930	pCurMmio->RamRange.GCPhysLast = NIL_RTGCPHYS;
3931	pCurMmio->fFlags &= ~(PGMREGMMIORANGE_F_OVERLAPPING \| PGMREGMMIORANGE_F_MAPPED);
3932	if (pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK)
3933	break;
3934	}
3935	}
3936
3937	/* Force a PGM pool flush as guest ram references have been changed. */
3938	/** @todo not entirely SMP safe; assuming for now the guest takes care
3939	* of this internally (not touch mapped mmio while changing the
3940	* mapping). */
3941	PVMCPU pVCpu = VMMGetCpu(pVM);
3942	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
3943	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3944
3945	pgmPhysInvalidatePageMapTLB(pVM);
3946	pgmPhysInvalidRamRangeTlbs(pVM);
3947
3948	/*
3949	* Notify NEM while holding the lock (experimental) and REM without (like always).
3950	*/
3951	uint32_t const fNemFlags = (fOldFlags & PGMREGMMIORANGE_F_MMIO2 ? NEM_NOTIFY_PHYS_MMIO_EX_F_MMIO2 : 0)
3952	\| (fOldFlags & PGMREGMMIORANGE_F_OVERLAPPING ? NEM_NOTIFY_PHYS_MMIO_EX_F_REPLACE : 0);
3953	rc = NEMR3NotifyPhysMmioExUnmap(pVM, GCPhysRangeNotify, cbRange, fNemFlags);
3954	pgmUnlock(pVM);
3955	#ifdef VBOX_WITH_REM
3956	if ((fOldFlags & (PGMREGMMIORANGE_F_OVERLAPPING \| PGMREGMMIORANGE_F_MMIO2)) == PGMREGMMIORANGE_F_MMIO2)
3957	REMR3NotifyPhysRamDeregister(pVM, GCPhysRangeNotify, cbRange);
3958	#endif
3959	return rc;
3960	}
3961
3962
3963	/**
3964	* Reduces the mapping size of a MMIO2 or pre-registered MMIO region.
3965	*
3966	* This is mainly for dealing with old saved states after changing the default
3967	* size of a mapping region. See PGMDevHlpMMIOExReduce and
3968	* PDMPCIDEV::pfnRegionLoadChangeHookR3.
3969	*
3970	* The region must not currently be mapped when making this call. The VM state
3971	* must be state restore or VM construction.
3972	*
3973	* @returns VBox status code.
3974	* @param pVM The cross context VM structure.
3975	* @param pDevIns The device instance owning the region.
3976	* @param iSubDev The sub-device number of the registered region.
3977	* @param iRegion The index of the registered region.
3978	* @param cbRegion The new mapping size.
3979	*/
3980	VMMR3_INT_DECL(int) PGMR3PhysMMIOExReduce(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion, RTGCPHYS cbRegion)
3981	{
3982	/*
3983	* Validate input
3984	*/
3985	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3986	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3987	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
3988	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3989	AssertReturn(cbRegion >= X86_PAGE_SIZE, VERR_INVALID_PARAMETER);
3990	AssertReturn(!(cbRegion & X86_PAGE_OFFSET_MASK), VERR_UNSUPPORTED_ALIGNMENT);
3991	VMSTATE enmVmState = VMR3GetState(pVM);
3992	AssertLogRelMsgReturn( enmVmState == VMSTATE_CREATING
3993	\|\| enmVmState == VMSTATE_LOADING,
3994	("enmVmState=%d (%s)\n", enmVmState, VMR3GetStateName(enmVmState)),
3995	VERR_VM_INVALID_VM_STATE);
3996
3997	int rc = pgmLock(pVM);
3998	AssertRCReturn(rc, rc);
3999
4000	PPGMREGMMIORANGE pFirstMmio = pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iRegion);
4001	if (pFirstMmio)
4002	{
4003	Assert(pFirstMmio->fFlags & PGMREGMMIORANGE_F_FIRST_CHUNK);
4004	if (!(pFirstMmio->fFlags & PGMREGMMIORANGE_F_MAPPED))
4005	{
4006	/*
4007	* NOTE! Current implementation does not support multiple ranges.
4008	* Implement when there is a real world need and thus a testcase.
4009	*/
4010	AssertLogRelMsgStmt(pFirstMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK,
4011	("%s: %#x\n", pFirstMmio->RamRange.pszDesc, pFirstMmio->fFlags),
4012	rc = VERR_NOT_SUPPORTED);
4013	if (RT_SUCCESS(rc))
4014	{
4015	/*
4016	* Make the change.
4017	*/
4018	Log(("PGMR3PhysMMIOExReduce: %s changes from %RGp bytes (%RGp) to %RGp bytes.\n",
4019	pFirstMmio->RamRange.pszDesc, pFirstMmio->RamRange.cb, pFirstMmio->cbReal, cbRegion));
4020
4021	AssertLogRelMsgStmt(cbRegion <= pFirstMmio->cbReal,
4022	("%s: cbRegion=%#RGp cbReal=%#RGp\n", pFirstMmio->RamRange.pszDesc, cbRegion, pFirstMmio->cbReal),
4023	rc = VERR_OUT_OF_RANGE);
4024	if (RT_SUCCESS(rc))
4025	{
4026	pFirstMmio->RamRange.cb = cbRegion;
4027	}
4028	}
4029	}
4030	else
4031	rc = VERR_WRONG_ORDER;
4032	}
4033	else
4034	rc = VERR_NOT_FOUND;
4035
4036	pgmUnlock(pVM);
4037	return rc;
4038	}
4039
4040
4041	/**
4042	* Checks if the given address is an MMIO2 or pre-registered MMIO base address
4043	* or not.
4044	*
4045	* @returns true/false accordingly.
4046	* @param pVM The cross context VM structure.
4047	* @param pDevIns The owner of the memory, optional.
4048	* @param GCPhys The address to check.
4049	*/
4050	VMMR3DECL(bool) PGMR3PhysMMIOExIsBase(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys)
4051	{
4052	/*
4053	* Validate input
4054	*/
4055	VM_ASSERT_EMT_RETURN(pVM, false);
4056	AssertPtrReturn(pDevIns, false);
4057	AssertReturn(GCPhys != NIL_RTGCPHYS, false);
4058	AssertReturn(GCPhys != 0, false);
4059	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), false);
4060
4061	/*
4062	* Search the list.
4063	*/
4064	pgmLock(pVM);
4065	for (PPGMREGMMIORANGE pCurMmio = pVM->pgm.s.pRegMmioRangesR3; pCurMmio; pCurMmio = pCurMmio->pNextR3)
4066	if (pCurMmio->RamRange.GCPhys == GCPhys)
4067	{
4068	Assert(pCurMmio->fFlags & PGMREGMMIORANGE_F_MAPPED);
4069	bool fRet = RT_BOOL(pCurMmio->fFlags & PGMREGMMIORANGE_F_FIRST_CHUNK);
4070	pgmUnlock(pVM);
4071	return fRet;
4072	}
4073	pgmUnlock(pVM);
4074	return false;
4075	}
4076
4077
4078	/**
4079	* Gets the HC physical address of a page in the MMIO2 region.
4080	*
4081	* This is API is intended for MMHyper and shouldn't be called
4082	* by anyone else...
4083	*
4084	* @returns VBox status code.
4085	* @param pVM The cross context VM structure.
4086	* @param pDevIns The owner of the memory, optional.
4087	* @param iSubDev Sub-device number.
4088	* @param iRegion The region.
4089	* @param off The page expressed an offset into the MMIO2 region.
4090	* @param pHCPhys Where to store the result.
4091	*/
4092	VMMR3_INT_DECL(int) PGMR3PhysMMIO2GetHCPhys(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion,
4093	RTGCPHYS off, PRTHCPHYS pHCPhys)
4094	{
4095	/*
4096	* Validate input
4097	*/
4098	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
4099	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
4100	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
4101	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
4102
4103	pgmLock(pVM);
4104	PPGMREGMMIORANGE pCurMmio = pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iRegion);
4105	AssertReturn(pCurMmio, VERR_NOT_FOUND);
4106	AssertReturn(pCurMmio->fFlags & (PGMREGMMIORANGE_F_MMIO2 \| PGMREGMMIORANGE_F_FIRST_CHUNK), VERR_WRONG_TYPE);
4107
4108	while ( off >= pCurMmio->RamRange.cb
4109	&& !(pCurMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK))
4110	{
4111	off -= pCurMmio->RamRange.cb;
4112	pCurMmio = pCurMmio->pNextR3;
4113	}
4114	AssertReturn(off < pCurMmio->RamRange.cb, VERR_INVALID_PARAMETER);
4115
4116	PCPGMPAGE pPage = &pCurMmio->RamRange.aPages[off >> PAGE_SHIFT];
4117	*pHCPhys = PGM_PAGE_GET_HCPHYS(pPage);
4118	pgmUnlock(pVM);
4119	return VINF_SUCCESS;
4120	}
4121
4122
4123	/**
4124	* Maps a portion of an MMIO2 region into kernel space (host).
4125	*
4126	* The kernel mapping will become invalid when the MMIO2 memory is deregistered
4127	* or the VM is terminated.
4128	*
4129	* @return VBox status code.
4130	*
4131	* @param pVM The cross context VM structure.
4132	* @param pDevIns The device owning the MMIO2 memory.
4133	* @param iSubDev The sub-device number.
4134	* @param iRegion The region.
4135	* @param off The offset into the region. Must be page aligned.
4136	* @param cb The number of bytes to map. Must be page aligned.
4137	* @param pszDesc Mapping description.
4138	* @param pR0Ptr Where to store the R0 address.
4139	*/
4140	VMMR3_INT_DECL(int) PGMR3PhysMMIO2MapKernel(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion,
4141	RTGCPHYS off, RTGCPHYS cb, const char *pszDesc, PRTR0PTR pR0Ptr)
4142	{
4143	/*
4144	* Validate input.
4145	*/
4146	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
4147	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
4148	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
4149	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
4150
4151	PPGMREGMMIORANGE pFirstRegMmio = pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iRegion);
4152	AssertReturn(pFirstRegMmio, VERR_NOT_FOUND);
4153	AssertReturn(pFirstRegMmio->fFlags & (PGMREGMMIORANGE_F_MMIO2 \| PGMREGMMIORANGE_F_FIRST_CHUNK), VERR_WRONG_TYPE);
4154	AssertReturn(off < pFirstRegMmio->RamRange.cb, VERR_INVALID_PARAMETER);
4155	AssertReturn(cb <= pFirstRegMmio->RamRange.cb, VERR_INVALID_PARAMETER);
4156	AssertReturn(off + cb <= pFirstRegMmio->RamRange.cb, VERR_INVALID_PARAMETER);
4157	NOREF(pszDesc);
4158
4159	/*
4160	* Pass the request on to the support library/driver.
4161	*/
4162	#if defined(RT_OS_WINDOWS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_OS2) /** @todo Fully implement RTR0MemObjMapKernelEx everywhere. */
4163	AssertLogRelReturn(off == 0, VERR_NOT_SUPPORTED);
4164	AssertLogRelReturn(pFirstRegMmio->fFlags & PGMREGMMIORANGE_F_LAST_CHUNK, VERR_NOT_SUPPORTED);
4165	int rc = SUPR3PageMapKernel(pFirstRegMmio->pvR3, 0 /off/, pFirstRegMmio->RamRange.cb, 0 /fFlags/, pR0Ptr);
4166	#else
4167	int rc = SUPR3PageMapKernel(pFirstRegMmio->pvR3, off, cb, 0 /fFlags/, pR0Ptr);
4168	#endif
4169
4170	return rc;
4171	}
4172
4173
4174	/**
4175	* Changes the region number of an MMIO2 or pre-registered MMIO region.
4176	*
4177	* This is only for dealing with save state issues, nothing else.
4178	*
4179	* @return VBox status code.
4180	*
4181	* @param pVM The cross context VM structure.
4182	* @param pDevIns The device owning the MMIO2 memory.
4183	* @param iSubDev The sub-device number.
4184	* @param iRegion The region.
4185	* @param iNewRegion The new region index.
4186	*
4187	* @sa @bugref{9359}
4188	*/
4189	VMMR3_INT_DECL(int) PGMR3PhysMMIOExChangeRegionNo(PVM pVM, PPDMDEVINS pDevIns, uint32_t iSubDev, uint32_t iRegion,
4190	uint32_t iNewRegion)
4191	{
4192	/*
4193	* Validate input.
4194	*/
4195	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
4196	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
4197	AssertReturn(iSubDev <= UINT8_MAX, VERR_INVALID_PARAMETER);
4198	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
4199	AssertReturn(iNewRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
4200
4201	AssertReturn(pVM->enmVMState == VMSTATE_LOADING, VERR_INVALID_STATE);
4202
4203	PPGMREGMMIORANGE pFirstRegMmio = pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iRegion);
4204	AssertReturn(pFirstRegMmio, VERR_NOT_FOUND);
4205	AssertReturn(pgmR3PhysMMIOExFind(pVM, pDevIns, iSubDev, iNewRegion) == NULL, VERR_RESOURCE_IN_USE);
4206
4207	/*
4208	* Make the change.
4209	*/
4210	pFirstRegMmio->iRegion = (uint8_t)iNewRegion;
4211
4212	return VINF_SUCCESS;
4213	}
4214
4215
4216	/**
4217	* Worker for PGMR3PhysRomRegister.
4218	*
4219	* This is here to simplify lock management, i.e. the caller does all the
4220	* locking and we can simply return without needing to remember to unlock
4221	* anything first.
4222	*
4223	* @returns VBox status code.
4224	* @param pVM The cross context VM structure.
4225	* @param pDevIns The device instance owning the ROM.
4226	* @param GCPhys First physical address in the range.
4227	* Must be page aligned!
4228	* @param cb The size of the range (in bytes).
4229	* Must be page aligned!
4230	* @param pvBinary Pointer to the binary data backing the ROM image.
4231	* @param cbBinary The size of the binary data pvBinary points to.
4232	* This must be less or equal to @a cb.
4233	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
4234	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
4235	* @param pszDesc Pointer to description string. This must not be freed.
4236	*/
4237	static int pgmR3PhysRomRegisterLocked(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
4238	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
4239	{
4240	/*
4241	* Validate input.
4242	*/
4243	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
4244	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
4245	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
4246	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
4247	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
4248	AssertPtrReturn(pvBinary, VERR_INVALID_PARAMETER);
4249	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
4250	AssertReturn(!(fFlags & ~(PGMPHYS_ROM_FLAGS_SHADOWED \| PGMPHYS_ROM_FLAGS_PERMANENT_BINARY)), VERR_INVALID_PARAMETER);
4251	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_CREATING, VERR_VM_INVALID_VM_STATE);
4252
4253	const uint32_t cPages = cb >> PAGE_SHIFT;
4254
4255	/*
4256	* Find the ROM location in the ROM list first.
4257	*/
4258	PPGMROMRANGE pRomPrev = NULL;
4259	PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3;
4260	while (pRom && GCPhysLast >= pRom->GCPhys)
4261	{
4262	if ( GCPhys <= pRom->GCPhysLast
4263	&& GCPhysLast >= pRom->GCPhys)
4264	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
4265	GCPhys, GCPhysLast, pszDesc,
4266	pRom->GCPhys, pRom->GCPhysLast, pRom->pszDesc),
4267	VERR_PGM_RAM_CONFLICT);
4268	/* next */
4269	pRomPrev = pRom;
4270	pRom = pRom->pNextR3;
4271	}
4272
4273	/*
4274	* Find the RAM location and check for conflicts.
4275	*
4276	* Conflict detection is a bit different than for RAM
4277	* registration since a ROM can be located within a RAM
4278	* range. So, what we have to check for is other memory
4279	* types (other than RAM that is) and that we don't span
4280	* more than one RAM range (layz).
4281	*/
4282	bool fRamExists = false;
4283	PPGMRAMRANGE pRamPrev = NULL;
4284	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
4285	while (pRam && GCPhysLast >= pRam->GCPhys)
4286	{
4287	if ( GCPhys <= pRam->GCPhysLast
4288	&& GCPhysLast >= pRam->GCPhys)
4289	{
4290	/* completely within? */
4291	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
4292	&& GCPhysLast <= pRam->GCPhysLast,
4293	("%RGp-%RGp (%s) falls partly outside %RGp-%RGp (%s)\n",
4294	GCPhys, GCPhysLast, pszDesc,
4295	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
4296	VERR_PGM_RAM_CONFLICT);
4297	fRamExists = true;
4298	break;
4299	}
4300
4301	/* next */
4302	pRamPrev = pRam;
4303	pRam = pRam->pNextR3;
4304	}
4305	if (fRamExists)
4306	{
4307	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
4308	uint32_t cPagesLeft = cPages;
4309	while (cPagesLeft-- > 0)
4310	{
4311	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
4312	("%RGp (%R[pgmpage]) isn't a RAM page - registering %RGp-%RGp (%s).\n",
4313	pRam->GCPhys + ((RTGCPHYS)(uintptr_t)(pPage - &pRam->aPages[0]) << PAGE_SHIFT),
4314	pPage, GCPhys, GCPhysLast, pszDesc), VERR_PGM_RAM_CONFLICT);
4315	Assert(PGM_PAGE_IS_ZERO(pPage));
4316	pPage++;
4317	}
4318	}
4319
4320	/*
4321	* Update the base memory reservation if necessary.
4322	*/
4323	uint32_t cExtraBaseCost = fRamExists ? 0 : cPages;
4324	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4325	cExtraBaseCost += cPages;
4326	if (cExtraBaseCost)
4327	{
4328	int rc = MMR3IncreaseBaseReservation(pVM, cExtraBaseCost);
4329	if (RT_FAILURE(rc))
4330	return rc;
4331	}
4332
4333	/*
4334	* Allocate memory for the virgin copy of the RAM.
4335	*/
4336	PGMMALLOCATEPAGESREQ pReq;
4337	int rc = GMMR3AllocatePagesPrepare(pVM, &pReq, cPages, GMMACCOUNT_BASE);
4338	AssertRCReturn(rc, rc);
4339
4340	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4341	{
4342	pReq->aPages[iPage].HCPhysGCPhys = GCPhys + (iPage << PAGE_SHIFT);
4343	pReq->aPages[iPage].idPage = NIL_GMM_PAGEID;
4344	pReq->aPages[iPage].idSharedPage = NIL_GMM_PAGEID;
4345	}
4346
4347	rc = GMMR3AllocatePagesPerform(pVM, pReq);
4348	if (RT_FAILURE(rc))
4349	{
4350	GMMR3AllocatePagesCleanup(pReq);
4351	return rc;
4352	}
4353
4354	/*
4355	* Allocate the new ROM range and RAM range (if necessary).
4356	*/
4357	PPGMROMRANGE pRomNew;
4358	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMROMRANGE, aPages[cPages]), 0, MM_TAG_PGM_PHYS, (void **)&pRomNew);
4359	if (RT_SUCCESS(rc))
4360	{
4361	PPGMRAMRANGE pRamNew = NULL;
4362	if (!fRamExists)
4363	rc = MMHyperAlloc(pVM, RT_UOFFSETOF_DYN(PGMRAMRANGE, aPages[cPages]), sizeof(PGMPAGE), MM_TAG_PGM_PHYS, (void **)&pRamNew);
4364	if (RT_SUCCESS(rc))
4365	{
4366	/*
4367	* Initialize and insert the RAM range (if required).
4368	*/
4369	PPGMROMPAGE pRomPage = &pRomNew->aPages[0];
4370	if (!fRamExists)
4371	{
4372	pRamNew->pSelfR0 = MMHyperCCToR0(pVM, pRamNew);
4373	pRamNew->pSelfRC = MMHyperCCToRC(pVM, pRamNew);
4374	pRamNew->GCPhys = GCPhys;
4375	pRamNew->GCPhysLast = GCPhysLast;
4376	pRamNew->cb = cb;
4377	pRamNew->pszDesc = pszDesc;
4378	pRamNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_ROM;
4379	pRamNew->pvR3 = NULL;
4380	pRamNew->paLSPages = NULL;
4381
4382	PPGMPAGE pPage = &pRamNew->aPages[0];
4383	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
4384	{
4385	PGM_PAGE_INIT(pPage,
4386	pReq->aPages[iPage].HCPhysGCPhys,
4387	pReq->aPages[iPage].idPage,
4388	PGMPAGETYPE_ROM,
4389	PGM_PAGE_STATE_ALLOCATED);
4390
4391	pRomPage->Virgin = *pPage;
4392	}
4393
4394	pVM->pgm.s.cAllPages += cPages;
4395	pgmR3PhysLinkRamRange(pVM, pRamNew, pRamPrev);
4396	}
4397	else
4398	{
4399	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
4400	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
4401	{
4402	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_ROM);
4403	PGM_PAGE_SET_HCPHYS(pVM, pPage, pReq->aPages[iPage].HCPhysGCPhys);
4404	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
4405	PGM_PAGE_SET_PAGEID(pVM, pPage, pReq->aPages[iPage].idPage);
4406	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4407	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
4408	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
4409
4410	pRomPage->Virgin = *pPage;
4411	}
4412
4413	pRamNew = pRam;
4414
4415	pVM->pgm.s.cZeroPages -= cPages;
4416	}
4417	pVM->pgm.s.cPrivatePages += cPages;
4418
4419	/* Flush physical page map TLB. */
4420	pgmPhysInvalidatePageMapTLB(pVM);
4421
4422
4423	/* Notify NEM before we register handlers. */
4424	uint32_t const fNemNotify = (fRamExists ? NEM_NOTIFY_PHYS_ROM_F_REPLACE : 0)
4425	\| (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED ? NEM_NOTIFY_PHYS_ROM_F_SHADOW : 0);
4426	rc = NEMR3NotifyPhysRomRegisterEarly(pVM, GCPhys, cb, fNemNotify);
4427
4428	/*
4429	* !HACK ALERT! REM + (Shadowed) ROM ==> mess.
4430	*
4431	* If it's shadowed we'll register the handler after the ROM notification
4432	* so we get the access handler callbacks that we should. If it isn't
4433	* shadowed we'll do it the other way around to make REM use the built-in
4434	* ROM behavior and not the handler behavior (which is to route all access
4435	* to PGM atm).
4436	*/
4437	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4438	{
4439	#ifdef VBOX_WITH_REM
4440	REMR3NotifyPhysRomRegister(pVM, GCPhys, cb, NULL, true /* fShadowed */);
4441	#endif
4442	if (RT_SUCCESS(rc))
4443	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, pVM->pgm.s.hRomPhysHandlerType,
4444	pRomNew, MMHyperCCToR0(pVM, pRomNew), MMHyperCCToRC(pVM, pRomNew),
4445	pszDesc);
4446	}
4447	else
4448	{
4449	if (RT_SUCCESS(rc))
4450	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, pVM->pgm.s.hRomPhysHandlerType,
4451	pRomNew, MMHyperCCToR0(pVM, pRomNew), MMHyperCCToRC(pVM, pRomNew),
4452	pszDesc);
4453	#ifdef VBOX_WITH_REM
4454	REMR3NotifyPhysRomRegister(pVM, GCPhys, cb, NULL, false /* fShadowed */);
4455	#endif
4456	}
4457	if (RT_SUCCESS(rc))
4458	{
4459	/*
4460	* Copy the image over to the virgin pages.
4461	* This must be done after linking in the RAM range.
4462	*/
4463	size_t cbBinaryLeft = cbBinary;
4464	PPGMPAGE pRamPage = &pRamNew->aPages[(GCPhys - pRamNew->GCPhys) >> PAGE_SHIFT];
4465	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
4466	{
4467	void *pvDstPage;
4468	rc = pgmPhysPageMap(pVM, pRamPage, GCPhys + (iPage << PAGE_SHIFT), &pvDstPage);
4469	if (RT_FAILURE(rc))
4470	{
4471	VMSetError(pVM, rc, RT_SRC_POS, "Failed to map virgin ROM page at %RGp", GCPhys);
4472	break;
4473	}
4474	if (cbBinaryLeft >= PAGE_SIZE)
4475	{
4476	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), PAGE_SIZE);
4477	cbBinaryLeft -= PAGE_SIZE;
4478	}
4479	else
4480	{
4481	ASMMemZeroPage(pvDstPage); /* (shouldn't be necessary, but can't hurt either) */
4482	if (cbBinaryLeft > 0)
4483	{
4484	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), cbBinaryLeft);
4485	cbBinaryLeft = 0;
4486	}
4487	}
4488	}
4489	if (RT_SUCCESS(rc))
4490	{
4491	/*
4492	* Initialize the ROM range.
4493	* Note that the Virgin member of the pages has already been initialized above.
4494	*/
4495	pRomNew->GCPhys = GCPhys;
4496	pRomNew->GCPhysLast = GCPhysLast;
4497	pRomNew->cb = cb;
4498	pRomNew->fFlags = fFlags;
4499	pRomNew->idSavedState = UINT8_MAX;
4500	pRomNew->cbOriginal = cbBinary;
4501	pRomNew->pszDesc = pszDesc;
4502	pRomNew->pvOriginal = fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY
4503	? pvBinary : RTMemDup(pvBinary, cbBinary);
4504	if (pRomNew->pvOriginal)
4505	{
4506	for (unsigned iPage = 0; iPage < cPages; iPage++)
4507	{
4508	PPGMROMPAGE pPage = &pRomNew->aPages[iPage];
4509	pPage->enmProt = PGMROMPROT_READ_ROM_WRITE_IGNORE;
4510	PGM_PAGE_INIT_ZERO(&pPage->Shadow, pVM, PGMPAGETYPE_ROM_SHADOW);
4511	}
4512
4513	/* update the page count stats for the shadow pages. */
4514	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4515	{
4516	pVM->pgm.s.cZeroPages += cPages;
4517	pVM->pgm.s.cAllPages += cPages;
4518	}
4519
4520	/*
4521	* Insert the ROM range, tell REM and return successfully.
4522	*/
4523	pRomNew->pNextR3 = pRom;
4524	pRomNew->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
4525	pRomNew->pNextRC = pRom ? MMHyperCCToRC(pVM, pRom) : NIL_RTRCPTR;
4526
4527	if (pRomPrev)
4528	{
4529	pRomPrev->pNextR3 = pRomNew;
4530	pRomPrev->pNextR0 = MMHyperCCToR0(pVM, pRomNew);
4531	pRomPrev->pNextRC = MMHyperCCToRC(pVM, pRomNew);
4532	}
4533	else
4534	{
4535	pVM->pgm.s.pRomRangesR3 = pRomNew;
4536	pVM->pgm.s.pRomRangesR0 = MMHyperCCToR0(pVM, pRomNew);
4537	pVM->pgm.s.pRomRangesRC = MMHyperCCToRC(pVM, pRomNew);
4538	}
4539
4540	pgmPhysInvalidatePageMapTLB(pVM);
4541	GMMR3AllocatePagesCleanup(pReq);
4542
4543	/* Notify NEM again. */
4544	return NEMR3NotifyPhysRomRegisterLate(pVM, GCPhys, cb, fNemNotify);
4545	}
4546
4547	/* bail out */
4548	rc = VERR_NO_MEMORY;
4549	}
4550
4551	int rc2 = PGMHandlerPhysicalDeregister(pVM, GCPhys);
4552	AssertRC(rc2);
4553	}
4554
4555	if (!fRamExists)
4556	{
4557	pgmR3PhysUnlinkRamRange2(pVM, pRamNew, pRamPrev);
4558	MMHyperFree(pVM, pRamNew);
4559	}
4560	}
4561	MMHyperFree(pVM, pRomNew);
4562	}
4563
4564	/** @todo Purge the mapping cache or something... */
4565	GMMR3FreeAllocatedPages(pVM, pReq);
4566	GMMR3AllocatePagesCleanup(pReq);
4567	return rc;
4568	}
4569
4570
4571	/**
4572	* Registers a ROM image.
4573	*
4574	* Shadowed ROM images requires double the amount of backing memory, so,
4575	* don't use that unless you have to. Shadowing of ROM images is process
4576	* where we can select where the reads go and where the writes go. On real
4577	* hardware the chipset provides means to configure this. We provide
4578	* PGMR3PhysProtectROM() for this purpose.
4579	*
4580	* A read-only copy of the ROM image will always be kept around while we
4581	* will allocate RAM pages for the changes on demand (unless all memory
4582	* is configured to be preallocated).
4583	*
4584	* @returns VBox status code.
4585	* @param pVM The cross context VM structure.
4586	* @param pDevIns The device instance owning the ROM.
4587	* @param GCPhys First physical address in the range.
4588	* Must be page aligned!
4589	* @param cb The size of the range (in bytes).
4590	* Must be page aligned!
4591	* @param pvBinary Pointer to the binary data backing the ROM image.
4592	* @param cbBinary The size of the binary data pvBinary points to.
4593	* This must be less or equal to @a cb.
4594	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
4595	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
4596	* @param pszDesc Pointer to description string. This must not be freed.
4597	*
4598	* @remark There is no way to remove the rom, automatically on device cleanup or
4599	* manually from the device yet. This isn't difficult in any way, it's
4600	* just not something we expect to be necessary for a while.
4601	*/
4602	VMMR3DECL(int) PGMR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
4603	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
4604	{
4605	Log(("PGMR3PhysRomRegister: pDevIns=%p GCPhys=%RGp(-%RGp) cb=%RGp pvBinary=%p cbBinary=%#x fFlags=%#x pszDesc=%s\n",
4606	pDevIns, GCPhys, GCPhys + cb, cb, pvBinary, cbBinary, fFlags, pszDesc));
4607	pgmLock(pVM);
4608	int rc = pgmR3PhysRomRegisterLocked(pVM, pDevIns, GCPhys, cb, pvBinary, cbBinary, fFlags, pszDesc);
4609	pgmUnlock(pVM);
4610	return rc;
4611	}
4612
4613
4614	/**
4615	* Called by PGMR3MemSetup to reset the shadow, switch to the virgin, and verify
4616	* that the virgin part is untouched.
4617	*
4618	* This is done after the normal memory has been cleared.
4619	*
4620	* ASSUMES that the caller owns the PGM lock.
4621	*
4622	* @param pVM The cross context VM structure.
4623	*/
4624	int pgmR3PhysRomReset(PVM pVM)
4625	{
4626	PGM_LOCK_ASSERT_OWNER(pVM);
4627	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4628	{
4629	const uint32_t cPages = pRom->cb >> PAGE_SHIFT;
4630
4631	if (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
4632	{
4633	/*
4634	* Reset the physical handler.
4635	*/
4636	int rc = PGMR3PhysRomProtect(pVM, pRom->GCPhys, pRom->cb, PGMROMPROT_READ_ROM_WRITE_IGNORE);
4637	AssertRCReturn(rc, rc);
4638
4639	/*
4640	* What we do with the shadow pages depends on the memory
4641	* preallocation option. If not enabled, we'll just throw
4642	* out all the dirty pages and replace them by the zero page.
4643	*/
4644	if (!pVM->pgm.s.fRamPreAlloc)
4645	{
4646	/* Free the dirty pages. */
4647	uint32_t cPendingPages = 0;
4648	PGMMFREEPAGESREQ pReq;
4649	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
4650	AssertRCReturn(rc, rc);
4651
4652	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4653	if ( !PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow)
4654	&& !PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow))
4655	{
4656	Assert(PGM_PAGE_GET_STATE(&pRom->aPages[iPage].Shadow) == PGM_PAGE_STATE_ALLOCATED);
4657	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, &pRom->aPages[iPage].Shadow,
4658	pRom->GCPhys + (iPage << PAGE_SHIFT),
4659	(PGMPAGETYPE)PGM_PAGE_GET_TYPE(&pRom->aPages[iPage].Shadow));
4660	AssertLogRelRCReturn(rc, rc);
4661	}
4662
4663	if (cPendingPages)
4664	{
4665	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
4666	AssertLogRelRCReturn(rc, rc);
4667	}
4668	GMMR3FreePagesCleanup(pReq);
4669	}
4670	else
4671	{
4672	/* clear all the shadow pages. */
4673	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4674	{
4675	if (PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow))
4676	continue;
4677	Assert(!PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow));
4678	void *pvDstPage;
4679	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4680	rc = pgmPhysPageMakeWritableAndMap(pVM, &pRom->aPages[iPage].Shadow, GCPhys, &pvDstPage);
4681	if (RT_FAILURE(rc))
4682	break;
4683	ASMMemZeroPage(pvDstPage);
4684	}
4685	AssertRCReturn(rc, rc);
4686	}
4687	}
4688
4689	/*
4690	* Restore the original ROM pages after a saved state load.
4691	* Also, in strict builds check that ROM pages remain unmodified.
4692	*/
4693	#ifndef VBOX_STRICT
4694	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4695	#endif
4696	{
4697	size_t cbSrcLeft = pRom->cbOriginal;
4698	uint8_t const pbSrcPage = (uint8_t const )pRom->pvOriginal;
4699	uint32_t cRestored = 0;
4700	for (uint32_t iPage = 0; iPage < cPages && cbSrcLeft > 0; iPage++, pbSrcPage += PAGE_SIZE)
4701	{
4702	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
4703	void const *pvDstPage;
4704	int rc = pgmPhysPageMapReadOnly(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPage);
4705	if (RT_FAILURE(rc))
4706	break;
4707
4708	if (memcmp(pvDstPage, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE)))
4709	{
4710	if (pVM->pgm.s.fRestoreRomPagesOnReset)
4711	{
4712	void *pvDstPageW;
4713	rc = pgmPhysPageMap(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPageW);
4714	AssertLogRelRCReturn(rc, rc);
4715	memcpy(pvDstPageW, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE));
4716	cRestored++;
4717	}
4718	else
4719	LogRel(("pgmR3PhysRomReset: %RGp: ROM page changed (%s)\n", GCPhys, pRom->pszDesc));
4720	}
4721	cbSrcLeft -= RT_MIN(cbSrcLeft, PAGE_SIZE);
4722	}
4723	if (cRestored > 0)
4724	LogRel(("PGM: ROM \"%s\": Reloaded %u of %u pages.\n", pRom->pszDesc, cRestored, cPages));
4725	}
4726	}
4727
4728	/* Clear the ROM restore flag now as we only need to do this once after
4729	loading saved state. */
4730	pVM->pgm.s.fRestoreRomPagesOnReset = false;
4731
4732	return VINF_SUCCESS;
4733	}
4734
4735
4736	/**
4737	* Called by PGMR3Term to free resources.
4738	*
4739	* ASSUMES that the caller owns the PGM lock.
4740	*
4741	* @param pVM The cross context VM structure.
4742	*/
4743	void pgmR3PhysRomTerm(PVM pVM)
4744	{
4745	/*
4746	* Free the heap copy of the original bits.
4747	*/
4748	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4749	{
4750	if ( pRom->pvOriginal
4751	&& !(pRom->fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY))
4752	{
4753	RTMemFree((void *)pRom->pvOriginal);
4754	pRom->pvOriginal = NULL;
4755	}
4756	}
4757	}
4758
4759
4760	/**
4761	* Change the shadowing of a range of ROM pages.
4762	*
4763	* This is intended for implementing chipset specific memory registers
4764	* and will not be very strict about the input. It will silently ignore
4765	* any pages that are not the part of a shadowed ROM.
4766	*
4767	* @returns VBox status code.
4768	* @retval VINF_PGM_SYNC_CR3
4769	*
4770	* @param pVM The cross context VM structure.
4771	* @param GCPhys Where to start. Page aligned.
4772	* @param cb How much to change. Page aligned.
4773	* @param enmProt The new ROM protection.
4774	*/
4775	VMMR3DECL(int) PGMR3PhysRomProtect(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMROMPROT enmProt)
4776	{
4777	/*
4778	* Check input
4779	*/
4780	if (!cb)
4781	return VINF_SUCCESS;
4782	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4783	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
4784	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
4785	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
4786	AssertReturn(enmProt >= PGMROMPROT_INVALID && enmProt <= PGMROMPROT_END, VERR_INVALID_PARAMETER);
4787
4788	/*
4789	* Process the request.
4790	*/
4791	pgmLock(pVM);
4792	int rc = VINF_SUCCESS;
4793	bool fFlushTLB = false;
4794	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
4795	{
4796	if ( GCPhys <= pRom->GCPhysLast
4797	&& GCPhysLast >= pRom->GCPhys
4798	&& (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED))
4799	{
4800	/*
4801	* Iterate the relevant pages and make necessary the changes.
4802	*/
4803	bool fChanges = false;
4804	uint32_t const cPages = pRom->GCPhysLast <= GCPhysLast
4805	? pRom->cb >> PAGE_SHIFT
4806	: (GCPhysLast - pRom->GCPhys + 1) >> PAGE_SHIFT;
4807	for (uint32_t iPage = (GCPhys - pRom->GCPhys) >> PAGE_SHIFT;
4808	iPage < cPages;
4809	iPage++)
4810	{
4811	PPGMROMPAGE pRomPage = &pRom->aPages[iPage];
4812	if (PGMROMPROT_IS_ROM(pRomPage->enmProt) != PGMROMPROT_IS_ROM(enmProt))
4813	{
4814	fChanges = true;
4815
4816	/* flush references to the page. */
4817	PPGMPAGE pRamPage = pgmPhysGetPage(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT));
4818	int rc2 = pgmPoolTrackUpdateGCPhys(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT), pRamPage,
4819	true /fFlushPTEs/, &fFlushTLB);
4820	if (rc2 != VINF_SUCCESS && (rc == VINF_SUCCESS \|\| RT_FAILURE(rc2)))
4821	rc = rc2;
4822	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pRamPage);
4823
4824	PPGMPAGE pOld = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Virgin : &pRomPage->Shadow;
4825	PPGMPAGE pNew = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Shadow : &pRomPage->Virgin;
4826
4827	pOld = pRamPage;
4828	pRamPage = pNew;
4829	/** @todo preserve the volatile flags (handlers) when these have been moved out of HCPhys! */
4830
4831	/* Tell NEM about the backing and protection change. */
4832	if (VM_IS_NEM_ENABLED(pVM))
4833	{
4834	PGMPAGETYPE enmType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pNew);
4835	NEMHCNotifyPhysPageChanged(pVM, GCPhys, PGM_PAGE_GET_HCPHYS(pOld), PGM_PAGE_GET_HCPHYS(pNew),
4836	pgmPhysPageCalcNemProtection(pRamPage, enmType), enmType, &u2State);
4837	PGM_PAGE_SET_NEM_STATE(pRamPage, u2State);
4838	}
4839	}
4840	pRomPage->enmProt = enmProt;
4841	}
4842
4843	/*
4844	* Reset the access handler if we made changes, no need
4845	* to optimize this.
4846	*/
4847	if (fChanges)
4848	{
4849	int rc2 = PGMHandlerPhysicalReset(pVM, pRom->GCPhys);
4850	if (RT_FAILURE(rc2))
4851	{
4852	pgmUnlock(pVM);
4853	AssertRC(rc);
4854	return rc2;
4855	}
4856	}
4857
4858	/* Advance - cb isn't updated. */
4859	GCPhys = pRom->GCPhys + (cPages << PAGE_SHIFT);
4860	}
4861	}
4862	pgmUnlock(pVM);
4863	if (fFlushTLB)
4864	PGM_INVL_ALL_VCPU_TLBS(pVM);
4865
4866	return rc;
4867	}
4868
4869
4870	/**
4871	* Sets the Address Gate 20 state.
4872	*
4873	* @param pVCpu The cross context virtual CPU structure.
4874	* @param fEnable True if the gate should be enabled.
4875	* False if the gate should be disabled.
4876	*/
4877	VMMDECL(void) PGMR3PhysSetA20(PVMCPU pVCpu, bool fEnable)
4878	{
4879	LogFlow(("PGMR3PhysSetA20 %d (was %d)\n", fEnable, pVCpu->pgm.s.fA20Enabled));
4880	if (pVCpu->pgm.s.fA20Enabled != fEnable)
4881	{
4882	#ifdef VBOX_WITH_NESTED_HWVIRT_VMX
4883	PCCPUMCTX pCtx = CPUMQueryGuestCtxPtr(pVCpu);
4884	if ( CPUMIsGuestInVmxRootMode(pCtx)
4885	&& !fEnable)
4886	{
4887	Log(("Cannot enter A20M mode while in VMX root mode\n"));
4888	return;
4889	}
4890	#endif
4891	pVCpu->pgm.s.fA20Enabled = fEnable;
4892	pVCpu->pgm.s.GCPhysA20Mask = ~((RTGCPHYS)!fEnable << 20);
4893	#ifdef VBOX_WITH_REM
4894	REMR3A20Set(pVCpu->pVMR3, pVCpu, fEnable);
4895	#endif
4896	NEMR3NotifySetA20(pVCpu, fEnable);
4897	#ifdef PGM_WITH_A20
4898	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_UPDATE_PAGE_BIT_VIRTUAL;
4899	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
4900	pgmR3RefreshShadowModeAfterA20Change(pVCpu);
4901	HMFlushTlb(pVCpu);
4902	#endif
4903	IEMTlbInvalidateAllPhysical(pVCpu);
4904	STAM_REL_COUNTER_INC(&pVCpu->pgm.s.cA20Changes);
4905	}
4906	}
4907
4908
4909	/**
4910	* Tree enumeration callback for dealing with age rollover.
4911	* It will perform a simple compression of the current age.
4912	*/
4913	static DECLCALLBACK(int) pgmR3PhysChunkAgeingRolloverCallback(PAVLU32NODECORE pNode, void *pvUser)
4914	{
4915	/* Age compression - ASSUMES iNow == 4. */
4916	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
4917	if (pChunk->iLastUsed >= UINT32_C(0xffffff00))
4918	pChunk->iLastUsed = 3;
4919	else if (pChunk->iLastUsed >= UINT32_C(0xfffff000))
4920	pChunk->iLastUsed = 2;
4921	else if (pChunk->iLastUsed)
4922	pChunk->iLastUsed = 1;
4923	else /* iLastUsed = 0 */
4924	pChunk->iLastUsed = 4;
4925
4926	NOREF(pvUser);
4927	return 0;
4928	}
4929
4930
4931	/**
4932	* The structure passed in the pvUser argument of pgmR3PhysChunkUnmapCandidateCallback().
4933	*/
4934	typedef struct PGMR3PHYSCHUNKUNMAPCB
4935	{
4936	PVM pVM; /*< Pointer to the VM. /
4937	PPGMCHUNKR3MAP pChunk; /*< The chunk to unmap. /
4938	} PGMR3PHYSCHUNKUNMAPCB, *PPGMR3PHYSCHUNKUNMAPCB;
4939
4940
4941	/**
4942	* Callback used to find the mapping that's been unused for
4943	* the longest time.
4944	*/
4945	static DECLCALLBACK(int) pgmR3PhysChunkUnmapCandidateCallback(PAVLU32NODECORE pNode, void *pvUser)
4946	{
4947	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
4948	PPGMR3PHYSCHUNKUNMAPCB pArg = (PPGMR3PHYSCHUNKUNMAPCB)pvUser;
4949
4950	/*
4951	* Check for locks and compare when last used.
4952	*/
4953	if (pChunk->cRefs)
4954	return 0;
4955	if (pChunk->cPermRefs)
4956	return 0;
4957	if ( pArg->pChunk
4958	&& pChunk->iLastUsed >= pArg->pChunk->iLastUsed)
4959	return 0;
4960
4961	/*
4962	* Check that it's not in any of the TLBs.
4963	*/
4964	PVM pVM = pArg->pVM;
4965	if ( pVM->pgm.s.ChunkR3Map.Tlb.aEntries[PGM_CHUNKR3MAPTLB_IDX(pChunk->Core.Key)].idChunk
4966	== pChunk->Core.Key)
4967	{
4968	pChunk = NULL;
4969	return 0;
4970	}
4971	#ifdef VBOX_STRICT
4972	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
4973	{
4974	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk != pChunk);
4975	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk != pChunk->Core.Key);
4976	}
4977	#endif
4978
4979	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.PhysTlbHC.aEntries); i++)
4980	if (pVM->pgm.s.PhysTlbHC.aEntries[i].pMap == pChunk)
4981	return 0;
4982
4983	pArg->pChunk = pChunk;
4984	return 0;
4985	}
4986
4987
4988	/**
4989	* Finds a good candidate for unmapping when the ring-3 mapping cache is full.
4990	*
4991	* The candidate will not be part of any TLBs, so no need to flush
4992	* anything afterwards.
4993	*
4994	* @returns Chunk id.
4995	* @param pVM The cross context VM structure.
4996	*/
4997	static int32_t pgmR3PhysChunkFindUnmapCandidate(PVM pVM)
4998	{
4999	PGM_LOCK_ASSERT_OWNER(pVM);
5000
5001	/*
5002	* Enumerate the age tree starting with the left most node.
5003	*/
5004	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
5005	PGMR3PHYSCHUNKUNMAPCB Args;
5006	Args.pVM = pVM;
5007	Args.pChunk = NULL;
5008	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkUnmapCandidateCallback, &Args);
5009	Assert(Args.pChunk);
5010	if (Args.pChunk)
5011	{
5012	Assert(Args.pChunk->cRefs == 0);
5013	Assert(Args.pChunk->cPermRefs == 0);
5014	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
5015	return Args.pChunk->Core.Key;
5016	}
5017
5018	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
5019	return INT32_MAX;
5020	}
5021
5022
5023	/**
5024	* Rendezvous callback used by pgmR3PhysUnmapChunk that unmaps a chunk
5025	*
5026	* This is only called on one of the EMTs while the other ones are waiting for
5027	* it to complete this function.
5028	*
5029	* @returns VINF_SUCCESS (VBox strict status code).
5030	* @param pVM The cross context VM structure.
5031	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
5032	* @param pvUser User pointer. Unused
5033	*
5034	*/
5035	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysUnmapChunkRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
5036	{
5037	int rc = VINF_SUCCESS;
5038	pgmLock(pVM);
5039	NOREF(pVCpu); NOREF(pvUser);
5040
5041	if (pVM->pgm.s.ChunkR3Map.c >= pVM->pgm.s.ChunkR3Map.cMax)
5042	{
5043	/* Flush the pgm pool cache; call the internal rendezvous handler as we're already in a rendezvous handler here. */
5044	/** @todo also not really efficient to unmap a chunk that contains PD
5045	* or PT pages. */
5046	pgmR3PoolClearAllRendezvous(pVM, &pVM->aCpus[0], NULL /* no need to flush the REM TLB as we already did that above */);
5047
5048	/*
5049	* Request the ring-0 part to unmap a chunk to make space in the mapping cache.
5050	*/
5051	GMMMAPUNMAPCHUNKREQ Req;
5052	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
5053	Req.Hdr.cbReq = sizeof(Req);
5054	Req.pvR3 = NULL;
5055	Req.idChunkMap = NIL_GMM_CHUNKID;
5056	Req.idChunkUnmap = pgmR3PhysChunkFindUnmapCandidate(pVM);
5057	if (Req.idChunkUnmap != INT32_MAX)
5058	{
5059	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkUnmap, a);
5060	rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
5061	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkUnmap, a);
5062	if (RT_SUCCESS(rc))
5063	{
5064	/*
5065	* Remove the unmapped one.
5066	*/
5067	PPGMCHUNKR3MAP pUnmappedChunk = (PPGMCHUNKR3MAP)RTAvlU32Remove(&pVM->pgm.s.ChunkR3Map.pTree, Req.idChunkUnmap);
5068	AssertRelease(pUnmappedChunk);
5069	AssertRelease(!pUnmappedChunk->cRefs);
5070	AssertRelease(!pUnmappedChunk->cPermRefs);
5071	pUnmappedChunk->pv = NULL;
5072	pUnmappedChunk->Core.Key = UINT32_MAX;
5073	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
5074	MMR3HeapFree(pUnmappedChunk);
5075	#else
5076	MMR3UkHeapFree(pVM, pUnmappedChunk, MM_TAG_PGM_CHUNK_MAPPING);
5077	#endif
5078	pVM->pgm.s.ChunkR3Map.c--;
5079	pVM->pgm.s.cUnmappedChunks++;
5080
5081	/*
5082	* Flush dangling PGM pointers (R3 & R0 ptrs to GC physical addresses).
5083	*/
5084	/** @todo We should not flush chunks which include cr3 mappings. */
5085	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
5086	{
5087	PPGMCPU pPGM = &pVM->aCpus[idCpu].pgm.s;
5088
5089	pPGM->pGst32BitPdR3 = NULL;
5090	pPGM->pGstPaePdptR3 = NULL;
5091	pPGM->pGstAmd64Pml4R3 = NULL;
5092	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
5093	pPGM->pGst32BitPdR0 = NIL_RTR0PTR;
5094	pPGM->pGstPaePdptR0 = NIL_RTR0PTR;
5095	pPGM->pGstAmd64Pml4R0 = NIL_RTR0PTR;
5096	#endif
5097	for (unsigned i = 0; i < RT_ELEMENTS(pPGM->apGstPaePDsR3); i++)
5098	{
5099	pPGM->apGstPaePDsR3[i] = NULL;
5100	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
5101	pPGM->apGstPaePDsR0[i] = NIL_RTR0PTR;
5102	#endif
5103	}
5104
5105	/* Flush REM TLBs. */
5106	CPUMSetChangedFlags(&pVM->aCpus[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
5107	}
5108	#ifdef VBOX_WITH_REM
5109	/* Flush REM translation blocks. */
5110	REMFlushTBs(pVM);
5111	#endif
5112	}
5113	}
5114	}
5115	pgmUnlock(pVM);
5116	return rc;
5117	}
5118
5119	/**
5120	* Unmap a chunk to free up virtual address space (request packet handler for pgmR3PhysChunkMap)
5121	*
5122	* @returns VBox status code.
5123	* @param pVM The cross context VM structure.
5124	*/
5125	void pgmR3PhysUnmapChunk(PVM pVM)
5126	{
5127	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysUnmapChunkRendezvous, NULL);
5128	AssertRC(rc);
5129	}
5130
5131
5132	/**
5133	* Maps the given chunk into the ring-3 mapping cache.
5134	*
5135	* This will call ring-0.
5136	*
5137	* @returns VBox status code.
5138	* @param pVM The cross context VM structure.
5139	* @param idChunk The chunk in question.
5140	* @param ppChunk Where to store the chunk tracking structure.
5141	*
5142	* @remarks Called from within the PGM critical section.
5143	* @remarks Can be called from any thread!
5144	*/
5145	int pgmR3PhysChunkMap(PVM pVM, uint32_t idChunk, PPPGMCHUNKR3MAP ppChunk)
5146	{
5147	int rc;
5148
5149	PGM_LOCK_ASSERT_OWNER(pVM);
5150
5151	/*
5152	* Move the chunk time forward.
5153	*/
5154	pVM->pgm.s.ChunkR3Map.iNow++;
5155	if (pVM->pgm.s.ChunkR3Map.iNow == 0)
5156	{
5157	pVM->pgm.s.ChunkR3Map.iNow = 4;
5158	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkAgeingRolloverCallback, NULL);
5159	}
5160
5161	/*
5162	* Allocate a new tracking structure first.
5163	*/
5164	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
5165	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3HeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk));
5166	#else
5167	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3UkHeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk), NULL);
5168	#endif
5169	AssertReturn(pChunk, VERR_NO_MEMORY);
5170	pChunk->Core.Key = idChunk;
5171	pChunk->iLastUsed = pVM->pgm.s.ChunkR3Map.iNow;
5172
5173	/*
5174	* Request the ring-0 part to map the chunk in question.
5175	*/
5176	GMMMAPUNMAPCHUNKREQ Req;
5177	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
5178	Req.Hdr.cbReq = sizeof(Req);
5179	Req.pvR3 = NULL;
5180	Req.idChunkMap = idChunk;
5181	Req.idChunkUnmap = NIL_GMM_CHUNKID;
5182
5183	/* Must be callable from any thread, so can't use VMMR3CallR0. */
5184	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkMap, a);
5185	rc = SUPR3CallVMMR0Ex(pVM->pVMR0, NIL_VMCPUID, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
5186	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkMap, a);
5187	if (RT_SUCCESS(rc))
5188	{
5189	pChunk->pv = Req.pvR3;
5190
5191	/*
5192	* If we're running out of virtual address space, then we should
5193	* unmap another chunk.
5194	*
5195	* Currently, an unmap operation requires that all other virtual CPUs
5196	* are idling and not by chance making use of the memory we're
5197	* unmapping. So, we create an async unmap operation here.
5198	*
5199	* Now, when creating or restoring a saved state this wont work very
5200	* well since we may want to restore all guest RAM + a little something.
5201	* So, we have to do the unmap synchronously. Fortunately for us
5202	* though, during these operations the other virtual CPUs are inactive
5203	* and it should be safe to do this.
5204	*/
5205	/** @todo Eventually we should lock all memory when used and do
5206	* map+unmap as one kernel call without any rendezvous or
5207	* other precautions. */
5208	if (pVM->pgm.s.ChunkR3Map.c + 1 >= pVM->pgm.s.ChunkR3Map.cMax)
5209	{
5210	switch (VMR3GetState(pVM))
5211	{
5212	case VMSTATE_LOADING:
5213	case VMSTATE_SAVING:
5214	{
5215	PVMCPU pVCpu = VMMGetCpu(pVM);
5216	if ( pVCpu
5217	&& pVM->pgm.s.cDeprecatedPageLocks == 0)
5218	{
5219	pgmR3PhysUnmapChunkRendezvous(pVM, pVCpu, NULL);
5220	break;
5221	}
5222	}
5223	RT_FALL_THRU();
5224	default:
5225	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysUnmapChunk, 1, pVM);
5226	AssertRC(rc);
5227	break;
5228	}
5229	}
5230
5231	/*
5232	* Update the tree. We must do this after any unmapping to make sure
5233	* the chunk we're going to return isn't unmapped by accident.
5234	*/
5235	AssertPtr(Req.pvR3);
5236	bool fRc = RTAvlU32Insert(&pVM->pgm.s.ChunkR3Map.pTree, &pChunk->Core);
5237	AssertRelease(fRc);
5238	pVM->pgm.s.ChunkR3Map.c++;
5239	pVM->pgm.s.cMappedChunks++;
5240	}
5241	else
5242	{
5243	/** @todo this may fail because of /proc/sys/vm/max_map_count, so we
5244	* should probably restrict ourselves on linux. */
5245	AssertRC(rc);
5246	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
5247	MMR3HeapFree(pChunk);
5248	#else
5249	MMR3UkHeapFree(pVM, pChunk, MM_TAG_PGM_CHUNK_MAPPING);
5250	#endif
5251	pChunk = NULL;
5252	}
5253
5254	*ppChunk = pChunk;
5255	return rc;
5256	}
5257
5258
5259	/**
5260	* For VMMCALLRING3_PGM_MAP_CHUNK, considered internal.
5261	*
5262	* @returns see pgmR3PhysChunkMap.
5263	* @param pVM The cross context VM structure.
5264	* @param idChunk The chunk to map.
5265	*/
5266	VMMR3DECL(int) PGMR3PhysChunkMap(PVM pVM, uint32_t idChunk)
5267	{
5268	PPGMCHUNKR3MAP pChunk;
5269	int rc;
5270
5271	pgmLock(pVM);
5272	rc = pgmR3PhysChunkMap(pVM, idChunk, &pChunk);
5273	pgmUnlock(pVM);
5274	return rc;
5275	}
5276
5277
5278	/**
5279	* Invalidates the TLB for the ring-3 mapping cache.
5280	*
5281	* @param pVM The cross context VM structure.
5282	*/
5283	VMMR3DECL(void) PGMR3PhysChunkInvalidateTLB(PVM pVM)
5284	{
5285	pgmLock(pVM);
5286	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
5287	{
5288	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk = NIL_GMM_CHUNKID;
5289	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk = NULL;
5290	}
5291	/* The page map TLB references chunks, so invalidate that one too. */
5292	pgmPhysInvalidatePageMapTLB(pVM);
5293	pgmUnlock(pVM);
5294	}
5295
5296
5297	/**
5298	* Response to VMMCALLRING3_PGM_ALLOCATE_LARGE_HANDY_PAGE to allocate a large
5299	* (2MB) page for use with a nested paging PDE.
5300	*
5301	* @returns The following VBox status codes.
5302	* @retval VINF_SUCCESS on success.
5303	* @retval VINF_EM_NO_MEMORY if we're out of memory.
5304	*
5305	* @param pVM The cross context VM structure.
5306	* @param GCPhys GC physical start address of the 2 MB range
5307	*/
5308	VMMR3DECL(int) PGMR3PhysAllocateLargeHandyPage(PVM pVM, RTGCPHYS GCPhys)
5309	{
5310	#ifdef PGM_WITH_LARGE_PAGES
5311	uint64_t u64TimeStamp1, u64TimeStamp2;
5312
5313	pgmLock(pVM);
5314
5315	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatAllocLargePage, a);
5316	u64TimeStamp1 = RTTimeMilliTS();
5317	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_LARGE_HANDY_PAGE, 0, NULL);
5318	u64TimeStamp2 = RTTimeMilliTS();
5319	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatAllocLargePage, a);
5320	if (RT_SUCCESS(rc))
5321	{
5322	Assert(pVM->pgm.s.cLargeHandyPages == 1);
5323
5324	uint32_t idPage = pVM->pgm.s.aLargeHandyPage[0].idPage;
5325	RTHCPHYS HCPhys = pVM->pgm.s.aLargeHandyPage[0].HCPhysGCPhys;
5326
5327	void *pv;
5328
5329	/* Map the large page into our address space.
5330	*
5331	* Note: assuming that within the 2 MB range:
5332	* - GCPhys + PAGE_SIZE = HCPhys + PAGE_SIZE (whole point of this exercise)
5333	* - user space mapping is continuous as well
5334	* - page id (GCPhys) + 1 = page id (GCPhys + PAGE_SIZE)
5335	*/
5336	rc = pgmPhysPageMapByPageID(pVM, idPage, HCPhys, &pv);
5337	AssertLogRelMsg(RT_SUCCESS(rc), ("idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n", idPage, HCPhys, rc));
5338
5339	if (RT_SUCCESS(rc))
5340	{
5341	/*
5342	* Clear the pages.
5343	*/
5344	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatClearLargePage, b);
5345	for (unsigned i = 0; i < _2M/PAGE_SIZE; i++)
5346	{
5347	ASMMemZeroPage(pv);
5348
5349	PPGMPAGE pPage;
5350	rc = pgmPhysGetPageEx(pVM, GCPhys, &pPage);
5351	AssertRC(rc);
5352
5353	Assert(PGM_PAGE_IS_ZERO(pPage));
5354	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatRZPageReplaceZero);
5355	pVM->pgm.s.cZeroPages--;
5356
5357	/*
5358	* Do the PGMPAGE modifications.
5359	*/
5360	pVM->pgm.s.cPrivatePages++;
5361	PGM_PAGE_SET_HCPHYS(pVM, pPage, HCPhys);
5362	PGM_PAGE_SET_PAGEID(pVM, pPage, idPage);
5363	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
5364	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_PDE);
5365	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
5366	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
5367
5368	/* Somewhat dirty assumption that page ids are increasing. */
5369	idPage++;
5370
5371	HCPhys += PAGE_SIZE;
5372	GCPhys += PAGE_SIZE;
5373
5374	pv = (void *)((uintptr_t)pv + PAGE_SIZE);
5375
5376	Log3(("PGMR3PhysAllocateLargePage: idPage=%#x HCPhys=%RGp\n", idPage, HCPhys));
5377	}
5378	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatClearLargePage, b);
5379
5380	/* Flush all TLBs. */
5381	PGM_INVL_ALL_VCPU_TLBS(pVM);
5382	pgmPhysInvalidatePageMapTLB(pVM);
5383	}
5384	pVM->pgm.s.cLargeHandyPages = 0;
5385	}
5386
5387	if (RT_SUCCESS(rc))
5388	{
5389	static uint32_t cTimeOut = 0;
5390	uint64_t u64TimeStampDelta = u64TimeStamp2 - u64TimeStamp1;
5391
5392	if (u64TimeStampDelta > 100)
5393	{
5394	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatLargePageOverflow);
5395	if ( ++cTimeOut > 10
5396	\|\| u64TimeStampDelta > 1000 /* more than one second forces an early retirement from allocating large pages. */)
5397	{
5398	/* If repeated attempts to allocate a large page takes more than 100 ms, then we fall back to normal 4k pages.
5399	* E.g. Vista 64 tries to move memory around, which takes a huge amount of time.
5400	*/
5401	LogRel(("PGMR3PhysAllocateLargePage: allocating large pages takes too long (last attempt %d ms; nr of timeouts %d); DISABLE\n", u64TimeStampDelta, cTimeOut));
5402	PGMSetLargePageUsage(pVM, false);
5403	}
5404	}
5405	else
5406	if (cTimeOut > 0)
5407	cTimeOut--;
5408	}
5409
5410	pgmUnlock(pVM);
5411	return rc;
5412	#else
5413	RT_NOREF(pVM, GCPhys);
5414	return VERR_NOT_IMPLEMENTED;
5415	#endif /* PGM_WITH_LARGE_PAGES */
5416	}
5417
5418
5419	/**
5420	* Response to VM_FF_PGM_NEED_HANDY_PAGES and VMMCALLRING3_PGM_ALLOCATE_HANDY_PAGES.
5421	*
5422	* This function will also work the VM_FF_PGM_NO_MEMORY force action flag, to
5423	* signal and clear the out of memory condition. When contracted, this API is
5424	* used to try clear the condition when the user wants to resume.
5425	*
5426	* @returns The following VBox status codes.
5427	* @retval VINF_SUCCESS on success. FFs cleared.
5428	* @retval VINF_EM_NO_MEMORY if we're out of memory. The FF is not cleared in
5429	* this case and it gets accompanied by VM_FF_PGM_NO_MEMORY.
5430	*
5431	* @param pVM The cross context VM structure.
5432	*
5433	* @remarks The VINF_EM_NO_MEMORY status is for the benefit of the FF processing
5434	* in EM.cpp and shouldn't be propagated outside TRPM, HM, EM and
5435	* pgmPhysEnsureHandyPage. There is one exception to this in the \#PF
5436	* handler.
5437	*/
5438	VMMR3DECL(int) PGMR3PhysAllocateHandyPages(PVM pVM)
5439	{
5440	pgmLock(pVM);
5441
5442	/*
5443	* Allocate more pages, noting down the index of the first new page.
5444	*/
5445	uint32_t iClear = pVM->pgm.s.cHandyPages;
5446	AssertMsgReturn(iClear <= RT_ELEMENTS(pVM->pgm.s.aHandyPages), ("%d", iClear), VERR_PGM_HANDY_PAGE_IPE);
5447	Log(("PGMR3PhysAllocateHandyPages: %d -> %d\n", iClear, RT_ELEMENTS(pVM->pgm.s.aHandyPages)));
5448	int rcAlloc = VINF_SUCCESS;
5449	int rcSeed = VINF_SUCCESS;
5450	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
5451	while (rc == VERR_GMM_SEED_ME)
5452	{
5453	void *pvChunk;
5454	rcAlloc = rc = SUPR3PageAlloc(GMM_CHUNK_SIZE >> PAGE_SHIFT, &pvChunk);
5455	if (RT_SUCCESS(rc))
5456	{
5457	rcSeed = rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_SEED_CHUNK, (uintptr_t)pvChunk, NULL);
5458	if (RT_FAILURE(rc))
5459	SUPR3PageFree(pvChunk, GMM_CHUNK_SIZE >> PAGE_SHIFT);
5460	}
5461	if (RT_SUCCESS(rc))
5462	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
5463	}
5464
5465	/** @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) */
5466	if ( rc == VERR_GMM_HIT_VM_ACCOUNT_LIMIT
5467	&& pVM->pgm.s.cHandyPages > 0)
5468	{
5469	/* Still handy pages left, so don't panic. */
5470	rc = VINF_SUCCESS;
5471	}
5472
5473	if (RT_SUCCESS(rc))
5474	{
5475	AssertMsg(rc == VINF_SUCCESS, ("%Rrc\n", rc));
5476	Assert(pVM->pgm.s.cHandyPages > 0);
5477	VM_FF_CLEAR(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
5478	VM_FF_CLEAR(pVM, VM_FF_PGM_NO_MEMORY);
5479
5480	#ifdef VBOX_STRICT
5481	uint32_t i;
5482	for (i = iClear; i < pVM->pgm.s.cHandyPages; i++)
5483	if ( pVM->pgm.s.aHandyPages[i].idPage == NIL_GMM_PAGEID
5484	\|\| pVM->pgm.s.aHandyPages[i].idSharedPage != NIL_GMM_PAGEID
5485	\|\| (pVM->pgm.s.aHandyPages[i].HCPhysGCPhys & PAGE_OFFSET_MASK))
5486	break;
5487	if (i != pVM->pgm.s.cHandyPages)
5488	{
5489	RTAssertMsg1Weak(NULL, __LINE__, __FILE__, __FUNCTION__);
5490	RTAssertMsg2Weak("i=%d iClear=%d cHandyPages=%d\n", i, iClear, pVM->pgm.s.cHandyPages);
5491	for (uint32_t j = iClear; j < pVM->pgm.s.cHandyPages; j++)
5492	RTAssertMsg2Add("%03d: idPage=%d HCPhysGCPhys=%RHp idSharedPage=%d%\n", j,
5493	pVM->pgm.s.aHandyPages[j].idPage,
5494	pVM->pgm.s.aHandyPages[j].HCPhysGCPhys,
5495	pVM->pgm.s.aHandyPages[j].idSharedPage,
5496	j == i ? " <---" : "");
5497	RTAssertPanic();
5498	}
5499	#endif
5500	/*
5501	* Clear the pages.
5502	*/
5503	while (iClear < pVM->pgm.s.cHandyPages)
5504	{
5505	PGMMPAGEDESC pPage = &pVM->pgm.s.aHandyPages[iClear];
5506	void *pv;
5507	rc = pgmPhysPageMapByPageID(pVM, pPage->idPage, pPage->HCPhysGCPhys, &pv);
5508	AssertLogRelMsgBreak(RT_SUCCESS(rc),
5509	("%u/%u: idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n",
5510	iClear, pVM->pgm.s.cHandyPages, pPage->idPage, pPage->HCPhysGCPhys, rc));
5511	ASMMemZeroPage(pv);
5512	iClear++;
5513	Log3(("PGMR3PhysAllocateHandyPages: idPage=%#x HCPhys=%RGp\n", pPage->idPage, pPage->HCPhysGCPhys));
5514	}
5515	}
5516	else
5517	{
5518	uint64_t cAllocPages, cMaxPages, cBalloonPages;
5519
5520	/*
5521	* We should never get here unless there is a genuine shortage of
5522	* memory (or some internal error). Flag the error so the VM can be
5523	* suspended ASAP and the user informed. If we're totally out of
5524	* handy pages we will return failure.
5525	*/
5526	/* Report the failure. */
5527	LogRel(("PGM: Failed to procure handy pages; rc=%Rrc rcAlloc=%Rrc rcSeed=%Rrc cHandyPages=%#x\n"
5528	" cAllPages=%#x cPrivatePages=%#x cSharedPages=%#x cZeroPages=%#x\n",
5529	rc, rcAlloc, rcSeed,
5530	pVM->pgm.s.cHandyPages,
5531	pVM->pgm.s.cAllPages,
5532	pVM->pgm.s.cPrivatePages,
5533	pVM->pgm.s.cSharedPages,
5534	pVM->pgm.s.cZeroPages));
5535
5536	if (GMMR3QueryMemoryStats(pVM, &cAllocPages, &cMaxPages, &cBalloonPages) == VINF_SUCCESS)
5537	{
5538	LogRel(("GMM: Statistics:\n"
5539	" Allocated pages: %RX64\n"
5540	" Maximum pages: %RX64\n"
5541	" Ballooned pages: %RX64\n", cAllocPages, cMaxPages, cBalloonPages));
5542	}
5543
5544	if ( rc != VERR_NO_MEMORY
5545	&& rc != VERR_NO_PHYS_MEMORY
5546	&& rc != VERR_LOCK_FAILED)
5547	{
5548	for (uint32_t i = 0; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
5549	{
5550	LogRel(("PGM: aHandyPages[#%#04x] = {.HCPhysGCPhys=%RHp, .idPage=%#08x, .idSharedPage=%#08x}\n",
5551	i, pVM->pgm.s.aHandyPages[i].HCPhysGCPhys, pVM->pgm.s.aHandyPages[i].idPage,
5552	pVM->pgm.s.aHandyPages[i].idSharedPage));
5553	uint32_t const idPage = pVM->pgm.s.aHandyPages[i].idPage;
5554	if (idPage != NIL_GMM_PAGEID)
5555	{
5556	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
5557	pRam;
5558	pRam = pRam->pNextR3)
5559	{
5560	uint32_t const cPages = pRam->cb >> PAGE_SHIFT;
5561	for (uint32_t iPage = 0; iPage < cPages; iPage++)
5562	if (PGM_PAGE_GET_PAGEID(&pRam->aPages[iPage]) == idPage)
5563	LogRel(("PGM: Used by %RGp %R[pgmpage] (%s)\n",
5564	pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pRam->aPages[iPage], pRam->pszDesc));
5565	}
5566	}
5567	}
5568	}
5569
5570	if (rc == VERR_NO_MEMORY)
5571	{
5572	uint64_t cbHostRamAvail = 0;
5573	int rc2 = RTSystemQueryAvailableRam(&cbHostRamAvail);
5574	if (RT_SUCCESS(rc2))
5575	LogRel(("Host RAM: %RU64MB available\n", cbHostRamAvail / _1M));
5576	else
5577	LogRel(("Cannot determine the amount of available host memory\n"));
5578	}
5579
5580	/* Set the FFs and adjust rc. */
5581	VM_FF_SET(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
5582	VM_FF_SET(pVM, VM_FF_PGM_NO_MEMORY);
5583	if ( rc == VERR_NO_MEMORY
5584	\|\| rc == VERR_NO_PHYS_MEMORY
5585	\|\| rc == VERR_LOCK_FAILED)
5586	rc = VINF_EM_NO_MEMORY;
5587	}
5588
5589	pgmUnlock(pVM);
5590	return rc;
5591	}
5592
5593
5594	/**
5595	* Frees the specified RAM page and replaces it with the ZERO page.
5596	*
5597	* This is used by ballooning, remapping MMIO2, RAM reset and state loading.
5598	*
5599	* @param pVM The cross context VM structure.
5600	* @param pReq Pointer to the request.
5601	* @param pcPendingPages Where the number of pages waiting to be freed are
5602	* kept. This will normally be incremented.
5603	* @param pPage Pointer to the page structure.
5604	* @param GCPhys The guest physical address of the page, if applicable.
5605	* @param enmNewType New page type for NEM notification, since several
5606	* callers will change the type upon successful return.
5607	*
5608	* @remarks The caller must own the PGM lock.
5609	*/
5610	int pgmPhysFreePage(PVM pVM, PGMMFREEPAGESREQ pReq, uint32_t *pcPendingPages, PPGMPAGE pPage, RTGCPHYS GCPhys,
5611	PGMPAGETYPE enmNewType)
5612	{
5613	/*
5614	* Assert sanity.
5615	*/
5616	PGM_LOCK_ASSERT_OWNER(pVM);
5617	if (RT_UNLIKELY( PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM
5618	&& PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_ROM_SHADOW))
5619	{
5620	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
5621	return VMSetError(pVM, VERR_PGM_PHYS_NOT_RAM, RT_SRC_POS, "GCPhys=%RGp type=%d", GCPhys, PGM_PAGE_GET_TYPE(pPage));
5622	}
5623
5624	/** @todo What about ballooning of large pages??! */
5625	Assert( PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE
5626	&& PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE_DISABLED);
5627
5628	if ( PGM_PAGE_IS_ZERO(pPage)
5629	\|\| PGM_PAGE_IS_BALLOONED(pPage))
5630	return VINF_SUCCESS;
5631
5632	const uint32_t idPage = PGM_PAGE_GET_PAGEID(pPage);
5633	Log3(("pgmPhysFreePage: idPage=%#x GCPhys=%RGp pPage=%R[pgmpage]\n", idPage, GCPhys, pPage));
5634	if (RT_UNLIKELY( idPage == NIL_GMM_PAGEID
5635	\|\| idPage > GMM_PAGEID_LAST
5636	\|\| PGM_PAGE_GET_CHUNKID(pPage) == NIL_GMM_CHUNKID))
5637	{
5638	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
5639	return VMSetError(pVM, VERR_PGM_PHYS_INVALID_PAGE_ID, RT_SRC_POS, "GCPhys=%RGp idPage=%#x", GCPhys, pPage);
5640	}
5641	const RTHCPHYS HCPhysPrev = PGM_PAGE_GET_HCPHYS(pPage);
5642
5643	/* update page count stats. */
5644	if (PGM_PAGE_IS_SHARED(pPage))
5645	pVM->pgm.s.cSharedPages--;
5646	else
5647	pVM->pgm.s.cPrivatePages--;
5648	pVM->pgm.s.cZeroPages++;
5649
5650	/* Deal with write monitored pages. */
5651	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
5652	{
5653	PGM_PAGE_SET_WRITTEN_TO(pVM, pPage);
5654	pVM->pgm.s.cWrittenToPages++;
5655	}
5656
5657	/*
5658	* pPage = ZERO page.
5659	*/
5660	PGM_PAGE_SET_HCPHYS(pVM, pPage, pVM->pgm.s.HCPhysZeroPg);
5661	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
5662	PGM_PAGE_SET_PAGEID(pVM, pPage, NIL_GMM_PAGEID);
5663	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
5664	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
5665	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
5666
5667	/* Flush physical page map TLB entry. */
5668	pgmPhysInvalidatePageMapTLBEntry(pVM, GCPhys);
5669
5670	/* Notify NEM. */
5671	/** @todo consider doing batch NEM notifications. */
5672	if (VM_IS_NEM_ENABLED(pVM))
5673	{
5674	uint8_t u2State = PGM_PAGE_GET_NEM_STATE(pPage);
5675	NEMHCNotifyPhysPageChanged(pVM, GCPhys, HCPhysPrev, pVM->pgm.s.HCPhysZeroPg,
5676	pgmPhysPageCalcNemProtection(pPage, enmNewType), enmNewType, &u2State);
5677	PGM_PAGE_SET_NEM_STATE(pPage, u2State);
5678	}
5679
5680	/*
5681	* Make sure it's not in the handy page array.
5682	*/
5683	for (uint32_t i = pVM->pgm.s.cHandyPages; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
5684	{
5685	if (pVM->pgm.s.aHandyPages[i].idPage == idPage)
5686	{
5687	pVM->pgm.s.aHandyPages[i].idPage = NIL_GMM_PAGEID;
5688	break;
5689	}
5690	if (pVM->pgm.s.aHandyPages[i].idSharedPage == idPage)
5691	{
5692	pVM->pgm.s.aHandyPages[i].idSharedPage = NIL_GMM_PAGEID;
5693	break;
5694	}
5695	}
5696
5697	/*
5698	* Push it onto the page array.
5699	*/
5700	uint32_t iPage = *pcPendingPages;
5701	Assert(iPage < PGMPHYS_FREE_PAGE_BATCH_SIZE);
5702	*pcPendingPages += 1;
5703
5704	pReq->aPages[iPage].idPage = idPage;
5705
5706	if (iPage + 1 < PGMPHYS_FREE_PAGE_BATCH_SIZE)
5707	return VINF_SUCCESS;
5708
5709	/*
5710	* Flush the pages.
5711	*/
5712	int rc = GMMR3FreePagesPerform(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE);
5713	if (RT_SUCCESS(rc))
5714	{
5715	GMMR3FreePagesRePrep(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
5716	*pcPendingPages = 0;
5717	}
5718	return rc;
5719	}
5720
5721
5722	/**
5723	* Converts a GC physical address to a HC ring-3 pointer, with some
5724	* additional checks.
5725	*
5726	* @returns VBox status code.
5727	* @retval VINF_SUCCESS on success.
5728	* @retval VINF_PGM_PHYS_TLB_CATCH_WRITE and *ppv set if the page has a write
5729	* access handler of some kind.
5730	* @retval VERR_PGM_PHYS_TLB_CATCH_ALL if the page has a handler catching all
5731	* accesses or is odd in any way.
5732	* @retval VERR_PGM_PHYS_TLB_UNASSIGNED if the page doesn't exist.
5733	*
5734	* @param pVM The cross context VM structure.
5735	* @param GCPhys The GC physical address to convert. Since this is only
5736	* used for filling the REM TLB, the A20 mask must be
5737	* applied before calling this API.
5738	* @param fWritable Whether write access is required.
5739	* @param ppv Where to store the pointer corresponding to GCPhys on
5740	* success.
5741	*/
5742	VMMR3DECL(int) PGMR3PhysTlbGCPhys2Ptr(PVM pVM, RTGCPHYS GCPhys, bool fWritable, void **ppv)
5743	{
5744	pgmLock(pVM);
5745	PGM_A20_ASSERT_MASKED(VMMGetCpu(pVM), GCPhys);
5746
5747	PPGMRAMRANGE pRam;
5748	PPGMPAGE pPage;
5749	int rc = pgmPhysGetPageAndRangeEx(pVM, GCPhys, &pPage, &pRam);
5750	if (RT_SUCCESS(rc))
5751	{
5752	if (PGM_PAGE_IS_BALLOONED(pPage))
5753	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5754	else if (!PGM_PAGE_HAS_ANY_HANDLERS(pPage))
5755	rc = VINF_SUCCESS;
5756	else
5757	{
5758	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)) /* catches MMIO */
5759	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
5760	else if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
5761	{
5762	/** @todo Handle TLB loads of virtual handlers so ./test.sh can be made to work
5763	* in -norawr0 mode. */
5764	if (fWritable)
5765	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5766	}
5767	else
5768	{
5769	/* Temporarily disabled physical handler(s), since the recompiler
5770	doesn't get notified when it's reset we'll have to pretend it's
5771	operating normally. */
5772	if (pgmHandlerPhysicalIsAll(pVM, GCPhys))
5773	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
5774	else
5775	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
5776	}
5777	}
5778	if (RT_SUCCESS(rc))
5779	{
5780	int rc2;
5781
5782	/* Make sure what we return is writable. */
5783	if (fWritable)
5784	switch (PGM_PAGE_GET_STATE(pPage))
5785	{
5786	case PGM_PAGE_STATE_ALLOCATED:
5787	break;
5788	case PGM_PAGE_STATE_BALLOONED:
5789	AssertFailed();
5790	break;
5791	case PGM_PAGE_STATE_ZERO:
5792	case PGM_PAGE_STATE_SHARED:
5793	if (rc == VINF_PGM_PHYS_TLB_CATCH_WRITE)
5794	break;
5795	RT_FALL_THRU();
5796	case PGM_PAGE_STATE_WRITE_MONITORED:
5797	rc2 = pgmPhysPageMakeWritable(pVM, pPage, GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK);
5798	AssertLogRelRCReturn(rc2, rc2);
5799	break;
5800	}
5801
5802	/* Get a ring-3 mapping of the address. */
5803	PPGMPAGER3MAPTLBE pTlbe;
5804	rc2 = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
5805	AssertLogRelRCReturn(rc2, rc2);
5806	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
5807	/** @todo mapping/locking hell; this isn't horribly efficient since
5808	* pgmPhysPageLoadIntoTlb will repeat the lookup we've done here. */
5809
5810	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage] ppv=%p\n", GCPhys, rc, pPage, ppv));
5811	}
5812	else
5813	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage]\n", GCPhys, rc, pPage));
5814
5815	/* else: handler catching all access, no pointer returned. */
5816	}
5817	else
5818	rc = VERR_PGM_PHYS_TLB_UNASSIGNED;
5819
5820	pgmUnlock(pVM);
5821	return rc;
5822	}
5823

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PGMPhys.cpp@ 77758

Download in other formats: