PGMPhys.cpp@ 49966

Last change on this file since 49966 was 48728, checked in by vboxsync, 11 years ago
PGMPhys.cpp: DEBUG_bird hack for loading new ROM on reset (snapshot fun).
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 169.5 KB

Line
1	/* $Id: PGMPhys.cpp 48728 2013-09-27 11:04:41Z vboxsync $ */
2	/** @file
3	* PGM - Page Manager and Monitor, Physical Memory Addressing.
4	*/
5
6	/*
7	* Copyright (C) 2006-2013 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/iom.h>
25	#include <VBox/vmm/mm.h>
26	#include <VBox/vmm/stam.h>
27	#ifdef VBOX_WITH_REM
28	# include <VBox/vmm/rem.h>
29	#endif
30	#include <VBox/vmm/pdmdev.h>
31	#include "PGMInternal.h"
32	#include <VBox/vmm/vm.h>
33	#include <VBox/vmm/uvm.h>
34	#include "PGMInline.h"
35	#include <VBox/sup.h>
36	#include <VBox/param.h>
37	#include <VBox/err.h>
38	#include <VBox/log.h>
39	#include <iprt/assert.h>
40	#include <iprt/alloc.h>
41	#include <iprt/asm.h>
42	#ifdef VBOX_STRICT
43	# include <iprt/crc.h>
44	#endif
45	#include <iprt/thread.h>
46	#include <iprt/string.h>
47	#include <iprt/system.h>
48
49
50	/*******************************************************************************
51	* Defined Constants And Macros *
52	*******************************************************************************/
53	/** The number of pages to free in one batch. */
54	#define PGMPHYS_FREE_PAGE_BATCH_SIZE 128
55
56
57	/*******************************************************************************
58	* Internal Functions *
59	*******************************************************************************/
60	static DECLCALLBACK(int) pgmR3PhysRomWriteHandler(PVM pVM, RTGCPHYS GCPhys, void pvPhys, void pvBuf, size_t cbBuf, PGMACCESSTYPE enmAccessType, void *pvUser);
61
62
63	/*
64	* PGMR3PhysReadU8-64
65	* PGMR3PhysWriteU8-64
66	*/
67	#define PGMPHYSFN_READNAME PGMR3PhysReadU8
68	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU8
69	#define PGMPHYS_DATASIZE 1
70	#define PGMPHYS_DATATYPE uint8_t
71	#include "PGMPhysRWTmpl.h"
72
73	#define PGMPHYSFN_READNAME PGMR3PhysReadU16
74	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU16
75	#define PGMPHYS_DATASIZE 2
76	#define PGMPHYS_DATATYPE uint16_t
77	#include "PGMPhysRWTmpl.h"
78
79	#define PGMPHYSFN_READNAME PGMR3PhysReadU32
80	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU32
81	#define PGMPHYS_DATASIZE 4
82	#define PGMPHYS_DATATYPE uint32_t
83	#include "PGMPhysRWTmpl.h"
84
85	#define PGMPHYSFN_READNAME PGMR3PhysReadU64
86	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU64
87	#define PGMPHYS_DATASIZE 8
88	#define PGMPHYS_DATATYPE uint64_t
89	#include "PGMPhysRWTmpl.h"
90
91
92	/**
93	* EMT worker for PGMR3PhysReadExternal.
94	*/
95	static DECLCALLBACK(int) pgmR3PhysReadExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, void *pvBuf, size_t cbRead)
96	{
97	PGMPhysRead(pVM, *pGCPhys, pvBuf, cbRead);
98	return VINF_SUCCESS;
99	}
100
101
102	/**
103	* Read from physical memory, external users.
104	*
105	* @returns VBox status code.
106	* @retval VINF_SUCCESS.
107	*
108	* @param pVM Pointer to the VM.
109	* @param GCPhys Physical address to read from.
110	* @param pvBuf Where to read into.
111	* @param cbRead How many bytes to read.
112	*
113	* @thread Any but EMTs.
114	*/
115	VMMR3DECL(int) PGMR3PhysReadExternal(PVM pVM, RTGCPHYS GCPhys, void *pvBuf, size_t cbRead)
116	{
117	VM_ASSERT_OTHER_THREAD(pVM);
118
119	AssertMsgReturn(cbRead > 0, ("don't even think about reading zero bytes!\n"), VINF_SUCCESS);
120	LogFlow(("PGMR3PhysReadExternal: %RGp %d\n", GCPhys, cbRead));
121
122	pgmLock(pVM);
123
124	/*
125	* Copy loop on ram ranges.
126	*/
127	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
128	for (;;)
129	{
130	/* Inside range or not? */
131	if (pRam && GCPhys >= pRam->GCPhys)
132	{
133	/*
134	* Must work our way thru this page by page.
135	*/
136	RTGCPHYS off = GCPhys - pRam->GCPhys;
137	while (off < pRam->cb)
138	{
139	unsigned iPage = off >> PAGE_SHIFT;
140	PPGMPAGE pPage = &pRam->aPages[iPage];
141
142	/*
143	* If the page has an ALL access handler, we'll have to
144	* delegate the job to EMT.
145	*/
146	if ( PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
147	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
148	{
149	pgmUnlock(pVM);
150
151	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysReadExternalEMT, 4,
152	pVM, &GCPhys, pvBuf, cbRead);
153	}
154	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
155
156	/*
157	* Simple stuff, go ahead.
158	*/
159	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
160	if (cb > cbRead)
161	cb = cbRead;
162	PGMPAGEMAPLOCK PgMpLck;
163	const void *pvSrc;
164	int rc = pgmPhysGCPhys2CCPtrInternalReadOnly(pVM, pPage, pRam->GCPhys + off, &pvSrc, &PgMpLck);
165	if (RT_SUCCESS(rc))
166	{
167	memcpy(pvBuf, pvSrc, cb);
168	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
169	}
170	else
171	{
172	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternalReadOnly failed on %RGp / %R[pgmpage] -> %Rrc\n",
173	pRam->GCPhys + off, pPage, rc));
174	memset(pvBuf, 0xff, cb);
175	}
176
177	/* next page */
178	if (cb >= cbRead)
179	{
180	pgmUnlock(pVM);
181	return VINF_SUCCESS;
182	}
183	cbRead -= cb;
184	off += cb;
185	GCPhys += cb;
186	pvBuf = (char *)pvBuf + cb;
187	} /* walk pages in ram range. */
188	}
189	else
190	{
191	LogFlow(("PGMPhysRead: Unassigned %RGp size=%u\n", GCPhys, cbRead));
192
193	/*
194	* Unassigned address space.
195	*/
196	size_t cb = pRam ? pRam->GCPhys - GCPhys : ~(size_t)0;
197	if (cb >= cbRead)
198	{
199	memset(pvBuf, 0xff, cbRead);
200	break;
201	}
202	memset(pvBuf, 0xff, cb);
203
204	cbRead -= cb;
205	pvBuf = (char *)pvBuf + cb;
206	GCPhys += cb;
207	}
208
209	/* Advance range if necessary. */
210	while (pRam && GCPhys > pRam->GCPhysLast)
211	pRam = pRam->CTX_SUFF(pNext);
212	} /* Ram range walk */
213
214	pgmUnlock(pVM);
215
216	return VINF_SUCCESS;
217	}
218
219
220	/**
221	* EMT worker for PGMR3PhysWriteExternal.
222	*/
223	static DECLCALLBACK(int) pgmR3PhysWriteExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, const void *pvBuf, size_t cbWrite)
224	{
225	/** @todo VERR_EM_NO_MEMORY */
226	PGMPhysWrite(pVM, *pGCPhys, pvBuf, cbWrite);
227	return VINF_SUCCESS;
228	}
229
230
231	/**
232	* Write to physical memory, external users.
233	*
234	* @returns VBox status code.
235	* @retval VINF_SUCCESS.
236	* @retval VERR_EM_NO_MEMORY.
237	*
238	* @param pVM Pointer to the VM.
239	* @param GCPhys Physical address to write to.
240	* @param pvBuf What to write.
241	* @param cbWrite How many bytes to write.
242	* @param pszWho Who is writing. For tracking down who is writing
243	* after we've saved the state.
244	*
245	* @thread Any but EMTs.
246	*/
247	VMMDECL(int) PGMR3PhysWriteExternal(PVM pVM, RTGCPHYS GCPhys, const void pvBuf, size_t cbWrite, const char pszWho)
248	{
249	VM_ASSERT_OTHER_THREAD(pVM);
250
251	AssertMsg(!pVM->pgm.s.fNoMorePhysWrites,
252	("Calling PGMR3PhysWriteExternal after pgmR3Save()! GCPhys=%RGp cbWrite=%#x pszWho=%s\n",
253	GCPhys, cbWrite, pszWho));
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);
291	else
292	{
293	pgmUnlock(pVM);
294
295	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysWriteExternalEMT, 4,
296	pVM, &GCPhys, pvBuf, cbWrite);
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 Pointer to the VM.
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 Pointer to the VM.
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)
474	#endif
475	)
476	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage);
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 Pointer to the VM.
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	#define MAKE_LEAF(a_pNode) \
598	do { \
599	(a_pNode)->pLeftR3 = NIL_RTR3PTR; \
600	(a_pNode)->pRightR3 = NIL_RTR3PTR; \
601	(a_pNode)->pLeftR0 = NIL_RTR0PTR; \
602	(a_pNode)->pRightR0 = NIL_RTR0PTR; \
603	(a_pNode)->pLeftRC = NIL_RTRCPTR; \
604	(a_pNode)->pRightRC = NIL_RTRCPTR; \
605	} while (0)
606
607	#define INSERT_LEFT(a_pParent, a_pNode) \
608	do { \
609	(a_pParent)->pLeftR3 = (a_pNode); \
610	(a_pParent)->pLeftR0 = (a_pNode)->pSelfR0; \
611	(a_pParent)->pLeftRC = (a_pNode)->pSelfRC; \
612	} while (0)
613	#define INSERT_RIGHT(a_pParent, a_pNode) \
614	do { \
615	(a_pParent)->pRightR3 = (a_pNode); \
616	(a_pParent)->pRightR0 = (a_pNode)->pSelfR0; \
617	(a_pParent)->pRightRC = (a_pNode)->pSelfRC; \
618	} while (0)
619
620
621	/**
622	* Recursive tree builder.
623	*
624	* @param ppRam Pointer to the iterator variable.
625	* @param iHeight The hight about normal leaf nodes. Inserts a leaf
626	* node if 0.
627	*/
628	static PPGMRAMRANGE pgmR3PhysRebuildRamRangeSearchTreesRecursively(PPGMRAMRANGE *ppRam, int iDepth)
629	{
630	PPGMRAMRANGE pRam;
631	if (iDepth <= 0)
632	{
633	/*
634	* Leaf node.
635	*/
636	pRam = *ppRam;
637	if (pRam)
638	{
639	*ppRam = pRam->pNextR3;
640	MAKE_LEAF(pRam);
641	}
642	}
643	else
644	{
645
646	/*
647	* Intermediate node.
648	*/
649	PPGMRAMRANGE pLeft = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
650
651	pRam = *ppRam;
652	if (!pRam)
653	return pLeft;
654	*ppRam = pRam->pNextR3;
655	MAKE_LEAF(pRam);
656	INSERT_LEFT(pRam, pLeft);
657
658	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
659	if (pRight)
660	INSERT_RIGHT(pRam, pRight);
661	}
662	return pRam;
663	}
664
665
666	/**
667	* Rebuilds the RAM range search trees.
668	*
669	* @param pVM Pointer to the VM.
670	*/
671	static void pgmR3PhysRebuildRamRangeSearchTrees(PVM pVM)
672	{
673
674	/*
675	* Create the reasonably balanced tree in a sequential fashion.
676	* For simplicity (laziness) we use standard recursion here.
677	*/
678	int iDepth = 0;
679	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
680	PPGMRAMRANGE pRoot = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, 0);
681	while (pRam)
682	{
683	PPGMRAMRANGE pLeft = pRoot;
684
685	pRoot = pRam;
686	pRam = pRam->pNextR3;
687	MAKE_LEAF(pRoot);
688	INSERT_LEFT(pRoot, pLeft);
689
690	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, iDepth);
691	if (pRight)
692	INSERT_RIGHT(pRoot, pRight);
693	/** @todo else: rotate the tree. */
694
695	iDepth++;
696	}
697
698	pVM->pgm.s.pRamRangeTreeR3 = pRoot;
699	pVM->pgm.s.pRamRangeTreeR0 = pRoot ? pRoot->pSelfR0 : NIL_RTR0PTR;
700	pVM->pgm.s.pRamRangeTreeRC = pRoot ? pRoot->pSelfRC : NIL_RTRCPTR;
701
702	#ifdef VBOX_STRICT
703	/*
704	* Verify that the above code works.
705	*/
706	unsigned cRanges = 0;
707	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
708	cRanges++;
709	Assert(cRanges > 0);
710
711	unsigned cMaxDepth = ASMBitLastSetU32(cRanges);
712	if ((1U << cMaxDepth) < cRanges)
713	cMaxDepth++;
714
715	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
716	{
717	unsigned cDepth = 0;
718	PPGMRAMRANGE pRam2 = pVM->pgm.s.pRamRangeTreeR3;
719	for (;;)
720	{
721	if (pRam == pRam2)
722	break;
723	Assert(pRam2);
724	if (pRam->GCPhys < pRam2->GCPhys)
725	pRam2 = pRam2->pLeftR3;
726	else
727	pRam2 = pRam2->pRightR3;
728	}
729	AssertMsg(cDepth <= cMaxDepth, ("cDepth=%d cMaxDepth=%d\n", cDepth, cMaxDepth));
730	}
731	#endif /* VBOX_STRICT */
732	}
733
734	#undef MAKE_LEAF
735	#undef INSERT_LEFT
736	#undef INSERT_RIGHT
737
738	/**
739	* Relinks the RAM ranges using the pSelfRC and pSelfR0 pointers.
740	*
741	* Called when anything was relocated.
742	*
743	* @param pVM Pointer to the VM.
744	*/
745	void pgmR3PhysRelinkRamRanges(PVM pVM)
746	{
747	PPGMRAMRANGE pCur;
748
749	#ifdef VBOX_STRICT
750	for (pCur = pVM->pgm.s.pRamRangesXR3; pCur; pCur = pCur->pNextR3)
751	{
752	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfR0 == MMHyperCCToR0(pVM, pCur));
753	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfRC == MMHyperCCToRC(pVM, pCur));
754	Assert((pCur->GCPhys & PAGE_OFFSET_MASK) == 0);
755	Assert((pCur->GCPhysLast & PAGE_OFFSET_MASK) == PAGE_OFFSET_MASK);
756	Assert((pCur->cb & PAGE_OFFSET_MASK) == 0);
757	Assert(pCur->cb == pCur->GCPhysLast - pCur->GCPhys + 1);
758	for (PPGMRAMRANGE pCur2 = pVM->pgm.s.pRamRangesXR3; pCur2; pCur2 = pCur2->pNextR3)
759	Assert( pCur2 == pCur
760	\|\| strcmp(pCur2->pszDesc, pCur->pszDesc)); /** @todo fix MMIO ranges!! */
761	}
762	#endif
763
764	pCur = pVM->pgm.s.pRamRangesXR3;
765	if (pCur)
766	{
767	pVM->pgm.s.pRamRangesXR0 = pCur->pSelfR0;
768	pVM->pgm.s.pRamRangesXRC = pCur->pSelfRC;
769
770	for (; pCur->pNextR3; pCur = pCur->pNextR3)
771	{
772	pCur->pNextR0 = pCur->pNextR3->pSelfR0;
773	pCur->pNextRC = pCur->pNextR3->pSelfRC;
774	}
775
776	Assert(pCur->pNextR0 == NIL_RTR0PTR);
777	Assert(pCur->pNextRC == NIL_RTRCPTR);
778	}
779	else
780	{
781	Assert(pVM->pgm.s.pRamRangesXR0 == NIL_RTR0PTR);
782	Assert(pVM->pgm.s.pRamRangesXRC == NIL_RTRCPTR);
783	}
784	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
785
786	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
787	}
788
789
790	/**
791	* Links a new RAM range into the list.
792	*
793	* @param pVM Pointer to the VM.
794	* @param pNew Pointer to the new list entry.
795	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
796	*/
797	static void pgmR3PhysLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, PPGMRAMRANGE pPrev)
798	{
799	AssertMsg(pNew->pszDesc, ("%RGp-%RGp\n", pNew->GCPhys, pNew->GCPhysLast));
800	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfR0 == MMHyperCCToR0(pVM, pNew));
801	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfRC == MMHyperCCToRC(pVM, pNew));
802
803	pgmLock(pVM);
804
805	PPGMRAMRANGE pRam = pPrev ? pPrev->pNextR3 : pVM->pgm.s.pRamRangesXR3;
806	pNew->pNextR3 = pRam;
807	pNew->pNextR0 = pRam ? pRam->pSelfR0 : NIL_RTR0PTR;
808	pNew->pNextRC = pRam ? pRam->pSelfRC : NIL_RTRCPTR;
809
810	if (pPrev)
811	{
812	pPrev->pNextR3 = pNew;
813	pPrev->pNextR0 = pNew->pSelfR0;
814	pPrev->pNextRC = pNew->pSelfRC;
815	}
816	else
817	{
818	pVM->pgm.s.pRamRangesXR3 = pNew;
819	pVM->pgm.s.pRamRangesXR0 = pNew->pSelfR0;
820	pVM->pgm.s.pRamRangesXRC = pNew->pSelfRC;
821	}
822	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
823
824	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
825	pgmUnlock(pVM);
826	}
827
828
829	/**
830	* Unlink an existing RAM range from the list.
831	*
832	* @param pVM Pointer to the VM.
833	* @param pRam Pointer to the new list entry.
834	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
835	*/
836	static void pgmR3PhysUnlinkRamRange2(PVM pVM, PPGMRAMRANGE pRam, PPGMRAMRANGE pPrev)
837	{
838	Assert(pPrev ? pPrev->pNextR3 == pRam : pVM->pgm.s.pRamRangesXR3 == pRam);
839	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfR0 == MMHyperCCToR0(pVM, pRam));
840	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfRC == MMHyperCCToRC(pVM, pRam));
841
842	pgmLock(pVM);
843
844	PPGMRAMRANGE pNext = pRam->pNextR3;
845	if (pPrev)
846	{
847	pPrev->pNextR3 = pNext;
848	pPrev->pNextR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
849	pPrev->pNextRC = pNext ? pNext->pSelfRC : NIL_RTRCPTR;
850	}
851	else
852	{
853	Assert(pVM->pgm.s.pRamRangesXR3 == pRam);
854	pVM->pgm.s.pRamRangesXR3 = pNext;
855	pVM->pgm.s.pRamRangesXR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
856	pVM->pgm.s.pRamRangesXRC = pNext ? pNext->pSelfRC : NIL_RTRCPTR;
857	}
858	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
859
860	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
861	pgmUnlock(pVM);
862	}
863
864
865	/**
866	* Unlink an existing RAM range from the list.
867	*
868	* @param pVM Pointer to the VM.
869	* @param pRam Pointer to the new list entry.
870	*/
871	static void pgmR3PhysUnlinkRamRange(PVM pVM, PPGMRAMRANGE pRam)
872	{
873	pgmLock(pVM);
874
875	/* find prev. */
876	PPGMRAMRANGE pPrev = NULL;
877	PPGMRAMRANGE pCur = pVM->pgm.s.pRamRangesXR3;
878	while (pCur != pRam)
879	{
880	pPrev = pCur;
881	pCur = pCur->pNextR3;
882	}
883	AssertFatal(pCur);
884
885	pgmR3PhysUnlinkRamRange2(pVM, pRam, pPrev);
886	pgmUnlock(pVM);
887	}
888
889
890	/**
891	* Frees a range of pages, replacing them with ZERO pages of the specified type.
892	*
893	* @returns VBox status code.
894	* @param pVM Pointer to the VM.
895	* @param pRam The RAM range in which the pages resides.
896	* @param GCPhys The address of the first page.
897	* @param GCPhysLast The address of the last page.
898	* @param uType The page type to replace then with.
899	*/
900	static int pgmR3PhysFreePageRange(PVM pVM, PPGMRAMRANGE pRam, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast, uint8_t uType)
901	{
902	PGM_LOCK_ASSERT_OWNER(pVM);
903	uint32_t cPendingPages = 0;
904	PGMMFREEPAGESREQ pReq;
905	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
906	AssertLogRelRCReturn(rc, rc);
907
908	/* Iterate the pages. */
909	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
910	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
911	while (cPagesLeft-- > 0)
912	{
913	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys);
914	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
915
916	PGM_PAGE_SET_TYPE(pVM, pPageDst, uType);
917
918	GCPhys += PAGE_SIZE;
919	pPageDst++;
920	}
921
922	if (cPendingPages)
923	{
924	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
925	AssertLogRelRCReturn(rc, rc);
926	}
927	GMMR3FreePagesCleanup(pReq);
928
929	return rc;
930	}
931
932	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
933
934	/**
935	* Rendezvous callback used by PGMR3ChangeMemBalloon that changes the memory balloon size
936	*
937	* This is only called on one of the EMTs while the other ones are waiting for
938	* it to complete this function.
939	*
940	* @returns VINF_SUCCESS (VBox strict status code).
941	* @param pVM Pointer to the VM.
942	* @param pVCpu The VMCPU for the EMT we're being called on. Unused.
943	* @param pvUser User parameter
944	*/
945	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysChangeMemBalloonRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
946	{
947	uintptr_t paUser = (uintptr_t )pvUser;
948	bool fInflate = !!paUser[0];
949	unsigned cPages = paUser[1];
950	RTGCPHYS paPhysPage = (RTGCPHYS )paUser[2];
951	uint32_t cPendingPages = 0;
952	PGMMFREEPAGESREQ pReq;
953	int rc;
954
955	Log(("pgmR3PhysChangeMemBalloonRendezvous: %s %x pages\n", (fInflate) ? "inflate" : "deflate", cPages));
956	pgmLock(pVM);
957
958	if (fInflate)
959	{
960	/* Flush the PGM pool cache as we might have stale references to pages that we just freed. */
961	pgmR3PoolClearAllRendezvous(pVM, pVCpu, NULL);
962
963	/* Replace pages with ZERO pages. */
964	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
965	if (RT_FAILURE(rc))
966	{
967	pgmUnlock(pVM);
968	AssertLogRelRC(rc);
969	return rc;
970	}
971
972	/* Iterate the pages. */
973	for (unsigned i = 0; i < cPages; i++)
974	{
975	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
976	if ( pPage == NULL
977	\|\| PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM)
978	{
979	Log(("pgmR3PhysChangeMemBalloonRendezvous: invalid physical page %RGp pPage->u3Type=%d\n", paPhysPage[i], pPage ? PGM_PAGE_GET_TYPE(pPage) : 0));
980	break;
981	}
982
983	LogFlow(("balloon page: %RGp\n", paPhysPage[i]));
984
985	/* Flush the shadow PT if this page was previously used as a guest page table. */
986	pgmPoolFlushPageByGCPhys(pVM, paPhysPage[i]);
987
988	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, paPhysPage[i]);
989	if (RT_FAILURE(rc))
990	{
991	pgmUnlock(pVM);
992	AssertLogRelRC(rc);
993	return rc;
994	}
995	Assert(PGM_PAGE_IS_ZERO(pPage));
996	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_BALLOONED);
997	}
998
999	if (cPendingPages)
1000	{
1001	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1002	if (RT_FAILURE(rc))
1003	{
1004	pgmUnlock(pVM);
1005	AssertLogRelRC(rc);
1006	return rc;
1007	}
1008	}
1009	GMMR3FreePagesCleanup(pReq);
1010	}
1011	else
1012	{
1013	/* Iterate the pages. */
1014	for (unsigned i = 0; i < cPages; i++)
1015	{
1016	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
1017	AssertBreak(pPage && PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM);
1018
1019	LogFlow(("Free ballooned page: %RGp\n", paPhysPage[i]));
1020
1021	Assert(PGM_PAGE_IS_BALLOONED(pPage));
1022
1023	/* Change back to zero page. */
1024	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
1025	}
1026
1027	/* Note that we currently do not map any ballooned pages in our shadow page tables, so no need to flush the pgm pool. */
1028	}
1029
1030	/* Notify GMM about the balloon change. */
1031	rc = GMMR3BalloonedPages(pVM, (fInflate) ? GMMBALLOONACTION_INFLATE : GMMBALLOONACTION_DEFLATE, cPages);
1032	if (RT_SUCCESS(rc))
1033	{
1034	if (!fInflate)
1035	{
1036	Assert(pVM->pgm.s.cBalloonedPages >= cPages);
1037	pVM->pgm.s.cBalloonedPages -= cPages;
1038	}
1039	else
1040	pVM->pgm.s.cBalloonedPages += cPages;
1041	}
1042
1043	pgmUnlock(pVM);
1044
1045	/* Flush the recompiler's TLB as well. */
1046	for (VMCPUID i = 0; i < pVM->cCpus; i++)
1047	CPUMSetChangedFlags(&pVM->aCpus[i], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1048
1049	AssertLogRelRC(rc);
1050	return rc;
1051	}
1052
1053
1054	/**
1055	* Frees a range of ram pages, replacing them with ZERO pages; helper for PGMR3PhysFreeRamPages
1056	*
1057	* @returns VBox status code.
1058	* @param pVM Pointer to the VM.
1059	* @param fInflate Inflate or deflate memory balloon
1060	* @param cPages Number of pages to free
1061	* @param paPhysPage Array of guest physical addresses
1062	*/
1063	static DECLCALLBACK(void) pgmR3PhysChangeMemBalloonHelper(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1064	{
1065	uintptr_t paUser[3];
1066
1067	paUser[0] = fInflate;
1068	paUser[1] = cPages;
1069	paUser[2] = (uintptr_t)paPhysPage;
1070	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1071	AssertRC(rc);
1072
1073	/* Made a copy in PGMR3PhysFreeRamPages; free it here. */
1074	RTMemFree(paPhysPage);
1075	}
1076
1077	#endif /* 64-bit host && (Windows \|\| Solaris \|\| Linux \|\| FreeBSD) */
1078
1079	/**
1080	* Inflate or deflate a memory balloon
1081	*
1082	* @returns VBox status code.
1083	* @param pVM Pointer to the VM.
1084	* @param fInflate Inflate or deflate memory balloon
1085	* @param cPages Number of pages to free
1086	* @param paPhysPage Array of guest physical addresses
1087	*/
1088	VMMR3DECL(int) PGMR3PhysChangeMemBalloon(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1089	{
1090	/* This must match GMMR0Init; currently we only support memory ballooning on all 64-bit hosts except Mac OS X */
1091	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
1092	int rc;
1093
1094	/* Older additions (ancient non-functioning balloon code) pass wrong physical addresses. */
1095	AssertReturn(!(paPhysPage[0] & 0xfff), VERR_INVALID_PARAMETER);
1096
1097	/* We own the IOM lock here and could cause a deadlock by waiting for another VCPU that is blocking on the IOM lock.
1098	* In the SMP case we post a request packet to postpone the job.
1099	*/
1100	if (pVM->cCpus > 1)
1101	{
1102	unsigned cbPhysPage = cPages * sizeof(paPhysPage[0]);
1103	RTGCPHYS paPhysPageCopy = (RTGCPHYS )RTMemAlloc(cbPhysPage);
1104	AssertReturn(paPhysPageCopy, VERR_NO_MEMORY);
1105
1106	memcpy(paPhysPageCopy, paPhysPage, cbPhysPage);
1107
1108	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysChangeMemBalloonHelper, 4, pVM, fInflate, cPages, paPhysPageCopy);
1109	AssertRC(rc);
1110	}
1111	else
1112	{
1113	uintptr_t paUser[3];
1114
1115	paUser[0] = fInflate;
1116	paUser[1] = cPages;
1117	paUser[2] = (uintptr_t)paPhysPage;
1118	rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1119	AssertRC(rc);
1120	}
1121	return rc;
1122
1123	#else
1124	NOREF(pVM); NOREF(fInflate); NOREF(cPages); NOREF(paPhysPage);
1125	return VERR_NOT_IMPLEMENTED;
1126	#endif
1127	}
1128
1129
1130	/**
1131	* Rendezvous callback used by PGMR3WriteProtectRAM that write protects all
1132	* physical RAM.
1133	*
1134	* This is only called on one of the EMTs while the other ones are waiting for
1135	* it to complete this function.
1136	*
1137	* @returns VINF_SUCCESS (VBox strict status code).
1138	* @param pVM Pointer to the VM.
1139	* @param pVCpu The VMCPU for the EMT we're being called on. Unused.
1140	* @param pvUser User parameter, unused.
1141	*/
1142	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysWriteProtectRAMRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
1143	{
1144	int rc = VINF_SUCCESS;
1145	NOREF(pvUser); NOREF(pVCpu);
1146
1147	pgmLock(pVM);
1148	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
1149	pgmPoolResetDirtyPages(pVM);
1150	#endif
1151
1152	/** @todo pointless to write protect the physical page pointed to by RSP. */
1153
1154	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
1155	pRam;
1156	pRam = pRam->CTX_SUFF(pNext))
1157	{
1158	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
1159	for (uint32_t iPage = 0; iPage < cPages; iPage++)
1160	{
1161	PPGMPAGE pPage = &pRam->aPages[iPage];
1162	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
1163
1164	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
1165	\|\| enmPageType == PGMPAGETYPE_MMIO2)
1166	{
1167	/*
1168	* A RAM page.
1169	*/
1170	switch (PGM_PAGE_GET_STATE(pPage))
1171	{
1172	case PGM_PAGE_STATE_ALLOCATED:
1173	/** @todo Optimize this: Don't always re-enable write
1174	* monitoring if the page is known to be very busy. */
1175	if (PGM_PAGE_IS_WRITTEN_TO(pPage))
1176	{
1177	PGM_PAGE_CLEAR_WRITTEN_TO(pVM, pPage);
1178	/* Remember this dirty page for the next (memory) sync. */
1179	PGM_PAGE_SET_FT_DIRTY(pPage);
1180	}
1181
1182	pgmPhysPageWriteMonitor(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
1183	break;
1184
1185	case PGM_PAGE_STATE_SHARED:
1186	AssertFailed();
1187	break;
1188
1189	case PGM_PAGE_STATE_WRITE_MONITORED: /* nothing to change. */
1190	default:
1191	break;
1192	}
1193	}
1194	}
1195	}
1196	pgmR3PoolWriteProtectPages(pVM);
1197	PGM_INVL_ALL_VCPU_TLBS(pVM);
1198	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
1199	CPUMSetChangedFlags(&pVM->aCpus[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1200
1201	pgmUnlock(pVM);
1202	return rc;
1203	}
1204
1205	/**
1206	* Protect all physical RAM to monitor writes
1207	*
1208	* @returns VBox status code.
1209	* @param pVM Pointer to the VM.
1210	*/
1211	VMMR3DECL(int) PGMR3PhysWriteProtectRAM(PVM pVM)
1212	{
1213	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1214
1215	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysWriteProtectRAMRendezvous, NULL);
1216	AssertRC(rc);
1217	return rc;
1218	}
1219
1220	/**
1221	* Enumerate all dirty FT pages.
1222	*
1223	* @returns VBox status code.
1224	* @param pVM Pointer to the VM.
1225	* @param pfnEnum Enumerate callback handler.
1226	* @param pvUser Enumerate callback handler parameter.
1227	*/
1228	VMMR3DECL(int) PGMR3PhysEnumDirtyFTPages(PVM pVM, PFNPGMENUMDIRTYFTPAGES pfnEnum, void *pvUser)
1229	{
1230	int rc = VINF_SUCCESS;
1231
1232	pgmLock(pVM);
1233	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
1234	pRam;
1235	pRam = pRam->CTX_SUFF(pNext))
1236	{
1237	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
1238	for (uint32_t iPage = 0; iPage < cPages; iPage++)
1239	{
1240	PPGMPAGE pPage = &pRam->aPages[iPage];
1241	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
1242
1243	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
1244	\|\| enmPageType == PGMPAGETYPE_MMIO2)
1245	{
1246	/*
1247	* A RAM page.
1248	*/
1249	switch (PGM_PAGE_GET_STATE(pPage))
1250	{
1251	case PGM_PAGE_STATE_ALLOCATED:
1252	case PGM_PAGE_STATE_WRITE_MONITORED:
1253	if ( !PGM_PAGE_IS_WRITTEN_TO(pPage) /* not very recently updated? */
1254	&& PGM_PAGE_IS_FT_DIRTY(pPage))
1255	{
1256	unsigned cbPageRange = PAGE_SIZE;
1257	unsigned iPageClean = iPage + 1;
1258	RTGCPHYS GCPhysPage = pRam->GCPhys + iPage * PAGE_SIZE;
1259	uint8_t *pu8Page = NULL;
1260	PGMPAGEMAPLOCK Lock;
1261
1262	/* Find the next clean page, so we can merge adjacent dirty pages. */
1263	for (; iPageClean < cPages; iPageClean++)
1264	{
1265	PPGMPAGE pPageNext = &pRam->aPages[iPageClean];
1266	if ( RT_UNLIKELY(PGM_PAGE_GET_TYPE(pPageNext) != PGMPAGETYPE_RAM)
1267	\|\| PGM_PAGE_GET_STATE(pPageNext) != PGM_PAGE_STATE_ALLOCATED
1268	\|\| PGM_PAGE_IS_WRITTEN_TO(pPageNext)
1269	\|\| !PGM_PAGE_IS_FT_DIRTY(pPageNext)
1270	/* Crossing a chunk boundary? */
1271	\|\| (GCPhysPage & GMM_PAGEID_IDX_MASK) != ((GCPhysPage + cbPageRange) & GMM_PAGEID_IDX_MASK)
1272	)
1273	break;
1274
1275	cbPageRange += PAGE_SIZE;
1276	}
1277
1278	rc = PGMPhysGCPhys2CCPtrReadOnly(pVM, GCPhysPage, (const void **)&pu8Page, &Lock);
1279	if (RT_SUCCESS(rc))
1280	{
1281	/** @todo this is risky; the range might be changed, but little choice as the sync
1282	* costs a lot of time. */
1283	pgmUnlock(pVM);
1284	pfnEnum(pVM, GCPhysPage, pu8Page, cbPageRange, pvUser);
1285	pgmLock(pVM);
1286	PGMPhysReleasePageMappingLock(pVM, &Lock);
1287	}
1288
1289	for (iPage; iPage < iPageClean; iPage++)
1290	PGM_PAGE_CLEAR_FT_DIRTY(&pRam->aPages[iPage]);
1291
1292	iPage = iPageClean - 1;
1293	}
1294	break;
1295	}
1296	}
1297	}
1298	}
1299	pgmUnlock(pVM);
1300	return rc;
1301	}
1302
1303
1304	/**
1305	* Gets the number of ram ranges.
1306	*
1307	* @returns Number of ram ranges. Returns UINT32_MAX if @a pVM is invalid.
1308	* @param pVM Pointer to the VM.
1309	*/
1310	VMMR3DECL(uint32_t) PGMR3PhysGetRamRangeCount(PVM pVM)
1311	{
1312	VM_ASSERT_VALID_EXT_RETURN(pVM, UINT32_MAX);
1313
1314	pgmLock(pVM);
1315	uint32_t cRamRanges = 0;
1316	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext))
1317	cRamRanges++;
1318	pgmUnlock(pVM);
1319	return cRamRanges;
1320	}
1321
1322
1323	/**
1324	* Get information about a range.
1325	*
1326	* @returns VINF_SUCCESS or VERR_OUT_OF_RANGE.
1327	* @param pVM Pointer to the VM.
1328	* @param iRange The ordinal of the range.
1329	* @param pGCPhysStart Where to return the start of the range. Optional.
1330	* @param pGCPhysLast Where to return the address of the last byte in the
1331	* range. Optional.
1332	* @param pfIsMmio Where to indicate that this is a pure MMIO range.
1333	* Optional.
1334	*/
1335	VMMR3DECL(int) PGMR3PhysGetRange(PVM pVM, uint32_t iRange, PRTGCPHYS pGCPhysStart, PRTGCPHYS pGCPhysLast,
1336	const char *ppszDesc, bool pfIsMmio)
1337	{
1338	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1339
1340	pgmLock(pVM);
1341	uint32_t iCurRange = 0;
1342	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext), iCurRange++)
1343	if (iCurRange == iRange)
1344	{
1345	if (pGCPhysStart)
1346	*pGCPhysStart = pCur->GCPhys;
1347	if (pGCPhysLast)
1348	*pGCPhysLast = pCur->GCPhysLast;
1349	if (ppszDesc)
1350	*ppszDesc = pCur->pszDesc;
1351	if (pfIsMmio)
1352	*pfIsMmio = !!(pCur->fFlags & PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO);
1353
1354	pgmUnlock(pVM);
1355	return VINF_SUCCESS;
1356	}
1357	pgmUnlock(pVM);
1358	return VERR_OUT_OF_RANGE;
1359	}
1360
1361
1362	/**
1363	* Query the amount of free memory inside VMMR0
1364	*
1365	* @returns VBox status code.
1366	* @param pUVM The user mode VM handle.
1367	* @param pcbAllocMem Where to return the amount of memory allocated
1368	* by VMs.
1369	* @param pcbFreeMem Where to return the amount of memory that is
1370	* allocated from the host but not currently used
1371	* by any VMs.
1372	* @param pcbBallonedMem Where to return the sum of memory that is
1373	* currently ballooned by the VMs.
1374	* @param pcbSharedMem Where to return the amount of memory that is
1375	* currently shared.
1376	*/
1377	VMMR3DECL(int) PGMR3QueryGlobalMemoryStats(PUVM pUVM, uint64_t pcbAllocMem, uint64_t pcbFreeMem,
1378	uint64_t pcbBallonedMem, uint64_t pcbSharedMem)
1379	{
1380	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1381	VM_ASSERT_VALID_EXT_RETURN(pUVM->pVM, VERR_INVALID_VM_HANDLE);
1382
1383	uint64_t cAllocPages = 0;
1384	uint64_t cFreePages = 0;
1385	uint64_t cBalloonPages = 0;
1386	uint64_t cSharedPages = 0;
1387	int rc = GMMR3QueryHypervisorMemoryStats(pUVM->pVM, &cAllocPages, &cFreePages, &cBalloonPages, &cSharedPages);
1388	AssertRCReturn(rc, rc);
1389
1390	if (pcbAllocMem)
1391	pcbAllocMem = cAllocPages _4K;
1392
1393	if (pcbFreeMem)
1394	pcbFreeMem = cFreePages _4K;
1395
1396	if (pcbBallonedMem)
1397	pcbBallonedMem = cBalloonPages _4K;
1398
1399	if (pcbSharedMem)
1400	pcbSharedMem = cSharedPages _4K;
1401
1402	Log(("PGMR3QueryVMMMemoryStats: all=%llx free=%llx ballooned=%llx shared=%llx\n",
1403	cAllocPages, cFreePages, cBalloonPages, cSharedPages));
1404	return VINF_SUCCESS;
1405	}
1406
1407
1408	/**
1409	* Query memory stats for the VM.
1410	*
1411	* @returns VBox status code.
1412	* @param pUVM The user mode VM handle.
1413	* @param pcbTotalMem Where to return total amount memory the VM may
1414	* possibly use.
1415	* @param pcbPrivateMem Where to return the amount of private memory
1416	* currently allocated.
1417	* @param pcbSharedMem Where to return the amount of actually shared
1418	* memory currently used by the VM.
1419	* @param pcbZeroMem Where to return the amount of memory backed by
1420	* zero pages.
1421	*
1422	* @remarks The total mem is normally larger than the sum of the three
1423	* components. There are two reasons for this, first the amount of
1424	* shared memory is what we're sure is shared instead of what could
1425	* possibly be shared with someone. Secondly, because the total may
1426	* include some pure MMIO pages that doesn't go into any of the three
1427	* sub-counts.
1428	*
1429	* @todo Why do we return reused shared pages instead of anything that could
1430	* potentially be shared? Doesn't this mean the first VM gets a much
1431	* lower number of shared pages?
1432	*/
1433	VMMR3DECL(int) PGMR3QueryMemoryStats(PUVM pUVM, uint64_t pcbTotalMem, uint64_t pcbPrivateMem,
1434	uint64_t pcbSharedMem, uint64_t pcbZeroMem)
1435	{
1436	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1437	PVM pVM = pUVM->pVM;
1438	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1439
1440	if (pcbTotalMem)
1441	pcbTotalMem = (uint64_t)pVM->pgm.s.cAllPages PAGE_SIZE;
1442
1443	if (pcbPrivateMem)
1444	pcbPrivateMem = (uint64_t)pVM->pgm.s.cPrivatePages PAGE_SIZE;
1445
1446	if (pcbSharedMem)
1447	pcbSharedMem = (uint64_t)pVM->pgm.s.cReusedSharedPages PAGE_SIZE;
1448
1449	if (pcbZeroMem)
1450	pcbZeroMem = (uint64_t)pVM->pgm.s.cZeroPages PAGE_SIZE;
1451
1452	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));
1453	return VINF_SUCCESS;
1454	}
1455
1456
1457	/**
1458	* PGMR3PhysRegisterRam worker that initializes and links a RAM range.
1459	*
1460	* @param pVM Pointer to the VM.
1461	* @param pNew The new RAM range.
1462	* @param GCPhys The address of the RAM range.
1463	* @param GCPhysLast The last address of the RAM range.
1464	* @param RCPtrNew The RC address if the range is floating. NIL_RTRCPTR
1465	* if in HMA.
1466	* @param R0PtrNew Ditto for R0.
1467	* @param pszDesc The description.
1468	* @param pPrev The previous RAM range (for linking).
1469	*/
1470	static void pgmR3PhysInitAndLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast,
1471	RTRCPTR RCPtrNew, RTR0PTR R0PtrNew, const char *pszDesc, PPGMRAMRANGE pPrev)
1472	{
1473	/*
1474	* Initialize the range.
1475	*/
1476	pNew->pSelfR0 = R0PtrNew != NIL_RTR0PTR ? R0PtrNew : MMHyperCCToR0(pVM, pNew);
1477	pNew->pSelfRC = RCPtrNew != NIL_RTRCPTR ? RCPtrNew : MMHyperCCToRC(pVM, pNew);
1478	pNew->GCPhys = GCPhys;
1479	pNew->GCPhysLast = GCPhysLast;
1480	pNew->cb = GCPhysLast - GCPhys + 1;
1481	pNew->pszDesc = pszDesc;
1482	pNew->fFlags = RCPtrNew != NIL_RTRCPTR ? PGM_RAM_RANGE_FLAGS_FLOATING : 0;
1483	pNew->pvR3 = NULL;
1484	pNew->paLSPages = NULL;
1485
1486	uint32_t const cPages = pNew->cb >> PAGE_SHIFT;
1487	RTGCPHYS iPage = cPages;
1488	while (iPage-- > 0)
1489	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_RAM);
1490
1491	/* Update the page count stats. */
1492	pVM->pgm.s.cZeroPages += cPages;
1493	pVM->pgm.s.cAllPages += cPages;
1494
1495	/*
1496	* Link it.
1497	*/
1498	pgmR3PhysLinkRamRange(pVM, pNew, pPrev);
1499	}
1500
1501
1502	/**
1503	* Relocate a floating RAM range.
1504	*
1505	* @copydoc FNPGMRELOCATE.
1506	*/
1507	static DECLCALLBACK(bool) pgmR3PhysRamRangeRelocate(PVM pVM, RTGCPTR GCPtrOld, RTGCPTR GCPtrNew, PGMRELOCATECALL enmMode, void *pvUser)
1508	{
1509	PPGMRAMRANGE pRam = (PPGMRAMRANGE)pvUser;
1510	Assert(pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING);
1511	Assert(pRam->pSelfRC == GCPtrOld + PAGE_SIZE);
1512
1513	switch (enmMode)
1514	{
1515	case PGMRELOCATECALL_SUGGEST:
1516	return true;
1517
1518	case PGMRELOCATECALL_RELOCATE:
1519	{
1520	/*
1521	* Update myself, then relink all the ranges and flush the RC TLB.
1522	*/
1523	pgmLock(pVM);
1524
1525	pRam->pSelfRC = (RTRCPTR)(GCPtrNew + PAGE_SIZE);
1526
1527	pgmR3PhysRelinkRamRanges(pVM);
1528	for (unsigned i = 0; i < PGM_RAMRANGE_TLB_ENTRIES; i++)
1529	pVM->pgm.s.apRamRangesTlbRC[i] = NIL_RTRCPTR;
1530
1531	pgmUnlock(pVM);
1532	return true;
1533	}
1534
1535	default:
1536	AssertFailedReturn(false);
1537	}
1538	}
1539
1540
1541	/**
1542	* PGMR3PhysRegisterRam worker that registers a high chunk.
1543	*
1544	* @returns VBox status code.
1545	* @param pVM Pointer to the VM.
1546	* @param GCPhys The address of the RAM.
1547	* @param cRamPages The number of RAM pages to register.
1548	* @param cbChunk The size of the PGMRAMRANGE guest mapping.
1549	* @param iChunk The chunk number.
1550	* @param pszDesc The RAM range description.
1551	* @param ppPrev Previous RAM range pointer. In/Out.
1552	*/
1553	static int pgmR3PhysRegisterHighRamChunk(PVM pVM, RTGCPHYS GCPhys, uint32_t cRamPages,
1554	uint32_t cbChunk, uint32_t iChunk, const char *pszDesc,
1555	PPGMRAMRANGE *ppPrev)
1556	{
1557	const char *pszDescChunk = iChunk == 0
1558	? pszDesc
1559	: MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s (#%u)", pszDesc, iChunk + 1);
1560	AssertReturn(pszDescChunk, VERR_NO_MEMORY);
1561
1562	/*
1563	* Allocate memory for the new chunk.
1564	*/
1565	size_t const cChunkPages = RT_ALIGN_Z(RT_UOFFSETOF(PGMRAMRANGE, aPages[cRamPages]), PAGE_SIZE) >> PAGE_SHIFT;
1566	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
1567	AssertReturn(paChunkPages, VERR_NO_TMP_MEMORY);
1568	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
1569	void *pvChunk = NULL;
1570	int rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk,
1571	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
1572	HMIsEnabled(pVM) ? &R0PtrChunk : NULL,
1573	#else
1574	NULL,
1575	#endif
1576	paChunkPages);
1577	if (RT_SUCCESS(rc))
1578	{
1579	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
1580	if (!HMIsEnabled(pVM))
1581	R0PtrChunk = NIL_RTR0PTR;
1582	#else
1583	R0PtrChunk = (uintptr_t)pvChunk;
1584	#endif
1585	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
1586
1587	PPGMRAMRANGE pNew = (PPGMRAMRANGE)pvChunk;
1588
1589	/*
1590	* Create a mapping and map the pages into it.
1591	* We push these in below the HMA.
1592	*/
1593	RTGCPTR GCPtrChunkMap = pVM->pgm.s.GCPtrPrevRamRangeMapping - cbChunk;
1594	rc = PGMR3MapPT(pVM, GCPtrChunkMap, cbChunk, 0 /fFlags/, pgmR3PhysRamRangeRelocate, pNew, pszDescChunk);
1595	if (RT_SUCCESS(rc))
1596	{
1597	pVM->pgm.s.GCPtrPrevRamRangeMapping = GCPtrChunkMap;
1598
1599	RTGCPTR const GCPtrChunk = GCPtrChunkMap + PAGE_SIZE;
1600	RTGCPTR GCPtrPage = GCPtrChunk;
1601	for (uint32_t iPage = 0; iPage < cChunkPages && RT_SUCCESS(rc); iPage++, GCPtrPage += PAGE_SIZE)
1602	rc = PGMMap(pVM, GCPtrPage, paChunkPages[iPage].Phys, PAGE_SIZE, 0);
1603	if (RT_SUCCESS(rc))
1604	{
1605	/*
1606	* Ok, init and link the range.
1607	*/
1608	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhys + ((RTGCPHYS)cRamPages << PAGE_SHIFT) - 1,
1609	(RTRCPTR)GCPtrChunk, R0PtrChunk, pszDescChunk, *ppPrev);
1610	*ppPrev = pNew;
1611	}
1612	}
1613
1614	if (RT_FAILURE(rc))
1615	SUPR3PageFreeEx(pvChunk, cChunkPages);
1616	}
1617
1618	RTMemTmpFree(paChunkPages);
1619	return rc;
1620	}
1621
1622
1623	/**
1624	* Sets up a range RAM.
1625	*
1626	* This will check for conflicting registrations, make a resource
1627	* reservation for the memory (with GMM), and setup the per-page
1628	* tracking structures (PGMPAGE).
1629	*
1630	* @returns VBox status code.
1631	* @param pVM Pointer to the VM.
1632	* @param GCPhys The physical address of the RAM.
1633	* @param cb The size of the RAM.
1634	* @param pszDesc The description - not copied, so, don't free or change it.
1635	*/
1636	VMMR3DECL(int) PGMR3PhysRegisterRam(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, const char *pszDesc)
1637	{
1638	/*
1639	* Validate input.
1640	*/
1641	Log(("PGMR3PhysRegisterRam: GCPhys=%RGp cb=%RGp pszDesc=%s\n", GCPhys, cb, pszDesc));
1642	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
1643	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
1644	AssertReturn(cb > 0, VERR_INVALID_PARAMETER);
1645	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
1646	AssertMsgReturn(GCPhysLast > GCPhys, ("The range wraps! GCPhys=%RGp cb=%RGp\n", GCPhys, cb), VERR_INVALID_PARAMETER);
1647	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
1648	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1649
1650	pgmLock(pVM);
1651
1652	/*
1653	* Find range location and check for conflicts.
1654	* (We don't lock here because the locking by EMT is only required on update.)
1655	*/
1656	PPGMRAMRANGE pPrev = NULL;
1657	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
1658	while (pRam && GCPhysLast >= pRam->GCPhys)
1659	{
1660	if ( GCPhysLast >= pRam->GCPhys
1661	&& GCPhys <= pRam->GCPhysLast)
1662	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
1663	GCPhys, GCPhysLast, pszDesc,
1664	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
1665	VERR_PGM_RAM_CONFLICT);
1666
1667	/* next */
1668	pPrev = pRam;
1669	pRam = pRam->pNextR3;
1670	}
1671
1672	/*
1673	* Register it with GMM (the API bitches).
1674	*/
1675	const RTGCPHYS cPages = cb >> PAGE_SHIFT;
1676	int rc = MMR3IncreaseBaseReservation(pVM, cPages);
1677	if (RT_FAILURE(rc))
1678	{
1679	pgmUnlock(pVM);
1680	return rc;
1681	}
1682
1683	if ( GCPhys >= _4G
1684	&& cPages > 256)
1685	{
1686	/*
1687	* The PGMRAMRANGE structures for the high memory can get very big.
1688	* In order to avoid SUPR3PageAllocEx allocation failures due to the
1689	* allocation size limit there and also to avoid being unable to find
1690	* guest mapping space for them, we split this memory up into 4MB in
1691	* (potential) raw-mode configs and 16MB chunks in forced AMD-V/VT-x
1692	* mode.
1693	*
1694	* The first and last page of each mapping are guard pages and marked
1695	* not-present. So, we've got 4186112 and 16769024 bytes available for
1696	* the PGMRAMRANGE structure.
1697	*
1698	* Note! The sizes used here will influence the saved state.
1699	*/
1700	uint32_t cbChunk;
1701	uint32_t cPagesPerChunk;
1702	if (HMIsEnabled(pVM))
1703	{
1704	cbChunk = 16U*_1M;
1705	cPagesPerChunk = 1048048; /* max ~1048059 */
1706	AssertCompile(sizeof(PGMRAMRANGE) + sizeof(PGMPAGE) * 1048048 < 16U_1M - PAGE_SIZE 2);
1707	}
1708	else
1709	{
1710	cbChunk = 4U*_1M;
1711	cPagesPerChunk = 261616; /* max ~261627 */
1712	AssertCompile(sizeof(PGMRAMRANGE) + sizeof(PGMPAGE) * 261616 < 4U_1M - PAGE_SIZE 2);
1713	}
1714	AssertRelease(RT_UOFFSETOF(PGMRAMRANGE, aPages[cPagesPerChunk]) + PAGE_SIZE * 2 <= cbChunk);
1715
1716	RTGCPHYS cPagesLeft = cPages;
1717	RTGCPHYS GCPhysChunk = GCPhys;
1718	uint32_t iChunk = 0;
1719	while (cPagesLeft > 0)
1720	{
1721	uint32_t cPagesInChunk = cPagesLeft;
1722	if (cPagesInChunk > cPagesPerChunk)
1723	cPagesInChunk = cPagesPerChunk;
1724
1725	rc = pgmR3PhysRegisterHighRamChunk(pVM, GCPhysChunk, cPagesInChunk, cbChunk, iChunk, pszDesc, &pPrev);
1726	AssertRCReturn(rc, rc);
1727
1728	/* advance */
1729	GCPhysChunk += (RTGCPHYS)cPagesInChunk << PAGE_SHIFT;
1730	cPagesLeft -= cPagesInChunk;
1731	iChunk++;
1732	}
1733	}
1734	else
1735	{
1736	/*
1737	* Allocate, initialize and link the new RAM range.
1738	*/
1739	const size_t cbRamRange = RT_OFFSETOF(PGMRAMRANGE, aPages[cPages]);
1740	PPGMRAMRANGE pNew;
1741	rc = MMR3HyperAllocOnceNoRel(pVM, cbRamRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
1742	AssertLogRelMsgRCReturn(rc, ("cbRamRange=%zu\n", cbRamRange), rc);
1743
1744	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhysLast, NIL_RTRCPTR, NIL_RTR0PTR, pszDesc, pPrev);
1745	}
1746	pgmPhysInvalidatePageMapTLB(pVM);
1747	pgmUnlock(pVM);
1748
1749	#ifdef VBOX_WITH_REM
1750	/*
1751	* Notify REM.
1752	*/
1753	REMR3NotifyPhysRamRegister(pVM, GCPhys, cb, REM_NOTIFY_PHYS_RAM_FLAGS_RAM);
1754	#endif
1755
1756	return VINF_SUCCESS;
1757	}
1758
1759
1760	/**
1761	* Worker called by PGMR3InitFinalize if we're configured to pre-allocate RAM.
1762	*
1763	* We do this late in the init process so that all the ROM and MMIO ranges have
1764	* been registered already and we don't go wasting memory on them.
1765	*
1766	* @returns VBox status code.
1767	*
1768	* @param pVM Pointer to the VM.
1769	*/
1770	int pgmR3PhysRamPreAllocate(PVM pVM)
1771	{
1772	Assert(pVM->pgm.s.fRamPreAlloc);
1773	Log(("pgmR3PhysRamPreAllocate: enter\n"));
1774
1775	/*
1776	* Walk the RAM ranges and allocate all RAM pages, halt at
1777	* the first allocation error.
1778	*/
1779	uint64_t cPages = 0;
1780	uint64_t NanoTS = RTTimeNanoTS();
1781	pgmLock(pVM);
1782	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1783	{
1784	PPGMPAGE pPage = &pRam->aPages[0];
1785	RTGCPHYS GCPhys = pRam->GCPhys;
1786	uint32_t cLeft = pRam->cb >> PAGE_SHIFT;
1787	while (cLeft-- > 0)
1788	{
1789	if (PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1790	{
1791	switch (PGM_PAGE_GET_STATE(pPage))
1792	{
1793	case PGM_PAGE_STATE_ZERO:
1794	{
1795	int rc = pgmPhysAllocPage(pVM, pPage, GCPhys);
1796	if (RT_FAILURE(rc))
1797	{
1798	LogRel(("PGM: RAM Pre-allocation failed at %RGp (in %s) with rc=%Rrc\n", GCPhys, pRam->pszDesc, rc));
1799	pgmUnlock(pVM);
1800	return rc;
1801	}
1802	cPages++;
1803	break;
1804	}
1805
1806	case PGM_PAGE_STATE_BALLOONED:
1807	case PGM_PAGE_STATE_ALLOCATED:
1808	case PGM_PAGE_STATE_WRITE_MONITORED:
1809	case PGM_PAGE_STATE_SHARED:
1810	/* nothing to do here. */
1811	break;
1812	}
1813	}
1814
1815	/* next */
1816	pPage++;
1817	GCPhys += PAGE_SIZE;
1818	}
1819	}
1820	pgmUnlock(pVM);
1821	NanoTS = RTTimeNanoTS() - NanoTS;
1822
1823	LogRel(("PGM: Pre-allocated %llu pages in %llu ms\n", cPages, NanoTS / 1000000));
1824	Log(("pgmR3PhysRamPreAllocate: returns VINF_SUCCESS\n"));
1825	return VINF_SUCCESS;
1826	}
1827
1828
1829	/**
1830	* Checks shared page checksums.
1831	*
1832	* @param pVM Pointer to the shared VM structure.
1833	*/
1834	void pgmR3PhysAssertSharedPageChecksums(PVM pVM)
1835	{
1836	#ifdef VBOX_STRICT
1837	pgmLock(pVM);
1838
1839	if (pVM->pgm.s.cSharedPages > 0)
1840	{
1841	/*
1842	* Walk the ram ranges.
1843	*/
1844	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1845	{
1846	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
1847	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
1848
1849	while (iPage-- > 0)
1850	{
1851	PPGMPAGE pPage = &pRam->aPages[iPage];
1852	if (PGM_PAGE_IS_SHARED(pPage))
1853	{
1854	uint32_t u32Checksum = pPage->s.u2Unused0 \| ((uint32_t)pPage->s.u2Unused1 << 8);
1855	if (!u32Checksum)
1856	{
1857	RTGCPHYS GCPhysPage = pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT);
1858	void const *pvPage;
1859	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhysPage, &pvPage);
1860	if (RT_SUCCESS(rc))
1861	{
1862	uint32_t u32Checksum2 = RTCrc32(pvPage, PAGE_SIZE);
1863	# if 0
1864	AssertMsg((u32Checksum2 & UINT32_C(0x00000303)) == u32Checksum, ("GCPhysPage=%RGp\n", GCPhysPage));
1865	# else
1866	if ((u32Checksum2 & UINT32_C(0x00000303)) == u32Checksum)
1867	LogFlow(("shpg %#x @ %RGp %#x [OK]\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
1868	else
1869	AssertMsgFailed(("shpg %#x @ %RGp %#x\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
1870	# endif
1871	}
1872	else
1873	AssertRC(rc);
1874	}
1875	}
1876
1877	} /* for each page */
1878
1879	} /* for each ram range */
1880	}
1881
1882	pgmUnlock(pVM);
1883	#endif /* VBOX_STRICT */
1884	NOREF(pVM);
1885	}
1886
1887
1888	/**
1889	* Resets the physical memory state.
1890	*
1891	* ASSUMES that the caller owns the PGM lock.
1892	*
1893	* @returns VBox status code.
1894	* @param pVM Pointer to the VM.
1895	*/
1896	int pgmR3PhysRamReset(PVM pVM)
1897	{
1898	PGM_LOCK_ASSERT_OWNER(pVM);
1899
1900	/* Reset the memory balloon. */
1901	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
1902	AssertRC(rc);
1903
1904	#ifdef VBOX_WITH_PAGE_SHARING
1905	/* Clear all registered shared modules. */
1906	pgmR3PhysAssertSharedPageChecksums(pVM);
1907	rc = GMMR3ResetSharedModules(pVM);
1908	AssertRC(rc);
1909	#endif
1910	/* Reset counters. */
1911	pVM->pgm.s.cReusedSharedPages = 0;
1912	pVM->pgm.s.cBalloonedPages = 0;
1913
1914	return VINF_SUCCESS;
1915	}
1916
1917
1918	/**
1919	* Resets (zeros) the RAM after all devices and components have been reset.
1920	*
1921	* ASSUMES that the caller owns the PGM lock.
1922	*
1923	* @returns VBox status code.
1924	* @param pVM Pointer to the VM.
1925	*/
1926	int pgmR3PhysRamZeroAll(PVM pVM)
1927	{
1928	PGM_LOCK_ASSERT_OWNER(pVM);
1929
1930	/*
1931	* We batch up pages that should be freed instead of calling GMM for
1932	* each and every one of them.
1933	*/
1934	uint32_t cPendingPages = 0;
1935	PGMMFREEPAGESREQ pReq;
1936	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1937	AssertLogRelRCReturn(rc, rc);
1938
1939	/*
1940	* Walk the ram ranges.
1941	*/
1942	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1943	{
1944	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
1945	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
1946
1947	#ifndef NO_RAM_RESET
1948	if (!pVM->pgm.s.fRamPreAlloc)
1949	#else
1950	if (0)
1951	#endif
1952	{
1953	/* Replace all RAM pages by ZERO pages. */
1954	while (iPage-- > 0)
1955	{
1956	PPGMPAGE pPage = &pRam->aPages[iPage];
1957	switch (PGM_PAGE_GET_TYPE(pPage))
1958	{
1959	case PGMPAGETYPE_RAM:
1960	/* Do not replace pages part of a 2 MB continuous range
1961	with zero pages, but zero them instead. */
1962	if ( PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE
1963	\|\| PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE_DISABLED)
1964	{
1965	void *pvPage;
1966	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
1967	AssertLogRelRCReturn(rc, rc);
1968	ASMMemZeroPage(pvPage);
1969	}
1970	else if (PGM_PAGE_IS_BALLOONED(pPage))
1971	{
1972	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
1973	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
1974	}
1975	else if (!PGM_PAGE_IS_ZERO(pPage))
1976	{
1977	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
1978	AssertLogRelRCReturn(rc, rc);
1979	}
1980	break;
1981
1982	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
1983	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
1984	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
1985	true /fDoAccounting/);
1986	break;
1987
1988	case PGMPAGETYPE_MMIO2:
1989	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
1990	case PGMPAGETYPE_ROM:
1991	case PGMPAGETYPE_MMIO:
1992	break;
1993	default:
1994	AssertFailed();
1995	}
1996	} /* for each page */
1997	}
1998	else
1999	{
2000	/* Zero the memory. */
2001	while (iPage-- > 0)
2002	{
2003	PPGMPAGE pPage = &pRam->aPages[iPage];
2004	switch (PGM_PAGE_GET_TYPE(pPage))
2005	{
2006	case PGMPAGETYPE_RAM:
2007	switch (PGM_PAGE_GET_STATE(pPage))
2008	{
2009	case PGM_PAGE_STATE_ZERO:
2010	break;
2011
2012	case PGM_PAGE_STATE_BALLOONED:
2013	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2014	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2015	break;
2016
2017	case PGM_PAGE_STATE_SHARED:
2018	case PGM_PAGE_STATE_WRITE_MONITORED:
2019	rc = pgmPhysPageMakeWritable(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2020	AssertLogRelRCReturn(rc, rc);
2021	/* no break */
2022
2023	case PGM_PAGE_STATE_ALLOCATED:
2024	{
2025	void *pvPage;
2026	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2027	AssertLogRelRCReturn(rc, rc);
2028	#ifndef NO_RAM_RESET
2029	ASMMemZeroPage(pvPage);
2030	#endif
2031	break;
2032	}
2033	}
2034	break;
2035
2036	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2037	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2038	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2039	true /fDoAccounting/);
2040	break;
2041
2042	case PGMPAGETYPE_MMIO2:
2043	case PGMPAGETYPE_ROM_SHADOW:
2044	case PGMPAGETYPE_ROM:
2045	case PGMPAGETYPE_MMIO:
2046	break;
2047	default:
2048	AssertFailed();
2049
2050	}
2051	} /* for each page */
2052	}
2053
2054	}
2055
2056	/*
2057	* Finish off any pages pending freeing.
2058	*/
2059	if (cPendingPages)
2060	{
2061	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2062	AssertLogRelRCReturn(rc, rc);
2063	}
2064	GMMR3FreePagesCleanup(pReq);
2065	return VINF_SUCCESS;
2066	}
2067
2068
2069	/**
2070	* Frees all RAM during VM termination
2071	*
2072	* ASSUMES that the caller owns the PGM lock.
2073	*
2074	* @returns VBox status code.
2075	* @param pVM Pointer to the VM.
2076	*/
2077	int pgmR3PhysRamTerm(PVM pVM)
2078	{
2079	PGM_LOCK_ASSERT_OWNER(pVM);
2080
2081	/* Reset the memory balloon. */
2082	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2083	AssertRC(rc);
2084
2085	#ifdef VBOX_WITH_PAGE_SHARING
2086	/*
2087	* Clear all registered shared modules.
2088	*/
2089	pgmR3PhysAssertSharedPageChecksums(pVM);
2090	rc = GMMR3ResetSharedModules(pVM);
2091	AssertRC(rc);
2092
2093	/*
2094	* Flush the handy pages updates to make sure no shared pages are hiding
2095	* in there. (No unlikely if the VM shuts down, apparently.)
2096	*/
2097	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_FLUSH_HANDY_PAGES, 0, NULL);
2098	#endif
2099
2100	/*
2101	* We batch up pages that should be freed instead of calling GMM for
2102	* each and every one of them.
2103	*/
2104	uint32_t cPendingPages = 0;
2105	PGMMFREEPAGESREQ pReq;
2106	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2107	AssertLogRelRCReturn(rc, rc);
2108
2109	/*
2110	* Walk the ram ranges.
2111	*/
2112	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2113	{
2114	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2115	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2116
2117	while (iPage-- > 0)
2118	{
2119	PPGMPAGE pPage = &pRam->aPages[iPage];
2120	switch (PGM_PAGE_GET_TYPE(pPage))
2121	{
2122	case PGMPAGETYPE_RAM:
2123	/* Free all shared pages. Private pages are automatically freed during GMM VM cleanup. */
2124	/** @todo change this to explicitly free private pages here. */
2125	if (PGM_PAGE_IS_SHARED(pPage))
2126	{
2127	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2128	AssertLogRelRCReturn(rc, rc);
2129	}
2130	break;
2131
2132	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2133	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO:
2134	case PGMPAGETYPE_MMIO2:
2135	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2136	case PGMPAGETYPE_ROM:
2137	case PGMPAGETYPE_MMIO:
2138	break;
2139	default:
2140	AssertFailed();
2141	}
2142	} /* for each page */
2143	}
2144
2145	/*
2146	* Finish off any pages pending freeing.
2147	*/
2148	if (cPendingPages)
2149	{
2150	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2151	AssertLogRelRCReturn(rc, rc);
2152	}
2153	GMMR3FreePagesCleanup(pReq);
2154	return VINF_SUCCESS;
2155	}
2156
2157
2158	/**
2159	* This is the interface IOM is using to register an MMIO region.
2160	*
2161	* It will check for conflicts and ensure that a RAM range structure
2162	* is present before calling the PGMR3HandlerPhysicalRegister API to
2163	* register the callbacks.
2164	*
2165	* @returns VBox status code.
2166	*
2167	* @param pVM Pointer to the VM.
2168	* @param GCPhys The start of the MMIO region.
2169	* @param cb The size of the MMIO region.
2170	* @param pfnHandlerR3 The address of the ring-3 handler. (IOMR3MMIOHandler)
2171	* @param pvUserR3 The user argument for R3.
2172	* @param pfnHandlerR0 The address of the ring-0 handler. (IOMMMIOHandler)
2173	* @param pvUserR0 The user argument for R0.
2174	* @param pfnHandlerRC The address of the RC handler. (IOMMMIOHandler)
2175	* @param pvUserRC The user argument for RC.
2176	* @param pszDesc The description of the MMIO region.
2177	*/
2178	VMMR3DECL(int) PGMR3PhysMMIORegister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb,
2179	R3PTRTYPE(PFNPGMR3PHYSHANDLER) pfnHandlerR3, RTR3PTR pvUserR3,
2180	R0PTRTYPE(PFNPGMR0PHYSHANDLER) pfnHandlerR0, RTR0PTR pvUserR0,
2181	RCPTRTYPE(PFNPGMRCPHYSHANDLER) pfnHandlerRC, RTRCPTR pvUserRC,
2182	R3PTRTYPE(const char *) pszDesc)
2183	{
2184	/*
2185	* Assert on some assumption.
2186	*/
2187	VM_ASSERT_EMT(pVM);
2188	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2189	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2190	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2191	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2192
2193	int rc = pgmLock(pVM);
2194	AssertRCReturn(rc, rc);
2195
2196	/*
2197	* Make sure there's a RAM range structure for the region.
2198	*/
2199	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2200	bool fRamExists = false;
2201	PPGMRAMRANGE pRamPrev = NULL;
2202	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2203	while (pRam && GCPhysLast >= pRam->GCPhys)
2204	{
2205	if ( GCPhysLast >= pRam->GCPhys
2206	&& GCPhys <= pRam->GCPhysLast)
2207	{
2208	/* Simplification: all within the same range. */
2209	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
2210	&& GCPhysLast <= pRam->GCPhysLast,
2211	("%RGp-%RGp (MMIO/%s) falls partly outside %RGp-%RGp (%s)\n",
2212	GCPhys, GCPhysLast, pszDesc,
2213	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2214	pgmUnlock(pVM),
2215	VERR_PGM_RAM_CONFLICT);
2216
2217	/* Check that it's all RAM or MMIO pages. */
2218	PCPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2219	uint32_t cLeft = cb >> PAGE_SHIFT;
2220	while (cLeft-- > 0)
2221	{
2222	AssertLogRelMsgReturnStmt( PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM
2223	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO,
2224	("%RGp-%RGp (MMIO/%s): %RGp is not a RAM or MMIO page - type=%d desc=%s\n",
2225	GCPhys, GCPhysLast, pszDesc, PGM_PAGE_GET_TYPE(pPage), pRam->pszDesc),
2226	pgmUnlock(pVM),
2227	VERR_PGM_RAM_CONFLICT);
2228	pPage++;
2229	}
2230
2231	/* Looks good. */
2232	fRamExists = true;
2233	break;
2234	}
2235
2236	/* next */
2237	pRamPrev = pRam;
2238	pRam = pRam->pNextR3;
2239	}
2240	PPGMRAMRANGE pNew;
2241	if (fRamExists)
2242	{
2243	pNew = NULL;
2244
2245	/*
2246	* Make all the pages in the range MMIO/ZERO pages, freeing any
2247	* RAM pages currently mapped here. This might not be 100% correct
2248	* for PCI memory, but we're doing the same thing for MMIO2 pages.
2249	*/
2250	rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, PGMPAGETYPE_MMIO);
2251	AssertRCReturnStmt(rc, pgmUnlock(pVM), rc);
2252
2253	/* Force a PGM pool flush as guest ram references have been changed. */
2254	/** @todo not entirely SMP safe; assuming for now the guest takes
2255	* care of this internally (not touch mapped mmio while changing the
2256	* mapping). */
2257	PVMCPU pVCpu = VMMGetCpu(pVM);
2258	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2259	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2260	}
2261	else
2262	{
2263
2264	/*
2265	* No RAM range, insert an ad hoc one.
2266	*
2267	* Note that we don't have to tell REM about this range because
2268	* PGMHandlerPhysicalRegisterEx will do that for us.
2269	*/
2270	Log(("PGMR3PhysMMIORegister: Adding ad hoc MMIO range for %RGp-%RGp %s\n", GCPhys, GCPhysLast, pszDesc));
2271
2272	const uint32_t cPages = cb >> PAGE_SHIFT;
2273	const size_t cbRamRange = RT_OFFSETOF(PGMRAMRANGE, aPages[cPages]);
2274	rc = MMHyperAlloc(pVM, RT_OFFSETOF(PGMRAMRANGE, aPages[cPages]), 16, MM_TAG_PGM_PHYS, (void **)&pNew);
2275	AssertLogRelMsgRCReturnStmt(rc, ("cbRamRange=%zu\n", cbRamRange), pgmUnlock(pVM), rc);
2276
2277	/* Initialize the range. */
2278	pNew->pSelfR0 = MMHyperCCToR0(pVM, pNew);
2279	pNew->pSelfRC = MMHyperCCToRC(pVM, pNew);
2280	pNew->GCPhys = GCPhys;
2281	pNew->GCPhysLast = GCPhysLast;
2282	pNew->cb = cb;
2283	pNew->pszDesc = pszDesc;
2284	pNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO;
2285	pNew->pvR3 = NULL;
2286	pNew->paLSPages = NULL;
2287
2288	uint32_t iPage = cPages;
2289	while (iPage-- > 0)
2290	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_MMIO);
2291	Assert(PGM_PAGE_GET_TYPE(&pNew->aPages[0]) == PGMPAGETYPE_MMIO);
2292
2293	/* update the page count stats. */
2294	pVM->pgm.s.cPureMmioPages += cPages;
2295	pVM->pgm.s.cAllPages += cPages;
2296
2297	/* link it */
2298	pgmR3PhysLinkRamRange(pVM, pNew, pRamPrev);
2299	}
2300
2301	/*
2302	* Register the access handler.
2303	*/
2304	rc = PGMHandlerPhysicalRegisterEx(pVM, PGMPHYSHANDLERTYPE_MMIO, GCPhys, GCPhysLast,
2305	pfnHandlerR3, pvUserR3,
2306	pfnHandlerR0, pvUserR0,
2307	pfnHandlerRC, pvUserRC, pszDesc);
2308	if ( RT_FAILURE(rc)
2309	&& !fRamExists)
2310	{
2311	pVM->pgm.s.cPureMmioPages -= cb >> PAGE_SHIFT;
2312	pVM->pgm.s.cAllPages -= cb >> PAGE_SHIFT;
2313
2314	/* remove the ad hoc range. */
2315	pgmR3PhysUnlinkRamRange2(pVM, pNew, pRamPrev);
2316	pNew->cb = pNew->GCPhys = pNew->GCPhysLast = NIL_RTGCPHYS;
2317	MMHyperFree(pVM, pRam);
2318	}
2319	pgmPhysInvalidatePageMapTLB(pVM);
2320
2321	pgmUnlock(pVM);
2322	return rc;
2323	}
2324
2325
2326	/**
2327	* This is the interface IOM is using to register an MMIO region.
2328	*
2329	* It will take care of calling PGMHandlerPhysicalDeregister and clean up
2330	* any ad hoc PGMRAMRANGE left behind.
2331	*
2332	* @returns VBox status code.
2333	* @param pVM Pointer to the VM.
2334	* @param GCPhys The start of the MMIO region.
2335	* @param cb The size of the MMIO region.
2336	*/
2337	VMMR3DECL(int) PGMR3PhysMMIODeregister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb)
2338	{
2339	VM_ASSERT_EMT(pVM);
2340
2341	int rc = pgmLock(pVM);
2342	AssertRCReturn(rc, rc);
2343
2344	/*
2345	* First deregister the handler, then check if we should remove the ram range.
2346	*/
2347	rc = PGMHandlerPhysicalDeregister(pVM, GCPhys);
2348	if (RT_SUCCESS(rc))
2349	{
2350	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2351	PPGMRAMRANGE pRamPrev = NULL;
2352	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2353	while (pRam && GCPhysLast >= pRam->GCPhys)
2354	{
2355	/** @todo We're being a bit too careful here. rewrite. */
2356	if ( GCPhysLast == pRam->GCPhysLast
2357	&& GCPhys == pRam->GCPhys)
2358	{
2359	Assert(pRam->cb == cb);
2360
2361	/*
2362	* See if all the pages are dead MMIO pages.
2363	*/
2364	uint32_t const cPages = cb >> PAGE_SHIFT;
2365	bool fAllMMIO = true;
2366	uint32_t iPage = 0;
2367	uint32_t cLeft = cPages;
2368	while (cLeft-- > 0)
2369	{
2370	PPGMPAGE pPage = &pRam->aPages[iPage];
2371	if ( !PGM_PAGE_IS_MMIO_OR_ALIAS(pPage)
2372	/\|\| not-out-of-action later /)
2373	{
2374	fAllMMIO = false;
2375	AssertMsgFailed(("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2376	break;
2377	}
2378	Assert( PGM_PAGE_IS_ZERO(pPage)
2379	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2380	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO);
2381	pPage++;
2382	}
2383	if (fAllMMIO)
2384	{
2385	/*
2386	* Ad-hoc range, unlink and free it.
2387	*/
2388	Log(("PGMR3PhysMMIODeregister: Freeing ad hoc MMIO range for %RGp-%RGp %s\n",
2389	GCPhys, GCPhysLast, pRam->pszDesc));
2390
2391	pVM->pgm.s.cAllPages -= cPages;
2392	pVM->pgm.s.cPureMmioPages -= cPages;
2393
2394	pgmR3PhysUnlinkRamRange2(pVM, pRam, pRamPrev);
2395	pRam->cb = pRam->GCPhys = pRam->GCPhysLast = NIL_RTGCPHYS;
2396	MMHyperFree(pVM, pRam);
2397	break;
2398	}
2399	}
2400
2401	/*
2402	* Range match? It will all be within one range (see PGMAllHandler.cpp).
2403	*/
2404	if ( GCPhysLast >= pRam->GCPhys
2405	&& GCPhys <= pRam->GCPhysLast)
2406	{
2407	Assert(GCPhys >= pRam->GCPhys);
2408	Assert(GCPhysLast <= pRam->GCPhysLast);
2409
2410	/*
2411	* Turn the pages back into RAM pages.
2412	*/
2413	uint32_t iPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2414	uint32_t cLeft = cb >> PAGE_SHIFT;
2415	while (cLeft--)
2416	{
2417	PPGMPAGE pPage = &pRam->aPages[iPage];
2418	AssertMsg( (PGM_PAGE_IS_MMIO(pPage) && PGM_PAGE_IS_ZERO(pPage))
2419	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2420	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO,
2421	("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2422	if (PGM_PAGE_IS_MMIO_OR_ALIAS(pPage))
2423	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_RAM);
2424	}
2425	break;
2426	}
2427
2428	/* next */
2429	pRamPrev = pRam;
2430	pRam = pRam->pNextR3;
2431	}
2432	}
2433
2434	/* Force a PGM pool flush as guest ram references have been changed. */
2435	/** @todo Not entirely SMP safe; assuming for now the guest takes care of
2436	* this internally (not touch mapped mmio while changing the mapping). */
2437	PVMCPU pVCpu = VMMGetCpu(pVM);
2438	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2439	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2440
2441	pgmPhysInvalidatePageMapTLB(pVM);
2442	pgmPhysInvalidRamRangeTlbs(pVM);
2443	pgmUnlock(pVM);
2444	return rc;
2445	}
2446
2447
2448	/**
2449	* Locate a MMIO2 range.
2450	*
2451	* @returns Pointer to the MMIO2 range.
2452	* @param pVM Pointer to the VM.
2453	* @param pDevIns The device instance owning the region.
2454	* @param iRegion The region.
2455	*/
2456	DECLINLINE(PPGMMMIO2RANGE) pgmR3PhysMMIO2Find(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion)
2457	{
2458	/*
2459	* Search the list.
2460	*/
2461	for (PPGMMMIO2RANGE pCur = pVM->pgm.s.pMmio2RangesR3; pCur; pCur = pCur->pNextR3)
2462	if ( pCur->pDevInsR3 == pDevIns
2463	&& pCur->iRegion == iRegion)
2464	return pCur;
2465	return NULL;
2466	}
2467
2468
2469	/**
2470	* Allocate and register an MMIO2 region.
2471	*
2472	* As mentioned elsewhere, MMIO2 is just RAM spelled differently. It's RAM
2473	* associated with a device. It is also non-shared memory with a permanent
2474	* ring-3 mapping and page backing (presently).
2475	*
2476	* A MMIO2 range may overlap with base memory if a lot of RAM is configured for
2477	* the VM, in which case we'll drop the base memory pages. Presently we will
2478	* make no attempt to preserve anything that happens to be present in the base
2479	* memory that is replaced, this is of course incorrectly but it's too much
2480	* effort.
2481	*
2482	* @returns VBox status code.
2483	* @retval VINF_SUCCESS on success, *ppv pointing to the R3 mapping of the
2484	* memory.
2485	* @retval VERR_ALREADY_EXISTS if the region already exists.
2486	*
2487	* @param pVM Pointer to the VM.
2488	* @param pDevIns The device instance owning the region.
2489	* @param iRegion The region number. If the MMIO2 memory is a PCI
2490	* I/O region this number has to be the number of that
2491	* region. Otherwise it can be any number safe
2492	* UINT8_MAX.
2493	* @param cb The size of the region. Must be page aligned.
2494	* @param fFlags Reserved for future use, must be zero.
2495	* @param ppv Where to store the pointer to the ring-3 mapping of
2496	* the memory.
2497	* @param pszDesc The description.
2498	*/
2499	VMMR3DECL(int) PGMR3PhysMMIO2Register(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS cb, uint32_t fFlags,
2500	void *ppv, const char pszDesc)
2501	{
2502	/*
2503	* Validate input.
2504	*/
2505	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2506	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2507	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
2508	AssertPtrReturn(ppv, VERR_INVALID_POINTER);
2509	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2510	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2511	AssertReturn(pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion) == NULL, VERR_ALREADY_EXISTS);
2512	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2513	AssertReturn(cb, VERR_INVALID_PARAMETER);
2514	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2515
2516	const uint32_t cPages = cb >> PAGE_SHIFT;
2517	AssertLogRelReturn(((RTGCPHYS)cPages << PAGE_SHIFT) == cb, VERR_INVALID_PARAMETER);
2518	AssertLogRelReturn(cPages <= PGM_MMIO2_MAX_PAGE_COUNT, VERR_NO_MEMORY);
2519
2520	/*
2521	* For the 2nd+ instance, mangle the description string so it's unique.
2522	*/
2523	if (pDevIns->iInstance > 0) /** @todo Move to PDMDevHlp.cpp and use a real string cache. */
2524	{
2525	pszDesc = MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s [%u]", pszDesc, pDevIns->iInstance);
2526	if (!pszDesc)
2527	return VERR_NO_MEMORY;
2528	}
2529
2530	/*
2531	* Allocate an MMIO2 range ID (not freed on failure).
2532	* The zero ID is not used as it could be confused with NIL_GMM_PAGEID.
2533	*/
2534	pgmLock(pVM);
2535	uint8_t idMmio2 = pVM->pgm.s.cMmio2Regions + 1;
2536	if (idMmio2 > PGM_MMIO2_MAX_RANGES)
2537	{
2538	pgmUnlock(pVM);
2539	AssertLogRelFailedReturn(VERR_PGM_TOO_MANY_MMIO2_RANGES);
2540	}
2541	pVM->pgm.s.cMmio2Regions = idMmio2;
2542	pgmUnlock(pVM);
2543
2544	/*
2545	* Try reserve and allocate the backing memory first as this is what is
2546	* most likely to fail.
2547	*/
2548	int rc = MMR3AdjustFixedReservation(pVM, cPages, pszDesc);
2549	if (RT_SUCCESS(rc))
2550	{
2551	void *pvPages;
2552	PSUPPAGE paPages = (PSUPPAGE)RTMemTmpAlloc(cPages * sizeof(SUPPAGE));
2553	if (RT_SUCCESS(rc))
2554	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, NULL /pR0Ptr/, paPages);
2555	if (RT_SUCCESS(rc))
2556	{
2557	memset(pvPages, 0, cPages * PAGE_SIZE);
2558
2559	/*
2560	* Create the MMIO2 range record for it.
2561	*/
2562	const size_t cbRange = RT_OFFSETOF(PGMMMIO2RANGE, RamRange.aPages[cPages]);
2563	PPGMMMIO2RANGE pNew;
2564	rc = MMR3HyperAllocOnceNoRel(pVM, cbRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
2565	AssertLogRelMsgRC(rc, ("cbRamRange=%zu\n", cbRange));
2566	if (RT_SUCCESS(rc))
2567	{
2568	pNew->pDevInsR3 = pDevIns;
2569	pNew->pvR3 = pvPages;
2570	//pNew->pNext = NULL;
2571	//pNew->fMapped = false;
2572	//pNew->fOverlapping = false;
2573	pNew->iRegion = iRegion;
2574	pNew->idSavedState = UINT8_MAX;
2575	pNew->idMmio2 = idMmio2;
2576	pNew->RamRange.pSelfR0 = MMHyperCCToR0(pVM, &pNew->RamRange);
2577	pNew->RamRange.pSelfRC = MMHyperCCToRC(pVM, &pNew->RamRange);
2578	pNew->RamRange.GCPhys = NIL_RTGCPHYS;
2579	pNew->RamRange.GCPhysLast = NIL_RTGCPHYS;
2580	pNew->RamRange.pszDesc = pszDesc;
2581	pNew->RamRange.cb = cb;
2582	pNew->RamRange.fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO2;
2583	pNew->RamRange.pvR3 = pvPages;
2584	//pNew->RamRange.paLSPages = NULL;
2585
2586	uint32_t iPage = cPages;
2587	while (iPage-- > 0)
2588	{
2589	PGM_PAGE_INIT(&pNew->RamRange.aPages[iPage],
2590	paPages[iPage].Phys,
2591	PGM_MMIO2_PAGEID_MAKE(idMmio2, iPage),
2592	PGMPAGETYPE_MMIO2, PGM_PAGE_STATE_ALLOCATED);
2593	}
2594
2595	/* update page count stats */
2596	pVM->pgm.s.cAllPages += cPages;
2597	pVM->pgm.s.cPrivatePages += cPages;
2598
2599	/*
2600	* Link it into the list.
2601	* Since there is no particular order, just push it.
2602	*/
2603	/** @todo we can save us the linked list now, just search the lookup table... */
2604	pgmLock(pVM);
2605	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == NULL);
2606	Assert(pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] == NIL_RTR0PTR);
2607	pNew->pNextR3 = pVM->pgm.s.pMmio2RangesR3;
2608	pVM->pgm.s.pMmio2RangesR3 = pNew;
2609	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = pNew;
2610	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = MMHyperCCToR0(pVM, pNew);
2611	pgmUnlock(pVM);
2612
2613	*ppv = pvPages;
2614	RTMemTmpFree(paPages);
2615	pgmPhysInvalidatePageMapTLB(pVM);
2616	return VINF_SUCCESS;
2617	}
2618
2619	SUPR3PageFreeEx(pvPages, cPages);
2620	}
2621	RTMemTmpFree(paPages);
2622	MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pszDesc);
2623	}
2624	if (pDevIns->iInstance > 0)
2625	MMR3HeapFree((void *)pszDesc);
2626	return rc;
2627	}
2628
2629
2630	/**
2631	* Deregisters and frees an MMIO2 region.
2632	*
2633	* Any physical (and virtual) access handlers registered for the region must
2634	* be deregistered before calling this function.
2635	*
2636	* @returns VBox status code.
2637	* @param pVM Pointer to the VM.
2638	* @param pDevIns The device instance owning the region.
2639	* @param iRegion The region. If it's UINT32_MAX it'll be a wildcard match.
2640	*/
2641	VMMR3DECL(int) PGMR3PhysMMIO2Deregister(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion)
2642	{
2643	/*
2644	* Validate input.
2645	*/
2646	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2647	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2648	AssertReturn(iRegion <= UINT8_MAX \|\| iRegion == UINT32_MAX, VERR_INVALID_PARAMETER);
2649
2650	pgmLock(pVM);
2651	int rc = VINF_SUCCESS;
2652	unsigned cFound = 0;
2653	PPGMMMIO2RANGE pPrev = NULL;
2654	PPGMMMIO2RANGE pCur = pVM->pgm.s.pMmio2RangesR3;
2655	while (pCur)
2656	{
2657	if ( pCur->pDevInsR3 == pDevIns
2658	&& ( iRegion == UINT32_MAX
2659	\|\| pCur->iRegion == iRegion))
2660	{
2661	cFound++;
2662
2663	/*
2664	* Unmap it if it's mapped.
2665	*/
2666	if (pCur->fMapped)
2667	{
2668	int rc2 = PGMR3PhysMMIO2Unmap(pVM, pCur->pDevInsR3, pCur->iRegion, pCur->RamRange.GCPhys);
2669	AssertRC(rc2);
2670	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
2671	rc = rc2;
2672	}
2673
2674	/*
2675	* Unlink it
2676	*/
2677	PPGMMMIO2RANGE pNext = pCur->pNextR3;
2678	if (pPrev)
2679	pPrev->pNextR3 = pNext;
2680	else
2681	pVM->pgm.s.pMmio2RangesR3 = pNext;
2682	pCur->pNextR3 = NULL;
2683
2684	uint8_t idMmio2 = pCur->idMmio2;
2685	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == pCur);
2686	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = NULL;
2687	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = NIL_RTR0PTR;
2688
2689	/*
2690	* Free the memory.
2691	*/
2692	int rc2 = SUPR3PageFreeEx(pCur->pvR3, pCur->RamRange.cb >> PAGE_SHIFT);
2693	AssertRC(rc2);
2694	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
2695	rc = rc2;
2696
2697	uint32_t const cPages = pCur->RamRange.cb >> PAGE_SHIFT;
2698	rc2 = MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pCur->RamRange.pszDesc);
2699	AssertRC(rc2);
2700	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
2701	rc = rc2;
2702
2703	/* we're leaking hyper memory here if done at runtime. */
2704	#ifdef VBOX_STRICT
2705	VMSTATE const enmState = VMR3GetState(pVM);
2706	AssertMsg( enmState == VMSTATE_POWERING_OFF
2707	\|\| enmState == VMSTATE_POWERING_OFF_LS
2708	\|\| enmState == VMSTATE_OFF
2709	\|\| enmState == VMSTATE_OFF_LS
2710	\|\| enmState == VMSTATE_DESTROYING
2711	\|\| enmState == VMSTATE_TERMINATED
2712	\|\| enmState == VMSTATE_CREATING
2713	, ("%s\n", VMR3GetStateName(enmState)));
2714	#endif
2715	/*rc = MMHyperFree(pVM, pCur);
2716	AssertRCReturn(rc, rc); - not safe, see the alloc call. */
2717
2718
2719	/* update page count stats */
2720	pVM->pgm.s.cAllPages -= cPages;
2721	pVM->pgm.s.cPrivatePages -= cPages;
2722
2723	/* next */
2724	pCur = pNext;
2725	}
2726	else
2727	{
2728	pPrev = pCur;
2729	pCur = pCur->pNextR3;
2730	}
2731	}
2732	pgmPhysInvalidatePageMapTLB(pVM);
2733	pgmUnlock(pVM);
2734	return !cFound && iRegion != UINT32_MAX ? VERR_NOT_FOUND : rc;
2735	}
2736
2737
2738	/**
2739	* Maps a MMIO2 region.
2740	*
2741	* This is done when a guest / the bios / state loading changes the
2742	* PCI config. The replacing of base memory has the same restrictions
2743	* as during registration, of course.
2744	*
2745	* @returns VBox status code.
2746	*
2747	* @param pVM Pointer to the VM.
2748	* @param pDevIns The device instance owning the region.
2749	*/
2750	VMMR3DECL(int) PGMR3PhysMMIO2Map(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS GCPhys)
2751	{
2752	/*
2753	* Validate input
2754	*/
2755	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2756	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2757	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
2758	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
2759	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
2760	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2761
2762	PPGMMMIO2RANGE pCur = pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion);
2763	AssertReturn(pCur, VERR_NOT_FOUND);
2764	AssertReturn(!pCur->fMapped, VERR_WRONG_ORDER);
2765	Assert(pCur->RamRange.GCPhys == NIL_RTGCPHYS);
2766	Assert(pCur->RamRange.GCPhysLast == NIL_RTGCPHYS);
2767
2768	const RTGCPHYS GCPhysLast = GCPhys + pCur->RamRange.cb - 1;
2769	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
2770
2771	/*
2772	* Find our location in the ram range list, checking for
2773	* restriction we don't bother implementing yet (partially overlapping).
2774	*/
2775	bool fRamExists = false;
2776	PPGMRAMRANGE pRamPrev = NULL;
2777	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2778	while (pRam && GCPhysLast >= pRam->GCPhys)
2779	{
2780	if ( GCPhys <= pRam->GCPhysLast
2781	&& GCPhysLast >= pRam->GCPhys)
2782	{
2783	/* completely within? */
2784	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
2785	&& GCPhysLast <= pRam->GCPhysLast,
2786	("%RGp-%RGp (MMIO2/%s) falls partly outside %RGp-%RGp (%s)\n",
2787	GCPhys, GCPhysLast, pCur->RamRange.pszDesc,
2788	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2789	VERR_PGM_RAM_CONFLICT);
2790	fRamExists = true;
2791	break;
2792	}
2793
2794	/* next */
2795	pRamPrev = pRam;
2796	pRam = pRam->pNextR3;
2797	}
2798	if (fRamExists)
2799	{
2800	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2801	uint32_t cPagesLeft = pCur->RamRange.cb >> PAGE_SHIFT;
2802	while (cPagesLeft-- > 0)
2803	{
2804	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
2805	("%RGp isn't a RAM page (%d) - mapping %RGp-%RGp (MMIO2/%s).\n",
2806	GCPhys, PGM_PAGE_GET_TYPE(pPage), GCPhys, GCPhysLast, pCur->RamRange.pszDesc),
2807	VERR_PGM_RAM_CONFLICT);
2808	pPage++;
2809	}
2810	}
2811	Log(("PGMR3PhysMMIO2Map: %RGp-%RGp fRamExists=%RTbool %s\n",
2812	GCPhys, GCPhysLast, fRamExists, pCur->RamRange.pszDesc));
2813
2814	/*
2815	* Make the changes.
2816	*/
2817	pgmLock(pVM);
2818
2819	pCur->RamRange.GCPhys = GCPhys;
2820	pCur->RamRange.GCPhysLast = GCPhysLast;
2821	pCur->fMapped = true;
2822	pCur->fOverlapping = fRamExists;
2823
2824	if (fRamExists)
2825	{
2826	/** @todo use pgmR3PhysFreePageRange here. */
2827	uint32_t cPendingPages = 0;
2828	PGMMFREEPAGESREQ pReq;
2829	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2830	AssertLogRelRCReturn(rc, rc);
2831
2832	/* replace the pages, freeing all present RAM pages. */
2833	PPGMPAGE pPageSrc = &pCur->RamRange.aPages[0];
2834	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2835	uint32_t cPagesLeft = pCur->RamRange.cb >> PAGE_SHIFT;
2836	while (cPagesLeft-- > 0)
2837	{
2838	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys);
2839	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
2840
2841	RTHCPHYS const HCPhys = PGM_PAGE_GET_HCPHYS(pPageSrc);
2842	PGM_PAGE_SET_HCPHYS(pVM, pPageDst, HCPhys);
2843	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO2);
2844	PGM_PAGE_SET_STATE(pVM, pPageDst, PGM_PAGE_STATE_ALLOCATED);
2845	PGM_PAGE_SET_PDE_TYPE(pVM, pPageDst, PGM_PAGE_PDE_TYPE_DONTCARE);
2846	PGM_PAGE_SET_PTE_INDEX(pVM, pPageDst, 0);
2847	PGM_PAGE_SET_TRACKING(pVM, pPageDst, 0);
2848
2849	pVM->pgm.s.cZeroPages--;
2850	GCPhys += PAGE_SIZE;
2851	pPageSrc++;
2852	pPageDst++;
2853	}
2854
2855	/* Flush physical page map TLB. */
2856	pgmPhysInvalidatePageMapTLB(pVM);
2857
2858	if (cPendingPages)
2859	{
2860	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2861	AssertLogRelRCReturn(rc, rc);
2862	}
2863	GMMR3FreePagesCleanup(pReq);
2864
2865	/* Force a PGM pool flush as guest ram references have been changed. */
2866	/** @todo not entirely SMP safe; assuming for now the guest takes care of
2867	* this internally (not touch mapped mmio while changing the mapping). */
2868	PVMCPU pVCpu = VMMGetCpu(pVM);
2869	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2870	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2871
2872	pgmUnlock(pVM);
2873	}
2874	else
2875	{
2876	RTGCPHYS cb = pCur->RamRange.cb;
2877
2878	/* Clear the tracking data of pages we're going to reactivate. */
2879	PPGMPAGE pPageSrc = &pCur->RamRange.aPages[0];
2880	uint32_t cPagesLeft = pCur->RamRange.cb >> PAGE_SHIFT;
2881	while (cPagesLeft-- > 0)
2882	{
2883	PGM_PAGE_SET_TRACKING(pVM, pPageSrc, 0);
2884	PGM_PAGE_SET_PTE_INDEX(pVM, pPageSrc, 0);
2885	pPageSrc++;
2886	}
2887
2888	/* link in the ram range */
2889	pgmR3PhysLinkRamRange(pVM, &pCur->RamRange, pRamPrev);
2890	pgmUnlock(pVM);
2891
2892	#ifdef VBOX_WITH_REM
2893	REMR3NotifyPhysRamRegister(pVM, GCPhys, cb, REM_NOTIFY_PHYS_RAM_FLAGS_MMIO2);
2894	#endif
2895	}
2896
2897	pgmPhysInvalidatePageMapTLB(pVM);
2898	return VINF_SUCCESS;
2899	}
2900
2901
2902	/**
2903	* Unmaps a MMIO2 region.
2904	*
2905	* This is done when a guest / the bios / state loading changes the
2906	* PCI config. The replacing of base memory has the same restrictions
2907	* as during registration, of course.
2908	*/
2909	VMMR3DECL(int) PGMR3PhysMMIO2Unmap(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS GCPhys)
2910	{
2911	/*
2912	* Validate input
2913	*/
2914	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2915	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2916	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
2917	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
2918	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
2919	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2920
2921	PPGMMMIO2RANGE pCur = pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion);
2922	AssertReturn(pCur, VERR_NOT_FOUND);
2923	AssertReturn(pCur->fMapped, VERR_WRONG_ORDER);
2924	AssertReturn(pCur->RamRange.GCPhys == GCPhys, VERR_INVALID_PARAMETER);
2925	Assert(pCur->RamRange.GCPhysLast != NIL_RTGCPHYS);
2926
2927	Log(("PGMR3PhysMMIO2Unmap: %RGp-%RGp %s\n",
2928	pCur->RamRange.GCPhys, pCur->RamRange.GCPhysLast, pCur->RamRange.pszDesc));
2929
2930	/*
2931	* Unmap it.
2932	*/
2933	pgmLock(pVM);
2934
2935	#ifdef VBOX_WITH_REM
2936	RTGCPHYS GCPhysRangeREM;
2937	RTGCPHYS cbRangeREM;
2938	bool fInformREM;
2939	#endif
2940	if (pCur->fOverlapping)
2941	{
2942	/* Restore the RAM pages we've replaced. */
2943	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2944	while (pRam->GCPhys > pCur->RamRange.GCPhysLast)
2945	pRam = pRam->pNextR3;
2946
2947	PPGMPAGE pPageDst = &pRam->aPages[(pCur->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2948	uint32_t cPagesLeft = pCur->RamRange.cb >> PAGE_SHIFT;
2949	while (cPagesLeft-- > 0)
2950	{
2951	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
2952	pVM->pgm.s.cZeroPages++;
2953	pPageDst++;
2954	}
2955
2956	/* Flush physical page map TLB. */
2957	pgmPhysInvalidatePageMapTLB(pVM);
2958	#ifdef VBOX_WITH_REM
2959	GCPhysRangeREM = NIL_RTGCPHYS; /* shuts up gcc */
2960	cbRangeREM = RTGCPHYS_MAX; /* ditto */
2961	fInformREM = false;
2962	#endif
2963	}
2964	else
2965	{
2966	#ifdef VBOX_WITH_REM
2967	GCPhysRangeREM = pCur->RamRange.GCPhys;
2968	cbRangeREM = pCur->RamRange.cb;
2969	fInformREM = true;
2970	#endif
2971	pgmR3PhysUnlinkRamRange(pVM, &pCur->RamRange);
2972	}
2973
2974	pCur->RamRange.GCPhys = NIL_RTGCPHYS;
2975	pCur->RamRange.GCPhysLast = NIL_RTGCPHYS;
2976	pCur->fOverlapping = false;
2977	pCur->fMapped = false;
2978
2979	/* Force a PGM pool flush as guest ram references have been changed. */
2980	/** @todo not entirely SMP safe; assuming for now the guest takes care
2981	* of this internally (not touch mapped mmio while changing the
2982	* mapping). */
2983	PVMCPU pVCpu = VMMGetCpu(pVM);
2984	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2985	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2986
2987	pgmPhysInvalidatePageMapTLB(pVM);
2988	pgmPhysInvalidRamRangeTlbs(pVM);
2989	pgmUnlock(pVM);
2990
2991	#ifdef VBOX_WITH_REM
2992	if (fInformREM)
2993	REMR3NotifyPhysRamDeregister(pVM, GCPhysRangeREM, cbRangeREM);
2994	#endif
2995
2996	return VINF_SUCCESS;
2997	}
2998
2999
3000	/**
3001	* Checks if the given address is an MMIO2 base address or not.
3002	*
3003	* @returns true/false accordingly.
3004	* @param pVM Pointer to the VM.
3005	* @param pDevIns The owner of the memory, optional.
3006	* @param GCPhys The address to check.
3007	*/
3008	VMMR3DECL(bool) PGMR3PhysMMIO2IsBase(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys)
3009	{
3010	/*
3011	* Validate input
3012	*/
3013	VM_ASSERT_EMT_RETURN(pVM, false);
3014	AssertPtrReturn(pDevIns, false);
3015	AssertReturn(GCPhys != NIL_RTGCPHYS, false);
3016	AssertReturn(GCPhys != 0, false);
3017	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), false);
3018
3019	/*
3020	* Search the list.
3021	*/
3022	pgmLock(pVM);
3023	for (PPGMMMIO2RANGE pCur = pVM->pgm.s.pMmio2RangesR3; pCur; pCur = pCur->pNextR3)
3024	if (pCur->RamRange.GCPhys == GCPhys)
3025	{
3026	Assert(pCur->fMapped);
3027	pgmUnlock(pVM);
3028	return true;
3029	}
3030	pgmUnlock(pVM);
3031	return false;
3032	}
3033
3034
3035	/**
3036	* Gets the HC physical address of a page in the MMIO2 region.
3037	*
3038	* This is API is intended for MMHyper and shouldn't be called
3039	* by anyone else...
3040	*
3041	* @returns VBox status code.
3042	* @param pVM Pointer to the VM.
3043	* @param pDevIns The owner of the memory, optional.
3044	* @param iRegion The region.
3045	* @param off The page expressed an offset into the MMIO2 region.
3046	* @param pHCPhys Where to store the result.
3047	*/
3048	VMMR3DECL(int) PGMR3PhysMMIO2GetHCPhys(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS off, PRTHCPHYS pHCPhys)
3049	{
3050	/*
3051	* Validate input
3052	*/
3053	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3054	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3055	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3056
3057	pgmLock(pVM);
3058	PPGMMMIO2RANGE pCur = pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion);
3059	AssertReturn(pCur, VERR_NOT_FOUND);
3060	AssertReturn(off < pCur->RamRange.cb, VERR_INVALID_PARAMETER);
3061
3062	PCPGMPAGE pPage = &pCur->RamRange.aPages[off >> PAGE_SHIFT];
3063	*pHCPhys = PGM_PAGE_GET_HCPHYS(pPage);
3064	pgmUnlock(pVM);
3065	return VINF_SUCCESS;
3066	}
3067
3068
3069	/**
3070	* Maps a portion of an MMIO2 region into kernel space (host).
3071	*
3072	* The kernel mapping will become invalid when the MMIO2 memory is deregistered
3073	* or the VM is terminated.
3074	*
3075	* @return VBox status code.
3076	*
3077	* @param pVM Pointer to the VM.
3078	* @param pDevIns The device owning the MMIO2 memory.
3079	* @param iRegion The region.
3080	* @param off The offset into the region. Must be page aligned.
3081	* @param cb The number of bytes to map. Must be page aligned.
3082	* @param pszDesc Mapping description.
3083	* @param pR0Ptr Where to store the R0 address.
3084	*/
3085	VMMR3DECL(int) PGMR3PhysMMIO2MapKernel(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS off, RTGCPHYS cb,
3086	const char *pszDesc, PRTR0PTR pR0Ptr)
3087	{
3088	/*
3089	* Validate input.
3090	*/
3091	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3092	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3093	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3094
3095	PPGMMMIO2RANGE pCur = pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion);
3096	AssertReturn(pCur, VERR_NOT_FOUND);
3097	AssertReturn(off < pCur->RamRange.cb, VERR_INVALID_PARAMETER);
3098	AssertReturn(cb <= pCur->RamRange.cb, VERR_INVALID_PARAMETER);
3099	AssertReturn(off + cb <= pCur->RamRange.cb, VERR_INVALID_PARAMETER);
3100	NOREF(pszDesc);
3101
3102	/*
3103	* Pass the request on to the support library/driver.
3104	*/
3105	int rc = SUPR3PageMapKernel(pCur->pvR3, off, cb, 0, pR0Ptr);
3106
3107	return rc;
3108	}
3109
3110
3111	/**
3112	* Worker for PGMR3PhysRomRegister.
3113	*
3114	* This is here to simplify lock management, i.e. the caller does all the
3115	* locking and we can simply return without needing to remember to unlock
3116	* anything first.
3117	*
3118	* @returns VBox status.
3119	* @param pVM Pointer to the VM.
3120	* @param pDevIns The device instance owning the ROM.
3121	* @param GCPhys First physical address in the range.
3122	* Must be page aligned!
3123	* @param cb The size of the range (in bytes).
3124	* Must be page aligned!
3125	* @param pvBinary Pointer to the binary data backing the ROM image.
3126	* @param cbBinary The size of the binary data pvBinary points to.
3127	* This must be less or equal to @a cb.
3128	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
3129	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
3130	* @param pszDesc Pointer to description string. This must not be freed.
3131	*/
3132	static int pgmR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
3133	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
3134	{
3135	/*
3136	* Validate input.
3137	*/
3138	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3139	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
3140	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
3141	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
3142	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3143	AssertPtrReturn(pvBinary, VERR_INVALID_PARAMETER);
3144	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
3145	AssertReturn(!(fFlags & ~(PGMPHYS_ROM_FLAGS_SHADOWED \| PGMPHYS_ROM_FLAGS_PERMANENT_BINARY)), VERR_INVALID_PARAMETER);
3146	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_CREATING, VERR_VM_INVALID_VM_STATE);
3147
3148	const uint32_t cPages = cb >> PAGE_SHIFT;
3149
3150	/*
3151	* Find the ROM location in the ROM list first.
3152	*/
3153	PPGMROMRANGE pRomPrev = NULL;
3154	PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3;
3155	while (pRom && GCPhysLast >= pRom->GCPhys)
3156	{
3157	if ( GCPhys <= pRom->GCPhysLast
3158	&& GCPhysLast >= pRom->GCPhys)
3159	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
3160	GCPhys, GCPhysLast, pszDesc,
3161	pRom->GCPhys, pRom->GCPhysLast, pRom->pszDesc),
3162	VERR_PGM_RAM_CONFLICT);
3163	/* next */
3164	pRomPrev = pRom;
3165	pRom = pRom->pNextR3;
3166	}
3167
3168	/*
3169	* Find the RAM location and check for conflicts.
3170	*
3171	* Conflict detection is a bit different than for RAM
3172	* registration since a ROM can be located within a RAM
3173	* range. So, what we have to check for is other memory
3174	* types (other than RAM that is) and that we don't span
3175	* more than one RAM range (layz).
3176	*/
3177	bool fRamExists = false;
3178	PPGMRAMRANGE pRamPrev = NULL;
3179	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3180	while (pRam && GCPhysLast >= pRam->GCPhys)
3181	{
3182	if ( GCPhys <= pRam->GCPhysLast
3183	&& GCPhysLast >= pRam->GCPhys)
3184	{
3185	/* completely within? */
3186	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
3187	&& GCPhysLast <= pRam->GCPhysLast,
3188	("%RGp-%RGp (%s) falls partly outside %RGp-%RGp (%s)\n",
3189	GCPhys, GCPhysLast, pszDesc,
3190	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3191	VERR_PGM_RAM_CONFLICT);
3192	fRamExists = true;
3193	break;
3194	}
3195
3196	/* next */
3197	pRamPrev = pRam;
3198	pRam = pRam->pNextR3;
3199	}
3200	if (fRamExists)
3201	{
3202	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3203	uint32_t cPagesLeft = cPages;
3204	while (cPagesLeft-- > 0)
3205	{
3206	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
3207	("%RGp (%R[pgmpage]) isn't a RAM page - registering %RGp-%RGp (%s).\n",
3208	pRam->GCPhys + ((RTGCPHYS)(uintptr_t)(pPage - &pRam->aPages[0]) << PAGE_SHIFT),
3209	pPage, GCPhys, GCPhysLast, pszDesc), VERR_PGM_RAM_CONFLICT);
3210	Assert(PGM_PAGE_IS_ZERO(pPage));
3211	pPage++;
3212	}
3213	}
3214
3215	/*
3216	* Update the base memory reservation if necessary.
3217	*/
3218	uint32_t cExtraBaseCost = fRamExists ? 0 : cPages;
3219	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
3220	cExtraBaseCost += cPages;
3221	if (cExtraBaseCost)
3222	{
3223	int rc = MMR3IncreaseBaseReservation(pVM, cExtraBaseCost);
3224	if (RT_FAILURE(rc))
3225	return rc;
3226	}
3227
3228	/*
3229	* Allocate memory for the virgin copy of the RAM.
3230	*/
3231	PGMMALLOCATEPAGESREQ pReq;
3232	int rc = GMMR3AllocatePagesPrepare(pVM, &pReq, cPages, GMMACCOUNT_BASE);
3233	AssertRCReturn(rc, rc);
3234
3235	for (uint32_t iPage = 0; iPage < cPages; iPage++)
3236	{
3237	pReq->aPages[iPage].HCPhysGCPhys = GCPhys + (iPage << PAGE_SHIFT);
3238	pReq->aPages[iPage].idPage = NIL_GMM_PAGEID;
3239	pReq->aPages[iPage].idSharedPage = NIL_GMM_PAGEID;
3240	}
3241
3242	rc = GMMR3AllocatePagesPerform(pVM, pReq);
3243	if (RT_FAILURE(rc))
3244	{
3245	GMMR3AllocatePagesCleanup(pReq);
3246	return rc;
3247	}
3248
3249	/*
3250	* Allocate the new ROM range and RAM range (if necessary).
3251	*/
3252	PPGMROMRANGE pRomNew;
3253	rc = MMHyperAlloc(pVM, RT_OFFSETOF(PGMROMRANGE, aPages[cPages]), 0, MM_TAG_PGM_PHYS, (void **)&pRomNew);
3254	if (RT_SUCCESS(rc))
3255	{
3256	PPGMRAMRANGE pRamNew = NULL;
3257	if (!fRamExists)
3258	rc = MMHyperAlloc(pVM, RT_OFFSETOF(PGMRAMRANGE, aPages[cPages]), sizeof(PGMPAGE), MM_TAG_PGM_PHYS, (void **)&pRamNew);
3259	if (RT_SUCCESS(rc))
3260	{
3261	/*
3262	* Initialize and insert the RAM range (if required).
3263	*/
3264	PPGMROMPAGE pRomPage = &pRomNew->aPages[0];
3265	if (!fRamExists)
3266	{
3267	pRamNew->pSelfR0 = MMHyperCCToR0(pVM, pRamNew);
3268	pRamNew->pSelfRC = MMHyperCCToRC(pVM, pRamNew);
3269	pRamNew->GCPhys = GCPhys;
3270	pRamNew->GCPhysLast = GCPhysLast;
3271	pRamNew->cb = cb;
3272	pRamNew->pszDesc = pszDesc;
3273	pRamNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_ROM;
3274	pRamNew->pvR3 = NULL;
3275	pRamNew->paLSPages = NULL;
3276
3277	PPGMPAGE pPage = &pRamNew->aPages[0];
3278	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
3279	{
3280	PGM_PAGE_INIT(pPage,
3281	pReq->aPages[iPage].HCPhysGCPhys,
3282	pReq->aPages[iPage].idPage,
3283	PGMPAGETYPE_ROM,
3284	PGM_PAGE_STATE_ALLOCATED);
3285
3286	pRomPage->Virgin = *pPage;
3287	}
3288
3289	pVM->pgm.s.cAllPages += cPages;
3290	pgmR3PhysLinkRamRange(pVM, pRamNew, pRamPrev);
3291	}
3292	else
3293	{
3294	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3295	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
3296	{
3297	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_ROM);
3298	PGM_PAGE_SET_HCPHYS(pVM, pPage, pReq->aPages[iPage].HCPhysGCPhys);
3299	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
3300	PGM_PAGE_SET_PAGEID(pVM, pPage, pReq->aPages[iPage].idPage);
3301	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
3302	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
3303	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
3304
3305	pRomPage->Virgin = *pPage;
3306	}
3307
3308	pRamNew = pRam;
3309
3310	pVM->pgm.s.cZeroPages -= cPages;
3311	}
3312	pVM->pgm.s.cPrivatePages += cPages;
3313
3314	/* Flush physical page map TLB. */
3315	pgmPhysInvalidatePageMapTLB(pVM);
3316
3317
3318	/*
3319	* !HACK ALERT! REM + (Shadowed) ROM ==> mess.
3320	*
3321	* If it's shadowed we'll register the handler after the ROM notification
3322	* so we get the access handler callbacks that we should. If it isn't
3323	* shadowed we'll do it the other way around to make REM use the built-in
3324	* ROM behavior and not the handler behavior (which is to route all access
3325	* to PGM atm).
3326	*/
3327	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
3328	{
3329	#ifdef VBOX_WITH_REM
3330	REMR3NotifyPhysRomRegister(pVM, GCPhys, cb, NULL, true /* fShadowed */);
3331	#endif
3332	rc = PGMR3HandlerPhysicalRegister(pVM, PGMPHYSHANDLERTYPE_PHYSICAL_WRITE, GCPhys, GCPhysLast,
3333	pgmR3PhysRomWriteHandler, pRomNew,
3334	NULL, "pgmPhysRomWriteHandler", MMHyperCCToR0(pVM, pRomNew),
3335	NULL, "pgmPhysRomWriteHandler", MMHyperCCToRC(pVM, pRomNew), pszDesc);
3336	}
3337	else
3338	{
3339	rc = PGMR3HandlerPhysicalRegister(pVM, PGMPHYSHANDLERTYPE_PHYSICAL_WRITE, GCPhys, GCPhysLast,
3340	pgmR3PhysRomWriteHandler, pRomNew,
3341	NULL, "pgmPhysRomWriteHandler", MMHyperCCToR0(pVM, pRomNew),
3342	NULL, "pgmPhysRomWriteHandler", MMHyperCCToRC(pVM, pRomNew), pszDesc);
3343	#ifdef VBOX_WITH_REM
3344	REMR3NotifyPhysRomRegister(pVM, GCPhys, cb, NULL, false /* fShadowed */);
3345	#endif
3346	}
3347	if (RT_SUCCESS(rc))
3348	{
3349	/*
3350	* Copy the image over to the virgin pages.
3351	* This must be done after linking in the RAM range.
3352	*/
3353	size_t cbBinaryLeft = cbBinary;
3354	PPGMPAGE pRamPage = &pRamNew->aPages[(GCPhys - pRamNew->GCPhys) >> PAGE_SHIFT];
3355	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
3356	{
3357	void *pvDstPage;
3358	rc = pgmPhysPageMap(pVM, pRamPage, GCPhys + (iPage << PAGE_SHIFT), &pvDstPage);
3359	if (RT_FAILURE(rc))
3360	{
3361	VMSetError(pVM, rc, RT_SRC_POS, "Failed to map virgin ROM page at %RGp", GCPhys);
3362	break;
3363	}
3364	if (cbBinaryLeft >= PAGE_SIZE)
3365	{
3366	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), PAGE_SIZE);
3367	cbBinaryLeft -= PAGE_SIZE;
3368	}
3369	else
3370	{
3371	ASMMemZeroPage(pvDstPage); /* (shouldn't be necessary, but can't hurt either) */
3372	if (cbBinaryLeft > 0)
3373	{
3374	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), cbBinaryLeft);
3375	cbBinaryLeft = 0;
3376	}
3377	}
3378	}
3379	if (RT_SUCCESS(rc))
3380	{
3381	/*
3382	* Initialize the ROM range.
3383	* Note that the Virgin member of the pages has already been initialized above.
3384	*/
3385	pRomNew->GCPhys = GCPhys;
3386	pRomNew->GCPhysLast = GCPhysLast;
3387	pRomNew->cb = cb;
3388	pRomNew->fFlags = fFlags;
3389	pRomNew->idSavedState = UINT8_MAX;
3390	pRomNew->cbOriginal = cbBinary;
3391	#ifdef VBOX_STRICT
3392	pRomNew->pvOriginal = fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY
3393	? pvBinary : RTMemDup(pvBinary, cbBinary);
3394	#else
3395	pRomNew->pvOriginal = fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY ? pvBinary : NULL;
3396	#endif
3397	pRomNew->pszDesc = pszDesc;
3398
3399	for (unsigned iPage = 0; iPage < cPages; iPage++)
3400	{
3401	PPGMROMPAGE pPage = &pRomNew->aPages[iPage];
3402	pPage->enmProt = PGMROMPROT_READ_ROM_WRITE_IGNORE;
3403	PGM_PAGE_INIT_ZERO(&pPage->Shadow, pVM, PGMPAGETYPE_ROM_SHADOW);
3404	}
3405
3406	/* update the page count stats for the shadow pages. */
3407	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
3408	{
3409	pVM->pgm.s.cZeroPages += cPages;
3410	pVM->pgm.s.cAllPages += cPages;
3411	}
3412
3413	/*
3414	* Insert the ROM range, tell REM and return successfully.
3415	*/
3416	pRomNew->pNextR3 = pRom;
3417	pRomNew->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
3418	pRomNew->pNextRC = pRom ? MMHyperCCToRC(pVM, pRom) : NIL_RTRCPTR;
3419
3420	if (pRomPrev)
3421	{
3422	pRomPrev->pNextR3 = pRomNew;
3423	pRomPrev->pNextR0 = MMHyperCCToR0(pVM, pRomNew);
3424	pRomPrev->pNextRC = MMHyperCCToRC(pVM, pRomNew);
3425	}
3426	else
3427	{
3428	pVM->pgm.s.pRomRangesR3 = pRomNew;
3429	pVM->pgm.s.pRomRangesR0 = MMHyperCCToR0(pVM, pRomNew);
3430	pVM->pgm.s.pRomRangesRC = MMHyperCCToRC(pVM, pRomNew);
3431	}
3432
3433	pgmPhysInvalidatePageMapTLB(pVM);
3434	GMMR3AllocatePagesCleanup(pReq);
3435	return VINF_SUCCESS;
3436	}
3437
3438	/* bail out */
3439
3440	int rc2 = PGMHandlerPhysicalDeregister(pVM, GCPhys);
3441	AssertRC(rc2);
3442	}
3443
3444	if (!fRamExists)
3445	{
3446	pgmR3PhysUnlinkRamRange2(pVM, pRamNew, pRamPrev);
3447	MMHyperFree(pVM, pRamNew);
3448	}
3449	}
3450	MMHyperFree(pVM, pRomNew);
3451	}
3452
3453	/** @todo Purge the mapping cache or something... */
3454	GMMR3FreeAllocatedPages(pVM, pReq);
3455	GMMR3AllocatePagesCleanup(pReq);
3456	return rc;
3457	}
3458
3459
3460	/**
3461	* Registers a ROM image.
3462	*
3463	* Shadowed ROM images requires double the amount of backing memory, so,
3464	* don't use that unless you have to. Shadowing of ROM images is process
3465	* where we can select where the reads go and where the writes go. On real
3466	* hardware the chipset provides means to configure this. We provide
3467	* PGMR3PhysProtectROM() for this purpose.
3468	*
3469	* A read-only copy of the ROM image will always be kept around while we
3470	* will allocate RAM pages for the changes on demand (unless all memory
3471	* is configured to be preallocated).
3472	*
3473	* @returns VBox status.
3474	* @param pVM Pointer to the VM.
3475	* @param pDevIns The device instance owning the ROM.
3476	* @param GCPhys First physical address in the range.
3477	* Must be page aligned!
3478	* @param cb The size of the range (in bytes).
3479	* Must be page aligned!
3480	* @param pvBinary Pointer to the binary data backing the ROM image.
3481	* @param cbBinary The size of the binary data pvBinary points to.
3482	* This must be less or equal to @a cb.
3483	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
3484	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
3485	* @param pszDesc Pointer to description string. This must not be freed.
3486	*
3487	* @remark There is no way to remove the rom, automatically on device cleanup or
3488	* manually from the device yet. This isn't difficult in any way, it's
3489	* just not something we expect to be necessary for a while.
3490	*/
3491	VMMR3DECL(int) PGMR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
3492	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
3493	{
3494	Log(("PGMR3PhysRomRegister: pDevIns=%p GCPhys=%RGp(-%RGp) cb=%RGp pvBinary=%p cbBinary=%#x fFlags=%#x pszDesc=%s\n",
3495	pDevIns, GCPhys, GCPhys + cb, cb, pvBinary, cbBinary, fFlags, pszDesc));
3496	pgmLock(pVM);
3497	int rc = pgmR3PhysRomRegister(pVM, pDevIns, GCPhys, cb, pvBinary, cbBinary, fFlags, pszDesc);
3498	pgmUnlock(pVM);
3499	return rc;
3500	}
3501
3502
3503	/**
3504	* \#PF Handler callback for ROM write accesses.
3505	*
3506	* @returns VINF_SUCCESS if the handler have carried out the operation.
3507	* @returns VINF_PGM_HANDLER_DO_DEFAULT if the caller should carry out the access operation.
3508	* @param pVM Pointer to the VM.
3509	* @param GCPhys The physical address the guest is writing to.
3510	* @param pvPhys The HC mapping of that address.
3511	* @param pvBuf What the guest is reading/writing.
3512	* @param cbBuf How much it's reading/writing.
3513	* @param enmAccessType The access type.
3514	* @param pvUser User argument.
3515	*/
3516	static DECLCALLBACK(int) pgmR3PhysRomWriteHandler(PVM pVM, RTGCPHYS GCPhys, void pvPhys, void pvBuf, size_t cbBuf,
3517	PGMACCESSTYPE enmAccessType, void *pvUser)
3518	{
3519	PPGMROMRANGE pRom = (PPGMROMRANGE)pvUser;
3520	const uint32_t iPage = (GCPhys - pRom->GCPhys) >> PAGE_SHIFT;
3521	Assert(iPage < (pRom->cb >> PAGE_SHIFT));
3522	PPGMROMPAGE pRomPage = &pRom->aPages[iPage];
3523	Log5(("pgmR3PhysRomWriteHandler: %d %c %#08RGp %#04zx\n", pRomPage->enmProt, enmAccessType == PGMACCESSTYPE_READ ? 'R' : 'W', GCPhys, cbBuf));
3524	NOREF(pvPhys);
3525
3526	if (enmAccessType == PGMACCESSTYPE_READ)
3527	{
3528	switch (pRomPage->enmProt)
3529	{
3530	/*
3531	* Take the default action.
3532	*/
3533	case PGMROMPROT_READ_ROM_WRITE_IGNORE:
3534	case PGMROMPROT_READ_RAM_WRITE_IGNORE:
3535	case PGMROMPROT_READ_ROM_WRITE_RAM:
3536	case PGMROMPROT_READ_RAM_WRITE_RAM:
3537	return VINF_PGM_HANDLER_DO_DEFAULT;
3538
3539	default:
3540	AssertMsgFailedReturn(("enmProt=%d iPage=%d GCPhys=%RGp\n",
3541	pRom->aPages[iPage].enmProt, iPage, GCPhys),
3542	VERR_IPE_NOT_REACHED_DEFAULT_CASE);
3543	}
3544	}
3545	else
3546	{
3547	Assert(enmAccessType == PGMACCESSTYPE_WRITE);
3548	switch (pRomPage->enmProt)
3549	{
3550	/*
3551	* Ignore writes.
3552	*/
3553	case PGMROMPROT_READ_ROM_WRITE_IGNORE:
3554	case PGMROMPROT_READ_RAM_WRITE_IGNORE:
3555	return VINF_SUCCESS;
3556
3557	/*
3558	* Write to the RAM page.
3559	*/
3560	case PGMROMPROT_READ_ROM_WRITE_RAM:
3561	case PGMROMPROT_READ_RAM_WRITE_RAM: /* yes this will get here too, it's way simpler that way. */
3562	{
3563	/* This should be impossible now, pvPhys doesn't work cross page anylonger. */
3564	Assert(((GCPhys - pRom->GCPhys + cbBuf - 1) >> PAGE_SHIFT) == iPage);
3565
3566	/*
3567	* Take the lock, do lazy allocation, map the page and copy the data.
3568	*
3569	* Note that we have to bypass the mapping TLB since it works on
3570	* guest physical addresses and entering the shadow page would
3571	* kind of screw things up...
3572	*/
3573	int rc = pgmLock(pVM);
3574	AssertRC(rc);
3575
3576	PPGMPAGE pShadowPage = &pRomPage->Shadow;
3577	if (!PGMROMPROT_IS_ROM(pRomPage->enmProt))
3578	{
3579	pShadowPage = pgmPhysGetPage(pVM, GCPhys);
3580	AssertLogRelReturn(pShadowPage, VERR_PGM_PHYS_PAGE_GET_IPE);
3581	}
3582
3583	void *pvDstPage;
3584	rc = pgmPhysPageMakeWritableAndMap(pVM, pShadowPage, GCPhys & X86_PTE_PG_MASK, &pvDstPage);
3585	if (RT_SUCCESS(rc))
3586	{
3587	memcpy((uint8_t *)pvDstPage + (GCPhys & PAGE_OFFSET_MASK), pvBuf, cbBuf);
3588	pRomPage->LiveSave.fWrittenTo = true;
3589	}
3590
3591	pgmUnlock(pVM);
3592	return rc;
3593	}
3594
3595	default:
3596	AssertMsgFailedReturn(("enmProt=%d iPage=%d GCPhys=%RGp\n",
3597	pRom->aPages[iPage].enmProt, iPage, GCPhys),
3598	VERR_IPE_NOT_REACHED_DEFAULT_CASE);
3599	}
3600	}
3601	}
3602
3603
3604	/**
3605	* Called by PGMR3MemSetup to reset the shadow, switch to the virgin, and verify
3606	* that the virgin part is untouched.
3607	*
3608	* This is done after the normal memory has been cleared.
3609	*
3610	* ASSUMES that the caller owns the PGM lock.
3611	*
3612	* @param pVM Pointer to the VM.
3613	*/
3614	int pgmR3PhysRomReset(PVM pVM)
3615	{
3616	PGM_LOCK_ASSERT_OWNER(pVM);
3617	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
3618	{
3619	const uint32_t cPages = pRom->cb >> PAGE_SHIFT;
3620
3621	if (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
3622	{
3623	/*
3624	* Reset the physical handler.
3625	*/
3626	int rc = PGMR3PhysRomProtect(pVM, pRom->GCPhys, pRom->cb, PGMROMPROT_READ_ROM_WRITE_IGNORE);
3627	AssertRCReturn(rc, rc);
3628
3629	/*
3630	* What we do with the shadow pages depends on the memory
3631	* preallocation option. If not enabled, we'll just throw
3632	* out all the dirty pages and replace them by the zero page.
3633	*/
3634	if (!pVM->pgm.s.fRamPreAlloc)
3635	{
3636	/* Free the dirty pages. */
3637	uint32_t cPendingPages = 0;
3638	PGMMFREEPAGESREQ pReq;
3639	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
3640	AssertRCReturn(rc, rc);
3641
3642	for (uint32_t iPage = 0; iPage < cPages; iPage++)
3643	if ( !PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow)
3644	&& !PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow))
3645	{
3646	Assert(PGM_PAGE_GET_STATE(&pRom->aPages[iPage].Shadow) == PGM_PAGE_STATE_ALLOCATED);
3647	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, &pRom->aPages[iPage].Shadow,
3648	pRom->GCPhys + (iPage << PAGE_SHIFT));
3649	AssertLogRelRCReturn(rc, rc);
3650	}
3651
3652	if (cPendingPages)
3653	{
3654	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
3655	AssertLogRelRCReturn(rc, rc);
3656	}
3657	GMMR3FreePagesCleanup(pReq);
3658	}
3659	else
3660	{
3661	/* clear all the shadow pages. */
3662	for (uint32_t iPage = 0; iPage < cPages; iPage++)
3663	{
3664	if (PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow))
3665	continue;
3666	Assert(!PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow));
3667	void *pvDstPage;
3668	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
3669	rc = pgmPhysPageMakeWritableAndMap(pVM, &pRom->aPages[iPage].Shadow, GCPhys, &pvDstPage);
3670	if (RT_FAILURE(rc))
3671	break;
3672	ASMMemZeroPage(pvDstPage);
3673	}
3674	AssertRCReturn(rc, rc);
3675	}
3676	}
3677
3678	#ifdef VBOX_STRICT
3679	/*
3680	* Verify that the virgin page is unchanged if possible.
3681	*/
3682	if (pRom->pvOriginal)
3683	{
3684	size_t cbSrcLeft = pRom->cbOriginal;
3685	uint8_t const pbSrcPage = (uint8_t const )pRom->pvOriginal;
3686	for (uint32_t iPage = 0; iPage < cPages && cbSrcLeft > 0; iPage++, pbSrcPage += PAGE_SIZE)
3687	{
3688	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
3689	void const *pvDstPage;
3690	int rc = pgmPhysPageMapReadOnly(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPage);
3691	if (RT_FAILURE(rc))
3692	break;
3693
3694	if (memcmp(pvDstPage, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE)))
3695	{
3696	# ifdef DEBUG_bird /* This is darn handy for EFI debugging w/ snapshots, should be made default later. */
3697	void *pvDstPageW;
3698	rc = pgmPhysPageMap(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPageW);
3699	AssertRCReturn(rc, rc);
3700	memcpy(pvDstPageW, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE));
3701	# else
3702	LogRel(("pgmR3PhysRomReset: %RGp rom page changed (%s) - loaded saved state?\n",
3703	GCPhys, pRom->pszDesc));
3704	# endif
3705	}
3706	cbSrcLeft -= RT_MIN(cbSrcLeft, PAGE_SIZE);
3707	}
3708	}
3709	#endif
3710	}
3711
3712	return VINF_SUCCESS;
3713	}
3714
3715
3716	/**
3717	* Called by PGMR3Term to free resources.
3718	*
3719	* ASSUMES that the caller owns the PGM lock.
3720	*
3721	* @param pVM Pointer to the VM.
3722	*/
3723	void pgmR3PhysRomTerm(PVM pVM)
3724	{
3725	#ifdef RT_STRICT
3726	/*
3727	* Free the heap copy of the original bits.
3728	*/
3729	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
3730	{
3731	if ( pRom->pvOriginal
3732	&& !(pRom->fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY))
3733	{
3734	RTMemFree((void *)pRom->pvOriginal);
3735	pRom->pvOriginal = NULL;
3736	}
3737	}
3738	#endif
3739	}
3740
3741
3742	/**
3743	* Change the shadowing of a range of ROM pages.
3744	*
3745	* This is intended for implementing chipset specific memory registers
3746	* and will not be very strict about the input. It will silently ignore
3747	* any pages that are not the part of a shadowed ROM.
3748	*
3749	* @returns VBox status code.
3750	* @retval VINF_PGM_SYNC_CR3
3751	*
3752	* @param pVM Pointer to the VM.
3753	* @param GCPhys Where to start. Page aligned.
3754	* @param cb How much to change. Page aligned.
3755	* @param enmProt The new ROM protection.
3756	*/
3757	VMMR3DECL(int) PGMR3PhysRomProtect(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMROMPROT enmProt)
3758	{
3759	/*
3760	* Check input
3761	*/
3762	if (!cb)
3763	return VINF_SUCCESS;
3764	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3765	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3766	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
3767	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3768	AssertReturn(enmProt >= PGMROMPROT_INVALID && enmProt <= PGMROMPROT_END, VERR_INVALID_PARAMETER);
3769
3770	/*
3771	* Process the request.
3772	*/
3773	pgmLock(pVM);
3774	int rc = VINF_SUCCESS;
3775	bool fFlushTLB = false;
3776	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
3777	{
3778	if ( GCPhys <= pRom->GCPhysLast
3779	&& GCPhysLast >= pRom->GCPhys
3780	&& (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED))
3781	{
3782	/*
3783	* Iterate the relevant pages and make necessary the changes.
3784	*/
3785	bool fChanges = false;
3786	uint32_t const cPages = pRom->GCPhysLast <= GCPhysLast
3787	? pRom->cb >> PAGE_SHIFT
3788	: (GCPhysLast - pRom->GCPhys + 1) >> PAGE_SHIFT;
3789	for (uint32_t iPage = (GCPhys - pRom->GCPhys) >> PAGE_SHIFT;
3790	iPage < cPages;
3791	iPage++)
3792	{
3793	PPGMROMPAGE pRomPage = &pRom->aPages[iPage];
3794	if (PGMROMPROT_IS_ROM(pRomPage->enmProt) != PGMROMPROT_IS_ROM(enmProt))
3795	{
3796	fChanges = true;
3797
3798	/* flush references to the page. */
3799	PPGMPAGE pRamPage = pgmPhysGetPage(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT));
3800	int rc2 = pgmPoolTrackUpdateGCPhys(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT), pRamPage,
3801	true /fFlushPTEs/, &fFlushTLB);
3802	if (rc2 != VINF_SUCCESS && (rc == VINF_SUCCESS \|\| RT_FAILURE(rc2)))
3803	rc = rc2;
3804
3805	PPGMPAGE pOld = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Virgin : &pRomPage->Shadow;
3806	PPGMPAGE pNew = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Shadow : &pRomPage->Virgin;
3807
3808	pOld = pRamPage;
3809	pRamPage = pNew;
3810	/** @todo preserve the volatile flags (handlers) when these have been moved out of HCPhys! */
3811	}
3812	pRomPage->enmProt = enmProt;
3813	}
3814
3815	/*
3816	* Reset the access handler if we made changes, no need
3817	* to optimize this.
3818	*/
3819	if (fChanges)
3820	{
3821	int rc2 = PGMHandlerPhysicalReset(pVM, pRom->GCPhys);
3822	if (RT_FAILURE(rc2))
3823	{
3824	pgmUnlock(pVM);
3825	AssertRC(rc);
3826	return rc2;
3827	}
3828	}
3829
3830	/* Advance - cb isn't updated. */
3831	GCPhys = pRom->GCPhys + (cPages << PAGE_SHIFT);
3832	}
3833	}
3834	pgmUnlock(pVM);
3835	if (fFlushTLB)
3836	PGM_INVL_ALL_VCPU_TLBS(pVM);
3837
3838	return rc;
3839	}
3840
3841
3842	/**
3843	* Sets the Address Gate 20 state.
3844	*
3845	* @param pVCpu Pointer to the VMCPU.
3846	* @param fEnable True if the gate should be enabled.
3847	* False if the gate should be disabled.
3848	*/
3849	VMMDECL(void) PGMR3PhysSetA20(PVMCPU pVCpu, bool fEnable)
3850	{
3851	LogFlow(("PGMR3PhysSetA20 %d (was %d)\n", fEnable, pVCpu->pgm.s.fA20Enabled));
3852	if (pVCpu->pgm.s.fA20Enabled != fEnable)
3853	{
3854	pVCpu->pgm.s.fA20Enabled = fEnable;
3855	pVCpu->pgm.s.GCPhysA20Mask = ~(RTGCPHYS)(!fEnable << 20);
3856	#ifdef VBOX_WITH_REM
3857	REMR3A20Set(pVCpu->pVMR3, pVCpu, fEnable);
3858	#endif
3859	#ifdef PGM_WITH_A20
3860	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_UPDATE_PAGE_BIT_VIRTUAL;
3861	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3862	pgmR3RefreshShadowModeAfterA20Change(pVCpu);
3863	HMFlushTLB(pVCpu);
3864	#endif
3865	STAM_REL_COUNTER_INC(&pVCpu->pgm.s.cA20Changes);
3866	}
3867	}
3868
3869
3870	/**
3871	* Tree enumeration callback for dealing with age rollover.
3872	* It will perform a simple compression of the current age.
3873	*/
3874	static DECLCALLBACK(int) pgmR3PhysChunkAgeingRolloverCallback(PAVLU32NODECORE pNode, void *pvUser)
3875	{
3876	/* Age compression - ASSUMES iNow == 4. */
3877	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
3878	if (pChunk->iLastUsed >= UINT32_C(0xffffff00))
3879	pChunk->iLastUsed = 3;
3880	else if (pChunk->iLastUsed >= UINT32_C(0xfffff000))
3881	pChunk->iLastUsed = 2;
3882	else if (pChunk->iLastUsed)
3883	pChunk->iLastUsed = 1;
3884	else /* iLastUsed = 0 */
3885	pChunk->iLastUsed = 4;
3886
3887	NOREF(pvUser);
3888	return 0;
3889	}
3890
3891
3892	/**
3893	* The structure passed in the pvUser argument of pgmR3PhysChunkUnmapCandidateCallback().
3894	*/
3895	typedef struct PGMR3PHYSCHUNKUNMAPCB
3896	{
3897	PVM pVM; /*< Pointer to the VM. /
3898	PPGMCHUNKR3MAP pChunk; /*< The chunk to unmap. /
3899	} PGMR3PHYSCHUNKUNMAPCB, *PPGMR3PHYSCHUNKUNMAPCB;
3900
3901
3902	/**
3903	* Callback used to find the mapping that's been unused for
3904	* the longest time.
3905	*/
3906	static DECLCALLBACK(int) pgmR3PhysChunkUnmapCandidateCallback(PAVLU32NODECORE pNode, void *pvUser)
3907	{
3908	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
3909	PPGMR3PHYSCHUNKUNMAPCB pArg = (PPGMR3PHYSCHUNKUNMAPCB)pvUser;
3910
3911	/*
3912	* Check for locks and compare when last used.
3913	*/
3914	if (pChunk->cRefs)
3915	return 0;
3916	if (pChunk->cPermRefs)
3917	return 0;
3918	if ( pArg->pChunk
3919	&& pChunk->iLastUsed >= pArg->pChunk->iLastUsed)
3920	return 0;
3921
3922	/*
3923	* Check that it's not in any of the TLBs.
3924	*/
3925	PVM pVM = pArg->pVM;
3926	if ( pVM->pgm.s.ChunkR3Map.Tlb.aEntries[PGM_CHUNKR3MAPTLB_IDX(pChunk->Core.Key)].idChunk
3927	== pChunk->Core.Key)
3928	{
3929	pChunk = NULL;
3930	return 0;
3931	}
3932	#ifdef VBOX_STRICT
3933	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
3934	{
3935	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk != pChunk);
3936	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk != pChunk->Core.Key);
3937	}
3938	#endif
3939
3940	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.PhysTlbHC.aEntries); i++)
3941	if (pVM->pgm.s.PhysTlbHC.aEntries[i].pMap == pChunk)
3942	return 0;
3943
3944	pArg->pChunk = pChunk;
3945	return 0;
3946	}
3947
3948
3949	/**
3950	* Finds a good candidate for unmapping when the ring-3 mapping cache is full.
3951	*
3952	* The candidate will not be part of any TLBs, so no need to flush
3953	* anything afterwards.
3954	*
3955	* @returns Chunk id.
3956	* @param pVM Pointer to the VM.
3957	*/
3958	static int32_t pgmR3PhysChunkFindUnmapCandidate(PVM pVM)
3959	{
3960	PGM_LOCK_ASSERT_OWNER(pVM);
3961
3962	/*
3963	* Enumerate the age tree starting with the left most node.
3964	*/
3965	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
3966	PGMR3PHYSCHUNKUNMAPCB Args;
3967	Args.pVM = pVM;
3968	Args.pChunk = NULL;
3969	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkUnmapCandidateCallback, &Args);
3970	Assert(Args.pChunk);
3971	if (Args.pChunk)
3972	{
3973	Assert(Args.pChunk->cRefs == 0);
3974	Assert(Args.pChunk->cPermRefs == 0);
3975	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
3976	return Args.pChunk->Core.Key;
3977	}
3978
3979	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
3980	return INT32_MAX;
3981	}
3982
3983
3984	/**
3985	* Rendezvous callback used by pgmR3PhysUnmapChunk that unmaps a chunk
3986	*
3987	* This is only called on one of the EMTs while the other ones are waiting for
3988	* it to complete this function.
3989	*
3990	* @returns VINF_SUCCESS (VBox strict status code).
3991	* @param pVM Pointer to the VM.
3992	* @param pVCpu The VMCPU for the EMT we're being called on. Unused.
3993	* @param pvUser User pointer. Unused
3994	*
3995	*/
3996	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysUnmapChunkRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
3997	{
3998	int rc = VINF_SUCCESS;
3999	pgmLock(pVM);
4000	NOREF(pVCpu); NOREF(pvUser);
4001
4002	if (pVM->pgm.s.ChunkR3Map.c >= pVM->pgm.s.ChunkR3Map.cMax)
4003	{
4004	/* Flush the pgm pool cache; call the internal rendezvous handler as we're already in a rendezvous handler here. */
4005	/** @todo also not really efficient to unmap a chunk that contains PD
4006	* or PT pages. */
4007	pgmR3PoolClearAllRendezvous(pVM, &pVM->aCpus[0], NULL /* no need to flush the REM TLB as we already did that above */);
4008
4009	/*
4010	* Request the ring-0 part to unmap a chunk to make space in the mapping cache.
4011	*/
4012	GMMMAPUNMAPCHUNKREQ Req;
4013	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
4014	Req.Hdr.cbReq = sizeof(Req);
4015	Req.pvR3 = NULL;
4016	Req.idChunkMap = NIL_GMM_CHUNKID;
4017	Req.idChunkUnmap = pgmR3PhysChunkFindUnmapCandidate(pVM);
4018	if (Req.idChunkUnmap != INT32_MAX)
4019	{
4020	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkUnmap, a);
4021	rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
4022	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkUnmap, a);
4023	if (RT_SUCCESS(rc))
4024	{
4025	/*
4026	* Remove the unmapped one.
4027	*/
4028	PPGMCHUNKR3MAP pUnmappedChunk = (PPGMCHUNKR3MAP)RTAvlU32Remove(&pVM->pgm.s.ChunkR3Map.pTree, Req.idChunkUnmap);
4029	AssertRelease(pUnmappedChunk);
4030	AssertRelease(!pUnmappedChunk->cRefs);
4031	AssertRelease(!pUnmappedChunk->cPermRefs);
4032	pUnmappedChunk->pv = NULL;
4033	pUnmappedChunk->Core.Key = UINT32_MAX;
4034	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
4035	MMR3HeapFree(pUnmappedChunk);
4036	#else
4037	MMR3UkHeapFree(pVM, pUnmappedChunk, MM_TAG_PGM_CHUNK_MAPPING);
4038	#endif
4039	pVM->pgm.s.ChunkR3Map.c--;
4040	pVM->pgm.s.cUnmappedChunks++;
4041
4042	/*
4043	* Flush dangling PGM pointers (R3 & R0 ptrs to GC physical addresses).
4044	*/
4045	/** todo: we should not flush chunks which include cr3 mappings. */
4046	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
4047	{
4048	PPGMCPU pPGM = &pVM->aCpus[idCpu].pgm.s;
4049
4050	pPGM->pGst32BitPdR3 = NULL;
4051	pPGM->pGstPaePdptR3 = NULL;
4052	pPGM->pGstAmd64Pml4R3 = NULL;
4053	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4054	pPGM->pGst32BitPdR0 = NIL_RTR0PTR;
4055	pPGM->pGstPaePdptR0 = NIL_RTR0PTR;
4056	pPGM->pGstAmd64Pml4R0 = NIL_RTR0PTR;
4057	#endif
4058	for (unsigned i = 0; i < RT_ELEMENTS(pPGM->apGstPaePDsR3); i++)
4059	{
4060	pPGM->apGstPaePDsR3[i] = NULL;
4061	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
4062	pPGM->apGstPaePDsR0[i] = NIL_RTR0PTR;
4063	#endif
4064	}
4065
4066	/* Flush REM TLBs. */
4067	CPUMSetChangedFlags(&pVM->aCpus[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
4068	}
4069	#ifdef VBOX_WITH_REM
4070	/* Flush REM translation blocks. */
4071	REMFlushTBs(pVM);
4072	#endif
4073	}
4074	}
4075	}
4076	pgmUnlock(pVM);
4077	return rc;
4078	}
4079
4080	/**
4081	* Unmap a chunk to free up virtual address space (request packet handler for pgmR3PhysChunkMap)
4082	*
4083	* @returns VBox status code.
4084	* @param pVM Pointer to the VM.
4085	*/
4086	void pgmR3PhysUnmapChunk(PVM pVM)
4087	{
4088	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysUnmapChunkRendezvous, NULL);
4089	AssertRC(rc);
4090	}
4091
4092
4093	/**
4094	* Maps the given chunk into the ring-3 mapping cache.
4095	*
4096	* This will call ring-0.
4097	*
4098	* @returns VBox status code.
4099	* @param pVM Pointer to the VM.
4100	* @param idChunk The chunk in question.
4101	* @param ppChunk Where to store the chunk tracking structure.
4102	*
4103	* @remarks Called from within the PGM critical section.
4104	* @remarks Can be called from any thread!
4105	*/
4106	int pgmR3PhysChunkMap(PVM pVM, uint32_t idChunk, PPPGMCHUNKR3MAP ppChunk)
4107	{
4108	int rc;
4109
4110	PGM_LOCK_ASSERT_OWNER(pVM);
4111
4112	/*
4113	* Move the chunk time forward.
4114	*/
4115	pVM->pgm.s.ChunkR3Map.iNow++;
4116	if (pVM->pgm.s.ChunkR3Map.iNow == 0)
4117	{
4118	pVM->pgm.s.ChunkR3Map.iNow = 4;
4119	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkAgeingRolloverCallback, NULL);
4120	}
4121
4122	/*
4123	* Allocate a new tracking structure first.
4124	*/
4125	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
4126	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3HeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk));
4127	#else
4128	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3UkHeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk), NULL);
4129	#endif
4130	AssertReturn(pChunk, VERR_NO_MEMORY);
4131	pChunk->Core.Key = idChunk;
4132	pChunk->iLastUsed = pVM->pgm.s.ChunkR3Map.iNow;
4133
4134	/*
4135	* Request the ring-0 part to map the chunk in question.
4136	*/
4137	GMMMAPUNMAPCHUNKREQ Req;
4138	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
4139	Req.Hdr.cbReq = sizeof(Req);
4140	Req.pvR3 = NULL;
4141	Req.idChunkMap = idChunk;
4142	Req.idChunkUnmap = NIL_GMM_CHUNKID;
4143
4144	/* Must be callable from any thread, so can't use VMMR3CallR0. */
4145	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkMap, a);
4146	rc = SUPR3CallVMMR0Ex(pVM->pVMR0, NIL_VMCPUID, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
4147	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkMap, a);
4148	if (RT_SUCCESS(rc))
4149	{
4150	pChunk->pv = Req.pvR3;
4151
4152	/*
4153	* If we're running out of virtual address space, then we should
4154	* unmap another chunk.
4155	*
4156	* Currently, an unmap operation requires that all other virtual CPUs
4157	* are idling and not by chance making use of the memory we're
4158	* unmapping. So, we create an async unmap operation here.
4159	*
4160	* Now, when creating or restoring a saved state this wont work very
4161	* well since we may want to restore all guest RAM + a little something.
4162	* So, we have to do the unmap synchronously. Fortunately for us
4163	* though, during these operations the other virtual CPUs are inactive
4164	* and it should be safe to do this.
4165	*/
4166	/** @todo Eventually we should lock all memory when used and do
4167	* map+unmap as one kernel call without any rendezvous or
4168	* other precautions. */
4169	if (pVM->pgm.s.ChunkR3Map.c + 1 >= pVM->pgm.s.ChunkR3Map.cMax)
4170	{
4171	switch (VMR3GetState(pVM))
4172	{
4173	case VMSTATE_LOADING:
4174	case VMSTATE_SAVING:
4175	{
4176	PVMCPU pVCpu = VMMGetCpu(pVM);
4177	if ( pVCpu
4178	&& pVM->pgm.s.cDeprecatedPageLocks == 0)
4179	{
4180	pgmR3PhysUnmapChunkRendezvous(pVM, pVCpu, NULL);
4181	break;
4182	}
4183	/* fall thru */
4184	}
4185	default:
4186	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysUnmapChunk, 1, pVM);
4187	AssertRC(rc);
4188	break;
4189	}
4190	}
4191
4192	/*
4193	* Update the tree. We must do this after any unmapping to make sure
4194	* the chunk we're going to return isn't unmapped by accident.
4195	*/
4196	AssertPtr(Req.pvR3);
4197	bool fRc = RTAvlU32Insert(&pVM->pgm.s.ChunkR3Map.pTree, &pChunk->Core);
4198	AssertRelease(fRc);
4199	pVM->pgm.s.ChunkR3Map.c++;
4200	pVM->pgm.s.cMappedChunks++;
4201	}
4202	else
4203	{
4204	/** @todo this may fail because of /proc/sys/vm/max_map_count, so we
4205	* should probably restrict ourselves on linux. */
4206	AssertRC(rc);
4207	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
4208	MMR3HeapFree(pChunk);
4209	#else
4210	MMR3UkHeapFree(pVM, pChunk, MM_TAG_PGM_CHUNK_MAPPING);
4211	#endif
4212	pChunk = NULL;
4213	}
4214
4215	*ppChunk = pChunk;
4216	return rc;
4217	}
4218
4219
4220	/**
4221	* For VMMCALLRING3_PGM_MAP_CHUNK, considered internal.
4222	*
4223	* @returns see pgmR3PhysChunkMap.
4224	* @param pVM Pointer to the VM.
4225	* @param idChunk The chunk to map.
4226	*/
4227	VMMR3DECL(int) PGMR3PhysChunkMap(PVM pVM, uint32_t idChunk)
4228	{
4229	PPGMCHUNKR3MAP pChunk;
4230	int rc;
4231
4232	pgmLock(pVM);
4233	rc = pgmR3PhysChunkMap(pVM, idChunk, &pChunk);
4234	pgmUnlock(pVM);
4235	return rc;
4236	}
4237
4238
4239	/**
4240	* Invalidates the TLB for the ring-3 mapping cache.
4241	*
4242	* @param pVM Pointer to the VM.
4243	*/
4244	VMMR3DECL(void) PGMR3PhysChunkInvalidateTLB(PVM pVM)
4245	{
4246	pgmLock(pVM);
4247	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
4248	{
4249	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk = NIL_GMM_CHUNKID;
4250	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk = NULL;
4251	}
4252	/* The page map TLB references chunks, so invalidate that one too. */
4253	pgmPhysInvalidatePageMapTLB(pVM);
4254	pgmUnlock(pVM);
4255	}
4256
4257
4258	/**
4259	* Response to VMMCALLRING3_PGM_ALLOCATE_LARGE_HANDY_PAGE to allocate a large
4260	* (2MB) page for use with a nested paging PDE.
4261	*
4262	* @returns The following VBox status codes.
4263	* @retval VINF_SUCCESS on success.
4264	* @retval VINF_EM_NO_MEMORY if we're out of memory.
4265	*
4266	* @param pVM Pointer to the VM.
4267	* @param GCPhys GC physical start address of the 2 MB range
4268	*/
4269	VMMR3DECL(int) PGMR3PhysAllocateLargeHandyPage(PVM pVM, RTGCPHYS GCPhys)
4270	{
4271	#ifdef PGM_WITH_LARGE_PAGES
4272	uint64_t u64TimeStamp1, u64TimeStamp2;
4273
4274	pgmLock(pVM);
4275
4276	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatAllocLargePage, a);
4277	u64TimeStamp1 = RTTimeMilliTS();
4278	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_LARGE_HANDY_PAGE, 0, NULL);
4279	u64TimeStamp2 = RTTimeMilliTS();
4280	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatAllocLargePage, a);
4281	if (RT_SUCCESS(rc))
4282	{
4283	Assert(pVM->pgm.s.cLargeHandyPages == 1);
4284
4285	uint32_t idPage = pVM->pgm.s.aLargeHandyPage[0].idPage;
4286	RTHCPHYS HCPhys = pVM->pgm.s.aLargeHandyPage[0].HCPhysGCPhys;
4287
4288	void *pv;
4289
4290	/* Map the large page into our address space.
4291	*
4292	* Note: assuming that within the 2 MB range:
4293	* - GCPhys + PAGE_SIZE = HCPhys + PAGE_SIZE (whole point of this exercise)
4294	* - user space mapping is continuous as well
4295	* - page id (GCPhys) + 1 = page id (GCPhys + PAGE_SIZE)
4296	*/
4297	rc = pgmPhysPageMapByPageID(pVM, idPage, HCPhys, &pv);
4298	AssertLogRelMsg(RT_SUCCESS(rc), ("idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n", idPage, HCPhys, rc));
4299
4300	if (RT_SUCCESS(rc))
4301	{
4302	/*
4303	* Clear the pages.
4304	*/
4305	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatClearLargePage, b);
4306	for (unsigned i = 0; i < _2M/PAGE_SIZE; i++)
4307	{
4308	ASMMemZeroPage(pv);
4309
4310	PPGMPAGE pPage;
4311	rc = pgmPhysGetPageEx(pVM, GCPhys, &pPage);
4312	AssertRC(rc);
4313
4314	Assert(PGM_PAGE_IS_ZERO(pPage));
4315	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatRZPageReplaceZero);
4316	pVM->pgm.s.cZeroPages--;
4317
4318	/*
4319	* Do the PGMPAGE modifications.
4320	*/
4321	pVM->pgm.s.cPrivatePages++;
4322	PGM_PAGE_SET_HCPHYS(pVM, pPage, HCPhys);
4323	PGM_PAGE_SET_PAGEID(pVM, pPage, idPage);
4324	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
4325	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_PDE);
4326	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
4327	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
4328
4329	/* Somewhat dirty assumption that page ids are increasing. */
4330	idPage++;
4331
4332	HCPhys += PAGE_SIZE;
4333	GCPhys += PAGE_SIZE;
4334
4335	pv = (void *)((uintptr_t)pv + PAGE_SIZE);
4336
4337	Log3(("PGMR3PhysAllocateLargePage: idPage=%#x HCPhys=%RGp\n", idPage, HCPhys));
4338	}
4339	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatClearLargePage, b);
4340
4341	/* Flush all TLBs. */
4342	PGM_INVL_ALL_VCPU_TLBS(pVM);
4343	pgmPhysInvalidatePageMapTLB(pVM);
4344	}
4345	pVM->pgm.s.cLargeHandyPages = 0;
4346	}
4347
4348	if (RT_SUCCESS(rc))
4349	{
4350	static uint32_t cTimeOut = 0;
4351	uint64_t u64TimeStampDelta = u64TimeStamp2 - u64TimeStamp1;
4352
4353	if (u64TimeStampDelta > 100)
4354	{
4355	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatLargePageOverflow);
4356	if ( ++cTimeOut > 10
4357	\|\| u64TimeStampDelta > 1000 /* more than one second forces an early retirement from allocating large pages. */)
4358	{
4359	/* If repeated attempts to allocate a large page takes more than 100 ms, then we fall back to normal 4k pages.
4360	* E.g. Vista 64 tries to move memory around, which takes a huge amount of time.
4361	*/
4362	LogRel(("PGMR3PhysAllocateLargePage: allocating large pages takes too long (last attempt %d ms; nr of timeouts %d); DISABLE\n", u64TimeStampDelta, cTimeOut));
4363	PGMSetLargePageUsage(pVM, false);
4364	}
4365	}
4366	else
4367	if (cTimeOut > 0)
4368	cTimeOut--;
4369	}
4370
4371	pgmUnlock(pVM);
4372	return rc;
4373	#else
4374	return VERR_NOT_IMPLEMENTED;
4375	#endif /* PGM_WITH_LARGE_PAGES */
4376	}
4377
4378
4379	/**
4380	* Response to VM_FF_PGM_NEED_HANDY_PAGES and VMMCALLRING3_PGM_ALLOCATE_HANDY_PAGES.
4381	*
4382	* This function will also work the VM_FF_PGM_NO_MEMORY force action flag, to
4383	* signal and clear the out of memory condition. When contracted, this API is
4384	* used to try clear the condition when the user wants to resume.
4385	*
4386	* @returns The following VBox status codes.
4387	* @retval VINF_SUCCESS on success. FFs cleared.
4388	* @retval VINF_EM_NO_MEMORY if we're out of memory. The FF is not cleared in
4389	* this case and it gets accompanied by VM_FF_PGM_NO_MEMORY.
4390	*
4391	* @param pVM Pointer to the VM.
4392	*
4393	* @remarks The VINF_EM_NO_MEMORY status is for the benefit of the FF processing
4394	* in EM.cpp and shouldn't be propagated outside TRPM, HM, EM and
4395	* pgmPhysEnsureHandyPage. There is one exception to this in the \#PF
4396	* handler.
4397	*/
4398	VMMR3DECL(int) PGMR3PhysAllocateHandyPages(PVM pVM)
4399	{
4400	pgmLock(pVM);
4401
4402	/*
4403	* Allocate more pages, noting down the index of the first new page.
4404	*/
4405	uint32_t iClear = pVM->pgm.s.cHandyPages;
4406	AssertMsgReturn(iClear <= RT_ELEMENTS(pVM->pgm.s.aHandyPages), ("%d", iClear), VERR_PGM_HANDY_PAGE_IPE);
4407	Log(("PGMR3PhysAllocateHandyPages: %d -> %d\n", iClear, RT_ELEMENTS(pVM->pgm.s.aHandyPages)));
4408	int rcAlloc = VINF_SUCCESS;
4409	int rcSeed = VINF_SUCCESS;
4410	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
4411	while (rc == VERR_GMM_SEED_ME)
4412	{
4413	void *pvChunk;
4414	rcAlloc = rc = SUPR3PageAlloc(GMM_CHUNK_SIZE >> PAGE_SHIFT, &pvChunk);
4415	if (RT_SUCCESS(rc))
4416	{
4417	rcSeed = rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_SEED_CHUNK, (uintptr_t)pvChunk, NULL);
4418	if (RT_FAILURE(rc))
4419	SUPR3PageFree(pvChunk, GMM_CHUNK_SIZE >> PAGE_SHIFT);
4420	}
4421	if (RT_SUCCESS(rc))
4422	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
4423	}
4424
4425	/* 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) */
4426	if ( rc == VERR_GMM_HIT_VM_ACCOUNT_LIMIT
4427	&& pVM->pgm.s.cHandyPages > 0)
4428	{
4429	/* Still handy pages left, so don't panic. */
4430	rc = VINF_SUCCESS;
4431	}
4432
4433	if (RT_SUCCESS(rc))
4434	{
4435	AssertMsg(rc == VINF_SUCCESS, ("%Rrc\n", rc));
4436	Assert(pVM->pgm.s.cHandyPages > 0);
4437	VM_FF_CLEAR(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
4438	VM_FF_CLEAR(pVM, VM_FF_PGM_NO_MEMORY);
4439
4440	#ifdef VBOX_STRICT
4441	uint32_t i;
4442	for (i = iClear; i < pVM->pgm.s.cHandyPages; i++)
4443	if ( pVM->pgm.s.aHandyPages[i].idPage == NIL_GMM_PAGEID
4444	\|\| pVM->pgm.s.aHandyPages[i].idSharedPage != NIL_GMM_PAGEID
4445	\|\| (pVM->pgm.s.aHandyPages[i].HCPhysGCPhys & PAGE_OFFSET_MASK))
4446	break;
4447	if (i != pVM->pgm.s.cHandyPages)
4448	{
4449	RTAssertMsg1Weak(NULL, __LINE__, __FILE__, __FUNCTION__);
4450	RTAssertMsg2Weak("i=%d iClear=%d cHandyPages=%d\n", i, iClear, pVM->pgm.s.cHandyPages);
4451	for (uint32_t j = iClear; j < pVM->pgm.s.cHandyPages; j++)
4452	RTAssertMsg2Add(("%03d: idPage=%d HCPhysGCPhys=%RHp idSharedPage=%d%\n", j,
4453	pVM->pgm.s.aHandyPages[j].idPage,
4454	pVM->pgm.s.aHandyPages[j].HCPhysGCPhys,
4455	pVM->pgm.s.aHandyPages[j].idSharedPage,
4456	j == i ? " <---" : ""));
4457	RTAssertPanic();
4458	}
4459	#endif
4460	/*
4461	* Clear the pages.
4462	*/
4463	while (iClear < pVM->pgm.s.cHandyPages)
4464	{
4465	PGMMPAGEDESC pPage = &pVM->pgm.s.aHandyPages[iClear];
4466	void *pv;
4467	rc = pgmPhysPageMapByPageID(pVM, pPage->idPage, pPage->HCPhysGCPhys, &pv);
4468	AssertLogRelMsgBreak(RT_SUCCESS(rc),
4469	("%u/%u: idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n",
4470	iClear, pVM->pgm.s.cHandyPages, pPage->idPage, pPage->HCPhysGCPhys, rc));
4471	ASMMemZeroPage(pv);
4472	iClear++;
4473	Log3(("PGMR3PhysAllocateHandyPages: idPage=%#x HCPhys=%RGp\n", pPage->idPage, pPage->HCPhysGCPhys));
4474	}
4475	}
4476	else
4477	{
4478	uint64_t cAllocPages, cMaxPages, cBalloonPages;
4479
4480	/*
4481	* We should never get here unless there is a genuine shortage of
4482	* memory (or some internal error). Flag the error so the VM can be
4483	* suspended ASAP and the user informed. If we're totally out of
4484	* handy pages we will return failure.
4485	*/
4486	/* Report the failure. */
4487	LogRel(("PGM: Failed to procure handy pages; rc=%Rrc rcAlloc=%Rrc rcSeed=%Rrc cHandyPages=%#x\n"
4488	" cAllPages=%#x cPrivatePages=%#x cSharedPages=%#x cZeroPages=%#x\n",
4489	rc, rcAlloc, rcSeed,
4490	pVM->pgm.s.cHandyPages,
4491	pVM->pgm.s.cAllPages,
4492	pVM->pgm.s.cPrivatePages,
4493	pVM->pgm.s.cSharedPages,
4494	pVM->pgm.s.cZeroPages));
4495
4496	if (GMMR3QueryMemoryStats(pVM, &cAllocPages, &cMaxPages, &cBalloonPages) == VINF_SUCCESS)
4497	{
4498	LogRel(("GMM: Statistics:\n"
4499	" Allocated pages: %RX64\n"
4500	" Maximum pages: %RX64\n"
4501	" Ballooned pages: %RX64\n", cAllocPages, cMaxPages, cBalloonPages));
4502	}
4503
4504	if ( rc != VERR_NO_MEMORY
4505	&& rc != VERR_NO_PHYS_MEMORY
4506	&& rc != VERR_LOCK_FAILED)
4507	{
4508	for (uint32_t i = 0; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
4509	{
4510	LogRel(("PGM: aHandyPages[#%#04x] = {.HCPhysGCPhys=%RHp, .idPage=%#08x, .idSharedPage=%#08x}\n",
4511	i, pVM->pgm.s.aHandyPages[i].HCPhysGCPhys, pVM->pgm.s.aHandyPages[i].idPage,
4512	pVM->pgm.s.aHandyPages[i].idSharedPage));
4513	uint32_t const idPage = pVM->pgm.s.aHandyPages[i].idPage;
4514	if (idPage != NIL_GMM_PAGEID)
4515	{
4516	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
4517	pRam;
4518	pRam = pRam->pNextR3)
4519	{
4520	uint32_t const cPages = pRam->cb >> PAGE_SHIFT;
4521	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4522	if (PGM_PAGE_GET_PAGEID(&pRam->aPages[iPage]) == idPage)
4523	LogRel(("PGM: Used by %RGp %R[pgmpage] (%s)\n",
4524	pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pRam->aPages[iPage], pRam->pszDesc));
4525	}
4526	}
4527	}
4528	}
4529
4530	/* Set the FFs and adjust rc. */
4531	VM_FF_SET(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
4532	VM_FF_SET(pVM, VM_FF_PGM_NO_MEMORY);
4533	if ( rc == VERR_NO_MEMORY
4534	\|\| rc == VERR_NO_PHYS_MEMORY
4535	\|\| rc == VERR_LOCK_FAILED)
4536	rc = VINF_EM_NO_MEMORY;
4537	}
4538
4539	pgmUnlock(pVM);
4540	return rc;
4541	}
4542
4543
4544	/**
4545	* Frees the specified RAM page and replaces it with the ZERO page.
4546	*
4547	* This is used by ballooning, remapping MMIO2, RAM reset and state loading.
4548	*
4549	* @param pVM Pointer to the VM.
4550	* @param pReq Pointer to the request.
4551	* @param pcPendingPages Where the number of pages waiting to be freed are
4552	* kept. This will normally be incremented.
4553	* @param pPage Pointer to the page structure.
4554	* @param GCPhys The guest physical address of the page, if applicable.
4555	*
4556	* @remarks The caller must own the PGM lock.
4557	*/
4558	int pgmPhysFreePage(PVM pVM, PGMMFREEPAGESREQ pReq, uint32_t *pcPendingPages, PPGMPAGE pPage, RTGCPHYS GCPhys)
4559	{
4560	/*
4561	* Assert sanity.
4562	*/
4563	PGM_LOCK_ASSERT_OWNER(pVM);
4564	if (RT_UNLIKELY( PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM
4565	&& PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_ROM_SHADOW))
4566	{
4567	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
4568	return VMSetError(pVM, VERR_PGM_PHYS_NOT_RAM, RT_SRC_POS, "GCPhys=%RGp type=%d", GCPhys, PGM_PAGE_GET_TYPE(pPage));
4569	}
4570
4571	/** @todo What about ballooning of large pages??! */
4572	Assert( PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE
4573	&& PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE_DISABLED);
4574
4575	if ( PGM_PAGE_IS_ZERO(pPage)
4576	\|\| PGM_PAGE_IS_BALLOONED(pPage))
4577	return VINF_SUCCESS;
4578
4579	const uint32_t idPage = PGM_PAGE_GET_PAGEID(pPage);
4580	Log3(("pgmPhysFreePage: idPage=%#x GCPhys=%RGp pPage=%R[pgmpage]\n", idPage, GCPhys, pPage));
4581	if (RT_UNLIKELY( idPage == NIL_GMM_PAGEID
4582	\|\| idPage > GMM_PAGEID_LAST
4583	\|\| PGM_PAGE_GET_CHUNKID(pPage) == NIL_GMM_CHUNKID))
4584	{
4585	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
4586	return VMSetError(pVM, VERR_PGM_PHYS_INVALID_PAGE_ID, RT_SRC_POS, "GCPhys=%RGp idPage=%#x", GCPhys, pPage);
4587	}
4588
4589	/* update page count stats. */
4590	if (PGM_PAGE_IS_SHARED(pPage))
4591	pVM->pgm.s.cSharedPages--;
4592	else
4593	pVM->pgm.s.cPrivatePages--;
4594	pVM->pgm.s.cZeroPages++;
4595
4596	/* Deal with write monitored pages. */
4597	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
4598	{
4599	PGM_PAGE_SET_WRITTEN_TO(pVM, pPage);
4600	pVM->pgm.s.cWrittenToPages++;
4601	}
4602
4603	/*
4604	* pPage = ZERO page.
4605	*/
4606	PGM_PAGE_SET_HCPHYS(pVM, pPage, pVM->pgm.s.HCPhysZeroPg);
4607	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
4608	PGM_PAGE_SET_PAGEID(pVM, pPage, NIL_GMM_PAGEID);
4609	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4610	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
4611	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
4612
4613	/* Flush physical page map TLB entry. */
4614	pgmPhysInvalidatePageMapTLBEntry(pVM, GCPhys);
4615
4616	/*
4617	* Make sure it's not in the handy page array.
4618	*/
4619	for (uint32_t i = pVM->pgm.s.cHandyPages; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
4620	{
4621	if (pVM->pgm.s.aHandyPages[i].idPage == idPage)
4622	{
4623	pVM->pgm.s.aHandyPages[i].idPage = NIL_GMM_PAGEID;
4624	break;
4625	}
4626	if (pVM->pgm.s.aHandyPages[i].idSharedPage == idPage)
4627	{
4628	pVM->pgm.s.aHandyPages[i].idSharedPage = NIL_GMM_PAGEID;
4629	break;
4630	}
4631	}
4632
4633	/*
4634	* Push it onto the page array.
4635	*/
4636	uint32_t iPage = *pcPendingPages;
4637	Assert(iPage < PGMPHYS_FREE_PAGE_BATCH_SIZE);
4638	*pcPendingPages += 1;
4639
4640	pReq->aPages[iPage].idPage = idPage;
4641
4642	if (iPage + 1 < PGMPHYS_FREE_PAGE_BATCH_SIZE)
4643	return VINF_SUCCESS;
4644
4645	/*
4646	* Flush the pages.
4647	*/
4648	int rc = GMMR3FreePagesPerform(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE);
4649	if (RT_SUCCESS(rc))
4650	{
4651	GMMR3FreePagesRePrep(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
4652	*pcPendingPages = 0;
4653	}
4654	return rc;
4655	}
4656
4657
4658	/**
4659	* Converts a GC physical address to a HC ring-3 pointer, with some
4660	* additional checks.
4661	*
4662	* @returns VBox status code.
4663	* @retval VINF_SUCCESS on success.
4664	* @retval VINF_PGM_PHYS_TLB_CATCH_WRITE and *ppv set if the page has a write
4665	* access handler of some kind.
4666	* @retval VERR_PGM_PHYS_TLB_CATCH_ALL if the page has a handler catching all
4667	* accesses or is odd in any way.
4668	* @retval VERR_PGM_PHYS_TLB_UNASSIGNED if the page doesn't exist.
4669	*
4670	* @param pVM Pointer to the VM.
4671	* @param GCPhys The GC physical address to convert. Since this is only
4672	* used for filling the REM TLB, the A20 mask must be
4673	* applied before calling this API.
4674	* @param fWritable Whether write access is required.
4675	* @param ppv Where to store the pointer corresponding to GCPhys on
4676	* success.
4677	*/
4678	VMMR3DECL(int) PGMR3PhysTlbGCPhys2Ptr(PVM pVM, RTGCPHYS GCPhys, bool fWritable, void **ppv)
4679	{
4680	pgmLock(pVM);
4681	PGM_A20_ASSERT_MASKED(VMMGetCpu(pVM), GCPhys);
4682
4683	PPGMRAMRANGE pRam;
4684	PPGMPAGE pPage;
4685	int rc = pgmPhysGetPageAndRangeEx(pVM, GCPhys, &pPage, &pRam);
4686	if (RT_SUCCESS(rc))
4687	{
4688	if (PGM_PAGE_IS_BALLOONED(pPage))
4689	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
4690	else if (!PGM_PAGE_HAS_ANY_HANDLERS(pPage))
4691	rc = VINF_SUCCESS;
4692	else
4693	{
4694	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)) /* catches MMIO */
4695	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
4696	else if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
4697	{
4698	/** @todo Handle TLB loads of virtual handlers so ./test.sh can be made to work
4699	* in -norawr0 mode. */
4700	if (fWritable)
4701	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
4702	}
4703	else
4704	{
4705	/* Temporarily disabled physical handler(s), since the recompiler
4706	doesn't get notified when it's reset we'll have to pretend it's
4707	operating normally. */
4708	if (pgmHandlerPhysicalIsAll(pVM, GCPhys))
4709	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
4710	else
4711	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
4712	}
4713	}
4714	if (RT_SUCCESS(rc))
4715	{
4716	int rc2;
4717
4718	/* Make sure what we return is writable. */
4719	if (fWritable)
4720	switch (PGM_PAGE_GET_STATE(pPage))
4721	{
4722	case PGM_PAGE_STATE_ALLOCATED:
4723	break;
4724	case PGM_PAGE_STATE_BALLOONED:
4725	AssertFailed();
4726	break;
4727	case PGM_PAGE_STATE_ZERO:
4728	case PGM_PAGE_STATE_SHARED:
4729	if (rc == VINF_PGM_PHYS_TLB_CATCH_WRITE)
4730	break;
4731	case PGM_PAGE_STATE_WRITE_MONITORED:
4732	rc2 = pgmPhysPageMakeWritable(pVM, pPage, GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK);
4733	AssertLogRelRCReturn(rc2, rc2);
4734	break;
4735	}
4736
4737	/* Get a ring-3 mapping of the address. */
4738	PPGMPAGER3MAPTLBE pTlbe;
4739	rc2 = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
4740	AssertLogRelRCReturn(rc2, rc2);
4741	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
4742	/** @todo mapping/locking hell; this isn't horribly efficient since
4743	* pgmPhysPageLoadIntoTlb will repeat the lookup we've done here. */
4744
4745	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage] ppv=%p\n", GCPhys, rc, pPage, ppv));
4746	}
4747	else
4748	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage]\n", GCPhys, rc, pPage));
4749
4750	/* else: handler catching all access, no pointer returned. */
4751	}
4752	else
4753	rc = VERR_PGM_PHYS_TLB_UNASSIGNED;
4754
4755	pgmUnlock(pVM);
4756	return rc;
4757	}
4758

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PGMPhys.cpp@ 49966

Download in other formats: