PGMPhys.cpp@ 63648

Last change on this file since 63648 was 63560, checked in by vboxsync, 8 years ago
scm: cleaning up todos
Property svn:eol-style set to `native` Property svn:keywords set to `Id Revision`
File size: 167.7 KB

Line
1	/* $Id: PGMPhys.cpp 63560 2016-08-16 14:01:20Z vboxsync $ */
2	/** @file
3	* PGM - Page Manager and Monitor, Physical Memory Addressing.
4	*/
5
6	/*
7	* Copyright (C) 2006-2016 Oracle Corporation
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*/
17
18
19	/*********************************************************************************************************************************
20	* Header Files *
21	*********************************************************************************************************************************/
22	#define LOG_GROUP LOG_GROUP_PGM_PHYS
23	#include <VBox/vmm/pgm.h>
24	#include <VBox/vmm/iem.h>
25	#include <VBox/vmm/iom.h>
26	#include <VBox/vmm/mm.h>
27	#include <VBox/vmm/stam.h>
28	#ifdef VBOX_WITH_REM
29	# include <VBox/vmm/rem.h>
30	#endif
31	#include <VBox/vmm/pdmdev.h>
32	#include "PGMInternal.h"
33	#include <VBox/vmm/vm.h>
34	#include <VBox/vmm/uvm.h>
35	#include "PGMInline.h"
36	#include <VBox/sup.h>
37	#include <VBox/param.h>
38	#include <VBox/err.h>
39	#include <VBox/log.h>
40	#include <iprt/assert.h>
41	#include <iprt/alloc.h>
42	#include <iprt/asm.h>
43	#ifdef VBOX_STRICT
44	# include <iprt/crc.h>
45	#endif
46	#include <iprt/thread.h>
47	#include <iprt/string.h>
48	#include <iprt/system.h>
49
50
51	/*********************************************************************************************************************************
52	* Defined Constants And Macros *
53	*********************************************************************************************************************************/
54	/** The number of pages to free in one batch. */
55	#define PGMPHYS_FREE_PAGE_BATCH_SIZE 128
56
57
58	/*
59	* PGMR3PhysReadU8-64
60	* PGMR3PhysWriteU8-64
61	*/
62	#define PGMPHYSFN_READNAME PGMR3PhysReadU8
63	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU8
64	#define PGMPHYS_DATASIZE 1
65	#define PGMPHYS_DATATYPE uint8_t
66	#include "PGMPhysRWTmpl.h"
67
68	#define PGMPHYSFN_READNAME PGMR3PhysReadU16
69	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU16
70	#define PGMPHYS_DATASIZE 2
71	#define PGMPHYS_DATATYPE uint16_t
72	#include "PGMPhysRWTmpl.h"
73
74	#define PGMPHYSFN_READNAME PGMR3PhysReadU32
75	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU32
76	#define PGMPHYS_DATASIZE 4
77	#define PGMPHYS_DATATYPE uint32_t
78	#include "PGMPhysRWTmpl.h"
79
80	#define PGMPHYSFN_READNAME PGMR3PhysReadU64
81	#define PGMPHYSFN_WRITENAME PGMR3PhysWriteU64
82	#define PGMPHYS_DATASIZE 8
83	#define PGMPHYS_DATATYPE uint64_t
84	#include "PGMPhysRWTmpl.h"
85
86
87	/**
88	* EMT worker for PGMR3PhysReadExternal.
89	*/
90	static DECLCALLBACK(int) pgmR3PhysReadExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, void *pvBuf, size_t cbRead,
91	PGMACCESSORIGIN enmOrigin)
92	{
93	VBOXSTRICTRC rcStrict = PGMPhysRead(pVM, *pGCPhys, pvBuf, cbRead, enmOrigin);
94	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
95	return VINF_SUCCESS;
96	}
97
98
99	/**
100	* Read from physical memory, external users.
101	*
102	* @returns VBox status code.
103	* @retval VINF_SUCCESS.
104	*
105	* @param pVM The cross context VM structure.
106	* @param GCPhys Physical address to read from.
107	* @param pvBuf Where to read into.
108	* @param cbRead How many bytes to read.
109	* @param enmOrigin Who is calling.
110	*
111	* @thread Any but EMTs.
112	*/
113	VMMR3DECL(int) PGMR3PhysReadExternal(PVM pVM, RTGCPHYS GCPhys, void *pvBuf, size_t cbRead, PGMACCESSORIGIN enmOrigin)
114	{
115	VM_ASSERT_OTHER_THREAD(pVM);
116
117	AssertMsgReturn(cbRead > 0, ("don't even think about reading zero bytes!\n"), VINF_SUCCESS);
118	LogFlow(("PGMR3PhysReadExternal: %RGp %d\n", GCPhys, cbRead));
119
120	pgmLock(pVM);
121
122	/*
123	* Copy loop on ram ranges.
124	*/
125	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
126	for (;;)
127	{
128	/* Inside range or not? */
129	if (pRam && GCPhys >= pRam->GCPhys)
130	{
131	/*
132	* Must work our way thru this page by page.
133	*/
134	RTGCPHYS off = GCPhys - pRam->GCPhys;
135	while (off < pRam->cb)
136	{
137	unsigned iPage = off >> PAGE_SHIFT;
138	PPGMPAGE pPage = &pRam->aPages[iPage];
139
140	/*
141	* If the page has an ALL access handler, we'll have to
142	* delegate the job to EMT.
143	*/
144	if ( PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)
145	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
146	{
147	pgmUnlock(pVM);
148
149	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysReadExternalEMT, 5,
150	pVM, &GCPhys, pvBuf, cbRead, enmOrigin);
151	}
152	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
153
154	/*
155	* Simple stuff, go ahead.
156	*/
157	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
158	if (cb > cbRead)
159	cb = cbRead;
160	PGMPAGEMAPLOCK PgMpLck;
161	const void *pvSrc;
162	int rc = pgmPhysGCPhys2CCPtrInternalReadOnly(pVM, pPage, pRam->GCPhys + off, &pvSrc, &PgMpLck);
163	if (RT_SUCCESS(rc))
164	{
165	memcpy(pvBuf, pvSrc, cb);
166	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
167	}
168	else
169	{
170	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternalReadOnly failed on %RGp / %R[pgmpage] -> %Rrc\n",
171	pRam->GCPhys + off, pPage, rc));
172	memset(pvBuf, 0xff, cb);
173	}
174
175	/* next page */
176	if (cb >= cbRead)
177	{
178	pgmUnlock(pVM);
179	return VINF_SUCCESS;
180	}
181	cbRead -= cb;
182	off += cb;
183	GCPhys += cb;
184	pvBuf = (char *)pvBuf + cb;
185	} /* walk pages in ram range. */
186	}
187	else
188	{
189	LogFlow(("PGMPhysRead: Unassigned %RGp size=%u\n", GCPhys, cbRead));
190
191	/*
192	* Unassigned address space.
193	*/
194	size_t cb = pRam ? pRam->GCPhys - GCPhys : ~(size_t)0;
195	if (cb >= cbRead)
196	{
197	memset(pvBuf, 0xff, cbRead);
198	break;
199	}
200	memset(pvBuf, 0xff, cb);
201
202	cbRead -= cb;
203	pvBuf = (char *)pvBuf + cb;
204	GCPhys += cb;
205	}
206
207	/* Advance range if necessary. */
208	while (pRam && GCPhys > pRam->GCPhysLast)
209	pRam = pRam->CTX_SUFF(pNext);
210	} /* Ram range walk */
211
212	pgmUnlock(pVM);
213
214	return VINF_SUCCESS;
215	}
216
217
218	/**
219	* EMT worker for PGMR3PhysWriteExternal.
220	*/
221	static DECLCALLBACK(int) pgmR3PhysWriteExternalEMT(PVM pVM, PRTGCPHYS pGCPhys, const void *pvBuf, size_t cbWrite,
222	PGMACCESSORIGIN enmOrigin)
223	{
224	/** @todo VERR_EM_NO_MEMORY */
225	VBOXSTRICTRC rcStrict = PGMPhysWrite(pVM, *pGCPhys, pvBuf, cbWrite, enmOrigin);
226	AssertMsg(rcStrict == VINF_SUCCESS, ("%Rrc\n", VBOXSTRICTRC_VAL(rcStrict))); NOREF(rcStrict);
227	return VINF_SUCCESS;
228	}
229
230
231	/**
232	* Write to physical memory, external users.
233	*
234	* @returns VBox status code.
235	* @retval VINF_SUCCESS.
236	* @retval VERR_EM_NO_MEMORY.
237	*
238	* @param pVM The cross context VM structure.
239	* @param GCPhys Physical address to write to.
240	* @param pvBuf What to write.
241	* @param cbWrite How many bytes to write.
242	* @param enmOrigin Who is calling.
243	*
244	* @thread Any but EMTs.
245	*/
246	VMMDECL(int) PGMR3PhysWriteExternal(PVM pVM, RTGCPHYS GCPhys, const void *pvBuf, size_t cbWrite, PGMACCESSORIGIN enmOrigin)
247	{
248	VM_ASSERT_OTHER_THREAD(pVM);
249
250	AssertMsg(!pVM->pgm.s.fNoMorePhysWrites,
251	("Calling PGMR3PhysWriteExternal after pgmR3Save()! GCPhys=%RGp cbWrite=%#x enmOrigin=%d\n",
252	GCPhys, cbWrite, enmOrigin));
253	AssertMsgReturn(cbWrite > 0, ("don't even think about writing zero bytes!\n"), VINF_SUCCESS);
254	LogFlow(("PGMR3PhysWriteExternal: %RGp %d\n", GCPhys, cbWrite));
255
256	pgmLock(pVM);
257
258	/*
259	* Copy loop on ram ranges, stop when we hit something difficult.
260	*/
261	PPGMRAMRANGE pRam = pgmPhysGetRangeAtOrAbove(pVM, GCPhys);
262	for (;;)
263	{
264	/* Inside range or not? */
265	if (pRam && GCPhys >= pRam->GCPhys)
266	{
267	/*
268	* Must work our way thru this page by page.
269	*/
270	RTGCPTR off = GCPhys - pRam->GCPhys;
271	while (off < pRam->cb)
272	{
273	RTGCPTR iPage = off >> PAGE_SHIFT;
274	PPGMPAGE pPage = &pRam->aPages[iPage];
275
276	/*
277	* Is the page problematic, we have to do the work on the EMT.
278	*
279	* Allocating writable pages and access handlers are
280	* problematic, write monitored pages are simple and can be
281	* dealt with here.
282	*/
283	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
284	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
285	\|\| PGM_PAGE_IS_SPECIAL_ALIAS_MMIO(pPage))
286	{
287	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
288	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
289	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage);
290	else
291	{
292	pgmUnlock(pVM);
293
294	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysWriteExternalEMT, 5,
295	pVM, &GCPhys, pvBuf, cbWrite, enmOrigin);
296	}
297	}
298	Assert(!PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage));
299
300	/*
301	* Simple stuff, go ahead.
302	*/
303	size_t cb = PAGE_SIZE - (off & PAGE_OFFSET_MASK);
304	if (cb > cbWrite)
305	cb = cbWrite;
306	PGMPAGEMAPLOCK PgMpLck;
307	void *pvDst;
308	int rc = pgmPhysGCPhys2CCPtrInternal(pVM, pPage, pRam->GCPhys + off, &pvDst, &PgMpLck);
309	if (RT_SUCCESS(rc))
310	{
311	memcpy(pvDst, pvBuf, cb);
312	pgmPhysReleaseInternalPageMappingLock(pVM, &PgMpLck);
313	}
314	else
315	AssertLogRelMsgFailed(("pgmPhysGCPhys2CCPtrInternal failed on %RGp / %R[pgmpage] -> %Rrc\n",
316	pRam->GCPhys + off, pPage, rc));
317
318	/* next page */
319	if (cb >= cbWrite)
320	{
321	pgmUnlock(pVM);
322	return VINF_SUCCESS;
323	}
324
325	cbWrite -= cb;
326	off += cb;
327	GCPhys += cb;
328	pvBuf = (const char *)pvBuf + cb;
329	} /* walk pages in ram range */
330	}
331	else
332	{
333	/*
334	* Unassigned address space, skip it.
335	*/
336	if (!pRam)
337	break;
338	size_t cb = pRam->GCPhys - GCPhys;
339	if (cb >= cbWrite)
340	break;
341	cbWrite -= cb;
342	pvBuf = (const char *)pvBuf + cb;
343	GCPhys += cb;
344	}
345
346	/* Advance range if necessary. */
347	while (pRam && GCPhys > pRam->GCPhysLast)
348	pRam = pRam->CTX_SUFF(pNext);
349	} /* Ram range walk */
350
351	pgmUnlock(pVM);
352	return VINF_SUCCESS;
353	}
354
355
356	/**
357	* VMR3ReqCall worker for PGMR3PhysGCPhys2CCPtrExternal to make pages writable.
358	*
359	* @returns see PGMR3PhysGCPhys2CCPtrExternal
360	* @param pVM The cross context VM structure.
361	* @param pGCPhys Pointer to the guest physical address.
362	* @param ppv Where to store the mapping address.
363	* @param pLock Where to store the lock.
364	*/
365	static DECLCALLBACK(int) pgmR3PhysGCPhys2CCPtrDelegated(PVM pVM, PRTGCPHYS pGCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
366	{
367	/*
368	* Just hand it to PGMPhysGCPhys2CCPtr and check that it's not a page with
369	* an access handler after it succeeds.
370	*/
371	int rc = pgmLock(pVM);
372	AssertRCReturn(rc, rc);
373
374	rc = PGMPhysGCPhys2CCPtr(pVM, *pGCPhys, ppv, pLock);
375	if (RT_SUCCESS(rc))
376	{
377	PPGMPAGEMAPTLBE pTlbe;
378	int rc2 = pgmPhysPageQueryTlbe(pVM, *pGCPhys, &pTlbe);
379	AssertFatalRC(rc2);
380	PPGMPAGE pPage = pTlbe->pPage;
381	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
382	{
383	PGMPhysReleasePageMappingLock(pVM, pLock);
384	rc = VERR_PGM_PHYS_PAGE_RESERVED;
385	}
386	else if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
387	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
388	\|\| pgmPoolIsDirtyPage(pVM, *pGCPhys)
389	#endif
390	)
391	{
392	/* We must flush any corresponding pgm pool page here, otherwise we'll
393	* not be informed about writes and keep bogus gst->shw mappings around.
394	*/
395	pgmPoolFlushPageByGCPhys(pVM, *pGCPhys);
396	Assert(!PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage));
397	/** @todo r=bird: return VERR_PGM_PHYS_PAGE_RESERVED here if it still has
398	* active handlers, see the PGMR3PhysGCPhys2CCPtrExternal docs. */
399	}
400	}
401
402	pgmUnlock(pVM);
403	return rc;
404	}
405
406
407	/**
408	* Requests the mapping of a guest page into ring-3, external threads.
409	*
410	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
411	* release it.
412	*
413	* This API will assume your intention is to write to the page, and will
414	* therefore replace shared and zero pages. If you do not intend to modify the
415	* page, use the PGMR3PhysGCPhys2CCPtrReadOnlyExternal() API.
416	*
417	* @returns VBox status code.
418	* @retval VINF_SUCCESS on success.
419	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
420	* backing or if the page has any active access handlers. The caller
421	* must fall back on using PGMR3PhysWriteExternal.
422	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
423	*
424	* @param pVM The cross context VM structure.
425	* @param GCPhys The guest physical address of the page that should be mapped.
426	* @param ppv Where to store the address corresponding to GCPhys.
427	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
428	*
429	* @remark Avoid calling this API from within critical sections (other than the
430	* PGM one) because of the deadlock risk when we have to delegating the
431	* task to an EMT.
432	* @thread Any.
433	*/
434	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrExternal(PVM pVM, RTGCPHYS GCPhys, void **ppv, PPGMPAGEMAPLOCK pLock)
435	{
436	AssertPtr(ppv);
437	AssertPtr(pLock);
438
439	Assert(VM_IS_EMT(pVM) \|\| !PGMIsLockOwner(pVM));
440
441	int rc = pgmLock(pVM);
442	AssertRCReturn(rc, rc);
443
444	/*
445	* Query the Physical TLB entry for the page (may fail).
446	*/
447	PPGMPAGEMAPTLBE pTlbe;
448	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
449	if (RT_SUCCESS(rc))
450	{
451	PPGMPAGE pPage = pTlbe->pPage;
452	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
453	rc = VERR_PGM_PHYS_PAGE_RESERVED;
454	else
455	{
456	/*
457	* If the page is shared, the zero page, or being write monitored
458	* it must be converted to an page that's writable if possible.
459	* We can only deal with write monitored pages here, the rest have
460	* to be on an EMT.
461	*/
462	if ( PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
463	\|\| PGM_PAGE_GET_STATE(pPage) != PGM_PAGE_STATE_ALLOCATED
464	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
465	\|\| pgmPoolIsDirtyPage(pVM, GCPhys)
466	#endif
467	)
468	{
469	if ( PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED
470	&& !PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage)
471	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
472	&& !pgmPoolIsDirtyPage(pVM, GCPhys)
473	#endif
474	)
475	pgmPhysPageMakeWriteMonitoredWritable(pVM, pPage);
476	else
477	{
478	pgmUnlock(pVM);
479
480	return VMR3ReqPriorityCallWait(pVM, VMCPUID_ANY, (PFNRT)pgmR3PhysGCPhys2CCPtrDelegated, 4,
481	pVM, &GCPhys, ppv, pLock);
482	}
483	}
484
485	/*
486	* Now, just perform the locking and calculate the return address.
487	*/
488	PPGMPAGEMAP pMap = pTlbe->pMap;
489	if (pMap)
490	pMap->cRefs++;
491
492	unsigned cLocks = PGM_PAGE_GET_WRITE_LOCKS(pPage);
493	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
494	{
495	if (cLocks == 0)
496	pVM->pgm.s.cWriteLockedPages++;
497	PGM_PAGE_INC_WRITE_LOCKS(pPage);
498	}
499	else if (cLocks != PGM_PAGE_GET_WRITE_LOCKS(pPage))
500	{
501	PGM_PAGE_INC_WRITE_LOCKS(pPage);
502	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent write locked state!\n", GCPhys, pPage));
503	if (pMap)
504	pMap->cRefs++; /* Extra ref to prevent it from going away. */
505	}
506
507	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
508	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_WRITE;
509	pLock->pvMap = pMap;
510	}
511	}
512
513	pgmUnlock(pVM);
514	return rc;
515	}
516
517
518	/**
519	* Requests the mapping of a guest page into ring-3, external threads.
520	*
521	* When you're done with the page, call PGMPhysReleasePageMappingLock() ASAP to
522	* release it.
523	*
524	* @returns VBox status code.
525	* @retval VINF_SUCCESS on success.
526	* @retval VERR_PGM_PHYS_PAGE_RESERVED it it's a valid page but has no physical
527	* backing or if the page as an active ALL access handler. The caller
528	* must fall back on using PGMPhysRead.
529	* @retval VERR_PGM_INVALID_GC_PHYSICAL_ADDRESS if it's not a valid physical address.
530	*
531	* @param pVM The cross context VM structure.
532	* @param GCPhys The guest physical address of the page that should be mapped.
533	* @param ppv Where to store the address corresponding to GCPhys.
534	* @param pLock Where to store the lock information that PGMPhysReleasePageMappingLock needs.
535	*
536	* @remark Avoid calling this API from within critical sections (other than
537	* the PGM one) because of the deadlock risk.
538	* @thread Any.
539	*/
540	VMMR3DECL(int) PGMR3PhysGCPhys2CCPtrReadOnlyExternal(PVM pVM, RTGCPHYS GCPhys, void const **ppv, PPGMPAGEMAPLOCK pLock)
541	{
542	int rc = pgmLock(pVM);
543	AssertRCReturn(rc, rc);
544
545	/*
546	* Query the Physical TLB entry for the page (may fail).
547	*/
548	PPGMPAGEMAPTLBE pTlbe;
549	rc = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
550	if (RT_SUCCESS(rc))
551	{
552	PPGMPAGE pPage = pTlbe->pPage;
553	#if 1
554	/* MMIO pages doesn't have any readable backing. */
555	if (PGM_PAGE_IS_MMIO_OR_SPECIAL_ALIAS(pPage))
556	rc = VERR_PGM_PHYS_PAGE_RESERVED;
557	#else
558	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage))
559	rc = VERR_PGM_PHYS_PAGE_RESERVED;
560	#endif
561	else
562	{
563	/*
564	* Now, just perform the locking and calculate the return address.
565	*/
566	PPGMPAGEMAP pMap = pTlbe->pMap;
567	if (pMap)
568	pMap->cRefs++;
569
570	unsigned cLocks = PGM_PAGE_GET_READ_LOCKS(pPage);
571	if (RT_LIKELY(cLocks < PGM_PAGE_MAX_LOCKS - 1))
572	{
573	if (cLocks == 0)
574	pVM->pgm.s.cReadLockedPages++;
575	PGM_PAGE_INC_READ_LOCKS(pPage);
576	}
577	else if (cLocks != PGM_PAGE_GET_READ_LOCKS(pPage))
578	{
579	PGM_PAGE_INC_READ_LOCKS(pPage);
580	AssertMsgFailed(("%RGp / %R[pgmpage] is entering permanent readonly locked state!\n", GCPhys, pPage));
581	if (pMap)
582	pMap->cRefs++; /* Extra ref to prevent it from going away. */
583	}
584
585	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
586	pLock->uPageAndType = (uintptr_t)pPage \| PGMPAGEMAPLOCK_TYPE_READ;
587	pLock->pvMap = pMap;
588	}
589	}
590
591	pgmUnlock(pVM);
592	return rc;
593	}
594
595
596	#define MAKE_LEAF(a_pNode) \
597	do { \
598	(a_pNode)->pLeftR3 = NIL_RTR3PTR; \
599	(a_pNode)->pRightR3 = NIL_RTR3PTR; \
600	(a_pNode)->pLeftR0 = NIL_RTR0PTR; \
601	(a_pNode)->pRightR0 = NIL_RTR0PTR; \
602	(a_pNode)->pLeftRC = NIL_RTRCPTR; \
603	(a_pNode)->pRightRC = NIL_RTRCPTR; \
604	} while (0)
605
606	#define INSERT_LEFT(a_pParent, a_pNode) \
607	do { \
608	(a_pParent)->pLeftR3 = (a_pNode); \
609	(a_pParent)->pLeftR0 = (a_pNode)->pSelfR0; \
610	(a_pParent)->pLeftRC = (a_pNode)->pSelfRC; \
611	} while (0)
612	#define INSERT_RIGHT(a_pParent, a_pNode) \
613	do { \
614	(a_pParent)->pRightR3 = (a_pNode); \
615	(a_pParent)->pRightR0 = (a_pNode)->pSelfR0; \
616	(a_pParent)->pRightRC = (a_pNode)->pSelfRC; \
617	} while (0)
618
619
620	/**
621	* Recursive tree builder.
622	*
623	* @param ppRam Pointer to the iterator variable.
624	* @param iDepth The current depth. Inserts a leaf node if 0.
625	*/
626	static PPGMRAMRANGE pgmR3PhysRebuildRamRangeSearchTreesRecursively(PPGMRAMRANGE *ppRam, int iDepth)
627	{
628	PPGMRAMRANGE pRam;
629	if (iDepth <= 0)
630	{
631	/*
632	* Leaf node.
633	*/
634	pRam = *ppRam;
635	if (pRam)
636	{
637	*ppRam = pRam->pNextR3;
638	MAKE_LEAF(pRam);
639	}
640	}
641	else
642	{
643
644	/*
645	* Intermediate node.
646	*/
647	PPGMRAMRANGE pLeft = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
648
649	pRam = *ppRam;
650	if (!pRam)
651	return pLeft;
652	*ppRam = pRam->pNextR3;
653	MAKE_LEAF(pRam);
654	INSERT_LEFT(pRam, pLeft);
655
656	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(ppRam, iDepth - 1);
657	if (pRight)
658	INSERT_RIGHT(pRam, pRight);
659	}
660	return pRam;
661	}
662
663
664	/**
665	* Rebuilds the RAM range search trees.
666	*
667	* @param pVM The cross context VM structure.
668	*/
669	static void pgmR3PhysRebuildRamRangeSearchTrees(PVM pVM)
670	{
671
672	/*
673	* Create the reasonably balanced tree in a sequential fashion.
674	* For simplicity (laziness) we use standard recursion here.
675	*/
676	int iDepth = 0;
677	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
678	PPGMRAMRANGE pRoot = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, 0);
679	while (pRam)
680	{
681	PPGMRAMRANGE pLeft = pRoot;
682
683	pRoot = pRam;
684	pRam = pRam->pNextR3;
685	MAKE_LEAF(pRoot);
686	INSERT_LEFT(pRoot, pLeft);
687
688	PPGMRAMRANGE pRight = pgmR3PhysRebuildRamRangeSearchTreesRecursively(&pRam, iDepth);
689	if (pRight)
690	INSERT_RIGHT(pRoot, pRight);
691	/** @todo else: rotate the tree. */
692
693	iDepth++;
694	}
695
696	pVM->pgm.s.pRamRangeTreeR3 = pRoot;
697	pVM->pgm.s.pRamRangeTreeR0 = pRoot ? pRoot->pSelfR0 : NIL_RTR0PTR;
698	pVM->pgm.s.pRamRangeTreeRC = pRoot ? pRoot->pSelfRC : NIL_RTRCPTR;
699
700	#ifdef VBOX_STRICT
701	/*
702	* Verify that the above code works.
703	*/
704	unsigned cRanges = 0;
705	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
706	cRanges++;
707	Assert(cRanges > 0);
708
709	unsigned cMaxDepth = ASMBitLastSetU32(cRanges);
710	if ((1U << cMaxDepth) < cRanges)
711	cMaxDepth++;
712
713	for (pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
714	{
715	unsigned cDepth = 0;
716	PPGMRAMRANGE pRam2 = pVM->pgm.s.pRamRangeTreeR3;
717	for (;;)
718	{
719	if (pRam == pRam2)
720	break;
721	Assert(pRam2);
722	if (pRam->GCPhys < pRam2->GCPhys)
723	pRam2 = pRam2->pLeftR3;
724	else
725	pRam2 = pRam2->pRightR3;
726	}
727	AssertMsg(cDepth <= cMaxDepth, ("cDepth=%d cMaxDepth=%d\n", cDepth, cMaxDepth));
728	}
729	#endif /* VBOX_STRICT */
730	}
731
732	#undef MAKE_LEAF
733	#undef INSERT_LEFT
734	#undef INSERT_RIGHT
735
736	/**
737	* Relinks the RAM ranges using the pSelfRC and pSelfR0 pointers.
738	*
739	* Called when anything was relocated.
740	*
741	* @param pVM The cross context VM structure.
742	*/
743	void pgmR3PhysRelinkRamRanges(PVM pVM)
744	{
745	PPGMRAMRANGE pCur;
746
747	#ifdef VBOX_STRICT
748	for (pCur = pVM->pgm.s.pRamRangesXR3; pCur; pCur = pCur->pNextR3)
749	{
750	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfR0 == MMHyperCCToR0(pVM, pCur));
751	Assert((pCur->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pCur->pSelfRC == MMHyperCCToRC(pVM, pCur));
752	Assert((pCur->GCPhys & PAGE_OFFSET_MASK) == 0);
753	Assert((pCur->GCPhysLast & PAGE_OFFSET_MASK) == PAGE_OFFSET_MASK);
754	Assert((pCur->cb & PAGE_OFFSET_MASK) == 0);
755	Assert(pCur->cb == pCur->GCPhysLast - pCur->GCPhys + 1);
756	for (PPGMRAMRANGE pCur2 = pVM->pgm.s.pRamRangesXR3; pCur2; pCur2 = pCur2->pNextR3)
757	Assert( pCur2 == pCur
758	\|\| strcmp(pCur2->pszDesc, pCur->pszDesc)); /** @todo fix MMIO ranges!! */
759	}
760	#endif
761
762	pCur = pVM->pgm.s.pRamRangesXR3;
763	if (pCur)
764	{
765	pVM->pgm.s.pRamRangesXR0 = pCur->pSelfR0;
766	pVM->pgm.s.pRamRangesXRC = pCur->pSelfRC;
767
768	for (; pCur->pNextR3; pCur = pCur->pNextR3)
769	{
770	pCur->pNextR0 = pCur->pNextR3->pSelfR0;
771	pCur->pNextRC = pCur->pNextR3->pSelfRC;
772	}
773
774	Assert(pCur->pNextR0 == NIL_RTR0PTR);
775	Assert(pCur->pNextRC == NIL_RTRCPTR);
776	}
777	else
778	{
779	Assert(pVM->pgm.s.pRamRangesXR0 == NIL_RTR0PTR);
780	Assert(pVM->pgm.s.pRamRangesXRC == NIL_RTRCPTR);
781	}
782	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
783
784	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
785	}
786
787
788	/**
789	* Links a new RAM range into the list.
790	*
791	* @param pVM The cross context VM structure.
792	* @param pNew Pointer to the new list entry.
793	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
794	*/
795	static void pgmR3PhysLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, PPGMRAMRANGE pPrev)
796	{
797	AssertMsg(pNew->pszDesc, ("%RGp-%RGp\n", pNew->GCPhys, pNew->GCPhysLast));
798	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfR0 == MMHyperCCToR0(pVM, pNew));
799	Assert((pNew->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pNew->pSelfRC == MMHyperCCToRC(pVM, pNew));
800
801	pgmLock(pVM);
802
803	PPGMRAMRANGE pRam = pPrev ? pPrev->pNextR3 : pVM->pgm.s.pRamRangesXR3;
804	pNew->pNextR3 = pRam;
805	pNew->pNextR0 = pRam ? pRam->pSelfR0 : NIL_RTR0PTR;
806	pNew->pNextRC = pRam ? pRam->pSelfRC : NIL_RTRCPTR;
807
808	if (pPrev)
809	{
810	pPrev->pNextR3 = pNew;
811	pPrev->pNextR0 = pNew->pSelfR0;
812	pPrev->pNextRC = pNew->pSelfRC;
813	}
814	else
815	{
816	pVM->pgm.s.pRamRangesXR3 = pNew;
817	pVM->pgm.s.pRamRangesXR0 = pNew->pSelfR0;
818	pVM->pgm.s.pRamRangesXRC = pNew->pSelfRC;
819	}
820	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
821
822	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
823	pgmUnlock(pVM);
824	}
825
826
827	/**
828	* Unlink an existing RAM range from the list.
829	*
830	* @param pVM The cross context VM structure.
831	* @param pRam Pointer to the new list entry.
832	* @param pPrev Pointer to the previous list entry. If NULL, insert as head.
833	*/
834	static void pgmR3PhysUnlinkRamRange2(PVM pVM, PPGMRAMRANGE pRam, PPGMRAMRANGE pPrev)
835	{
836	Assert(pPrev ? pPrev->pNextR3 == pRam : pVM->pgm.s.pRamRangesXR3 == pRam);
837	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfR0 == MMHyperCCToR0(pVM, pRam));
838	Assert((pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING) \|\| pRam->pSelfRC == MMHyperCCToRC(pVM, pRam));
839
840	pgmLock(pVM);
841
842	PPGMRAMRANGE pNext = pRam->pNextR3;
843	if (pPrev)
844	{
845	pPrev->pNextR3 = pNext;
846	pPrev->pNextR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
847	pPrev->pNextRC = pNext ? pNext->pSelfRC : NIL_RTRCPTR;
848	}
849	else
850	{
851	Assert(pVM->pgm.s.pRamRangesXR3 == pRam);
852	pVM->pgm.s.pRamRangesXR3 = pNext;
853	pVM->pgm.s.pRamRangesXR0 = pNext ? pNext->pSelfR0 : NIL_RTR0PTR;
854	pVM->pgm.s.pRamRangesXRC = pNext ? pNext->pSelfRC : NIL_RTRCPTR;
855	}
856	ASMAtomicIncU32(&pVM->pgm.s.idRamRangesGen);
857
858	pgmR3PhysRebuildRamRangeSearchTrees(pVM);
859	pgmUnlock(pVM);
860	}
861
862
863	/**
864	* Unlink an existing RAM range from the list.
865	*
866	* @param pVM The cross context VM structure.
867	* @param pRam Pointer to the new list entry.
868	*/
869	static void pgmR3PhysUnlinkRamRange(PVM pVM, PPGMRAMRANGE pRam)
870	{
871	pgmLock(pVM);
872
873	/* find prev. */
874	PPGMRAMRANGE pPrev = NULL;
875	PPGMRAMRANGE pCur = pVM->pgm.s.pRamRangesXR3;
876	while (pCur != pRam)
877	{
878	pPrev = pCur;
879	pCur = pCur->pNextR3;
880	}
881	AssertFatal(pCur);
882
883	pgmR3PhysUnlinkRamRange2(pVM, pRam, pPrev);
884	pgmUnlock(pVM);
885	}
886
887
888	/**
889	* Frees a range of pages, replacing them with ZERO pages of the specified type.
890	*
891	* @returns VBox status code.
892	* @param pVM The cross context VM structure.
893	* @param pRam The RAM range in which the pages resides.
894	* @param GCPhys The address of the first page.
895	* @param GCPhysLast The address of the last page.
896	* @param uType The page type to replace then with.
897	*/
898	static int pgmR3PhysFreePageRange(PVM pVM, PPGMRAMRANGE pRam, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast, uint8_t uType)
899	{
900	PGM_LOCK_ASSERT_OWNER(pVM);
901	uint32_t cPendingPages = 0;
902	PGMMFREEPAGESREQ pReq;
903	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
904	AssertLogRelRCReturn(rc, rc);
905
906	/* Iterate the pages. */
907	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
908	uint32_t cPagesLeft = ((GCPhysLast - GCPhys) >> PAGE_SHIFT) + 1;
909	while (cPagesLeft-- > 0)
910	{
911	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys);
912	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
913
914	PGM_PAGE_SET_TYPE(pVM, pPageDst, uType);
915
916	GCPhys += PAGE_SIZE;
917	pPageDst++;
918	}
919
920	if (cPendingPages)
921	{
922	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
923	AssertLogRelRCReturn(rc, rc);
924	}
925	GMMR3FreePagesCleanup(pReq);
926
927	return rc;
928	}
929
930	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
931
932	/**
933	* Rendezvous callback used by PGMR3ChangeMemBalloon that changes the memory balloon size
934	*
935	* This is only called on one of the EMTs while the other ones are waiting for
936	* it to complete this function.
937	*
938	* @returns VINF_SUCCESS (VBox strict status code).
939	* @param pVM The cross context VM structure.
940	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
941	* @param pvUser User parameter
942	*/
943	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysChangeMemBalloonRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
944	{
945	uintptr_t paUser = (uintptr_t )pvUser;
946	bool fInflate = !!paUser[0];
947	unsigned cPages = paUser[1];
948	RTGCPHYS paPhysPage = (RTGCPHYS )paUser[2];
949	uint32_t cPendingPages = 0;
950	PGMMFREEPAGESREQ pReq;
951	int rc;
952
953	Log(("pgmR3PhysChangeMemBalloonRendezvous: %s %x pages\n", (fInflate) ? "inflate" : "deflate", cPages));
954	pgmLock(pVM);
955
956	if (fInflate)
957	{
958	/* Flush the PGM pool cache as we might have stale references to pages that we just freed. */
959	pgmR3PoolClearAllRendezvous(pVM, pVCpu, NULL);
960
961	/* Replace pages with ZERO pages. */
962	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
963	if (RT_FAILURE(rc))
964	{
965	pgmUnlock(pVM);
966	AssertLogRelRC(rc);
967	return rc;
968	}
969
970	/* Iterate the pages. */
971	for (unsigned i = 0; i < cPages; i++)
972	{
973	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
974	if ( pPage == NULL
975	\|\| PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM)
976	{
977	Log(("pgmR3PhysChangeMemBalloonRendezvous: invalid physical page %RGp pPage->u3Type=%d\n", paPhysPage[i], pPage ? PGM_PAGE_GET_TYPE(pPage) : 0));
978	break;
979	}
980
981	LogFlow(("balloon page: %RGp\n", paPhysPage[i]));
982
983	/* Flush the shadow PT if this page was previously used as a guest page table. */
984	pgmPoolFlushPageByGCPhys(pVM, paPhysPage[i]);
985
986	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, paPhysPage[i]);
987	if (RT_FAILURE(rc))
988	{
989	pgmUnlock(pVM);
990	AssertLogRelRC(rc);
991	return rc;
992	}
993	Assert(PGM_PAGE_IS_ZERO(pPage));
994	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_BALLOONED);
995	}
996
997	if (cPendingPages)
998	{
999	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
1000	if (RT_FAILURE(rc))
1001	{
1002	pgmUnlock(pVM);
1003	AssertLogRelRC(rc);
1004	return rc;
1005	}
1006	}
1007	GMMR3FreePagesCleanup(pReq);
1008	}
1009	else
1010	{
1011	/* Iterate the pages. */
1012	for (unsigned i = 0; i < cPages; i++)
1013	{
1014	PPGMPAGE pPage = pgmPhysGetPage(pVM, paPhysPage[i]);
1015	AssertBreak(pPage && PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM);
1016
1017	LogFlow(("Free ballooned page: %RGp\n", paPhysPage[i]));
1018
1019	Assert(PGM_PAGE_IS_BALLOONED(pPage));
1020
1021	/* Change back to zero page. */
1022	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
1023	}
1024
1025	/* Note that we currently do not map any ballooned pages in our shadow page tables, so no need to flush the pgm pool. */
1026	}
1027
1028	/* Notify GMM about the balloon change. */
1029	rc = GMMR3BalloonedPages(pVM, (fInflate) ? GMMBALLOONACTION_INFLATE : GMMBALLOONACTION_DEFLATE, cPages);
1030	if (RT_SUCCESS(rc))
1031	{
1032	if (!fInflate)
1033	{
1034	Assert(pVM->pgm.s.cBalloonedPages >= cPages);
1035	pVM->pgm.s.cBalloonedPages -= cPages;
1036	}
1037	else
1038	pVM->pgm.s.cBalloonedPages += cPages;
1039	}
1040
1041	pgmUnlock(pVM);
1042
1043	/* Flush the recompiler's TLB as well. */
1044	for (VMCPUID i = 0; i < pVM->cCpus; i++)
1045	CPUMSetChangedFlags(&pVM->aCpus[i], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1046
1047	AssertLogRelRC(rc);
1048	return rc;
1049	}
1050
1051
1052	/**
1053	* Frees a range of ram pages, replacing them with ZERO pages; helper for PGMR3PhysFreeRamPages
1054	*
1055	* @returns VBox status code.
1056	* @param pVM The cross context VM structure.
1057	* @param fInflate Inflate or deflate memory balloon
1058	* @param cPages Number of pages to free
1059	* @param paPhysPage Array of guest physical addresses
1060	*/
1061	static DECLCALLBACK(void) pgmR3PhysChangeMemBalloonHelper(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1062	{
1063	uintptr_t paUser[3];
1064
1065	paUser[0] = fInflate;
1066	paUser[1] = cPages;
1067	paUser[2] = (uintptr_t)paPhysPage;
1068	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1069	AssertRC(rc);
1070
1071	/* Made a copy in PGMR3PhysFreeRamPages; free it here. */
1072	RTMemFree(paPhysPage);
1073	}
1074
1075	#endif /* 64-bit host && (Windows \|\| Solaris \|\| Linux \|\| FreeBSD) */
1076
1077	/**
1078	* Inflate or deflate a memory balloon
1079	*
1080	* @returns VBox status code.
1081	* @param pVM The cross context VM structure.
1082	* @param fInflate Inflate or deflate memory balloon
1083	* @param cPages Number of pages to free
1084	* @param paPhysPage Array of guest physical addresses
1085	*/
1086	VMMR3DECL(int) PGMR3PhysChangeMemBalloon(PVM pVM, bool fInflate, unsigned cPages, RTGCPHYS *paPhysPage)
1087	{
1088	/* This must match GMMR0Init; currently we only support memory ballooning on all 64-bit hosts except Mac OS X */
1089	#if HC_ARCH_BITS == 64 && (defined(RT_OS_WINDOWS) \|\| defined(RT_OS_SOLARIS) \|\| defined(RT_OS_LINUX) \|\| defined(RT_OS_FREEBSD))
1090	int rc;
1091
1092	/* Older additions (ancient non-functioning balloon code) pass wrong physical addresses. */
1093	AssertReturn(!(paPhysPage[0] & 0xfff), VERR_INVALID_PARAMETER);
1094
1095	/* We own the IOM lock here and could cause a deadlock by waiting for another VCPU that is blocking on the IOM lock.
1096	* In the SMP case we post a request packet to postpone the job.
1097	*/
1098	if (pVM->cCpus > 1)
1099	{
1100	unsigned cbPhysPage = cPages * sizeof(paPhysPage[0]);
1101	RTGCPHYS paPhysPageCopy = (RTGCPHYS )RTMemAlloc(cbPhysPage);
1102	AssertReturn(paPhysPageCopy, VERR_NO_MEMORY);
1103
1104	memcpy(paPhysPageCopy, paPhysPage, cbPhysPage);
1105
1106	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysChangeMemBalloonHelper, 4, pVM, fInflate, cPages, paPhysPageCopy);
1107	AssertRC(rc);
1108	}
1109	else
1110	{
1111	uintptr_t paUser[3];
1112
1113	paUser[0] = fInflate;
1114	paUser[1] = cPages;
1115	paUser[2] = (uintptr_t)paPhysPage;
1116	rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysChangeMemBalloonRendezvous, (void *)paUser);
1117	AssertRC(rc);
1118	}
1119	return rc;
1120
1121	#else
1122	NOREF(pVM); NOREF(fInflate); NOREF(cPages); NOREF(paPhysPage);
1123	return VERR_NOT_IMPLEMENTED;
1124	#endif
1125	}
1126
1127
1128	/**
1129	* Rendezvous callback used by PGMR3WriteProtectRAM that write protects all
1130	* physical RAM.
1131	*
1132	* This is only called on one of the EMTs while the other ones are waiting for
1133	* it to complete this function.
1134	*
1135	* @returns VINF_SUCCESS (VBox strict status code).
1136	* @param pVM The cross context VM structure.
1137	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
1138	* @param pvUser User parameter, unused.
1139	*/
1140	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysWriteProtectRAMRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
1141	{
1142	int rc = VINF_SUCCESS;
1143	NOREF(pvUser); NOREF(pVCpu);
1144
1145	pgmLock(pVM);
1146	#ifdef PGMPOOL_WITH_OPTIMIZED_DIRTY_PT
1147	pgmPoolResetDirtyPages(pVM);
1148	#endif
1149
1150	/** @todo pointless to write protect the physical page pointed to by RSP. */
1151
1152	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
1153	pRam;
1154	pRam = pRam->CTX_SUFF(pNext))
1155	{
1156	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
1157	for (uint32_t iPage = 0; iPage < cPages; iPage++)
1158	{
1159	PPGMPAGE pPage = &pRam->aPages[iPage];
1160	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
1161
1162	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
1163	\|\| enmPageType == PGMPAGETYPE_MMIO2)
1164	{
1165	/*
1166	* A RAM page.
1167	*/
1168	switch (PGM_PAGE_GET_STATE(pPage))
1169	{
1170	case PGM_PAGE_STATE_ALLOCATED:
1171	/** @todo Optimize this: Don't always re-enable write
1172	* monitoring if the page is known to be very busy. */
1173	if (PGM_PAGE_IS_WRITTEN_TO(pPage))
1174	{
1175	PGM_PAGE_CLEAR_WRITTEN_TO(pVM, pPage);
1176	/* Remember this dirty page for the next (memory) sync. */
1177	PGM_PAGE_SET_FT_DIRTY(pPage);
1178	}
1179
1180	pgmPhysPageWriteMonitor(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
1181	break;
1182
1183	case PGM_PAGE_STATE_SHARED:
1184	AssertFailed();
1185	break;
1186
1187	case PGM_PAGE_STATE_WRITE_MONITORED: /* nothing to change. */
1188	default:
1189	break;
1190	}
1191	}
1192	}
1193	}
1194	pgmR3PoolWriteProtectPages(pVM);
1195	PGM_INVL_ALL_VCPU_TLBS(pVM);
1196	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
1197	CPUMSetChangedFlags(&pVM->aCpus[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
1198
1199	pgmUnlock(pVM);
1200	return rc;
1201	}
1202
1203	/**
1204	* Protect all physical RAM to monitor writes
1205	*
1206	* @returns VBox status code.
1207	* @param pVM The cross context VM structure.
1208	*/
1209	VMMR3DECL(int) PGMR3PhysWriteProtectRAM(PVM pVM)
1210	{
1211	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1212
1213	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysWriteProtectRAMRendezvous, NULL);
1214	AssertRC(rc);
1215	return rc;
1216	}
1217
1218	/**
1219	* Enumerate all dirty FT pages.
1220	*
1221	* @returns VBox status code.
1222	* @param pVM The cross context VM structure.
1223	* @param pfnEnum Enumerate callback handler.
1224	* @param pvUser Enumerate callback handler parameter.
1225	*/
1226	VMMR3DECL(int) PGMR3PhysEnumDirtyFTPages(PVM pVM, PFNPGMENUMDIRTYFTPAGES pfnEnum, void *pvUser)
1227	{
1228	int rc = VINF_SUCCESS;
1229
1230	pgmLock(pVM);
1231	for (PPGMRAMRANGE pRam = pVM->pgm.s.CTX_SUFF(pRamRangesX);
1232	pRam;
1233	pRam = pRam->CTX_SUFF(pNext))
1234	{
1235	uint32_t cPages = pRam->cb >> PAGE_SHIFT;
1236	for (uint32_t iPage = 0; iPage < cPages; iPage++)
1237	{
1238	PPGMPAGE pPage = &pRam->aPages[iPage];
1239	PGMPAGETYPE enmPageType = (PGMPAGETYPE)PGM_PAGE_GET_TYPE(pPage);
1240
1241	if ( RT_LIKELY(enmPageType == PGMPAGETYPE_RAM)
1242	\|\| enmPageType == PGMPAGETYPE_MMIO2)
1243	{
1244	/*
1245	* A RAM page.
1246	*/
1247	switch (PGM_PAGE_GET_STATE(pPage))
1248	{
1249	case PGM_PAGE_STATE_ALLOCATED:
1250	case PGM_PAGE_STATE_WRITE_MONITORED:
1251	if ( !PGM_PAGE_IS_WRITTEN_TO(pPage) /* not very recently updated? */
1252	&& PGM_PAGE_IS_FT_DIRTY(pPage))
1253	{
1254	uint32_t cbPageRange = PAGE_SIZE;
1255	uint32_t iPageClean = iPage + 1;
1256	RTGCPHYS GCPhysPage = pRam->GCPhys + iPage * PAGE_SIZE;
1257	uint8_t *pu8Page = NULL;
1258	PGMPAGEMAPLOCK Lock;
1259
1260	/* Find the next clean page, so we can merge adjacent dirty pages. */
1261	for (; iPageClean < cPages; iPageClean++)
1262	{
1263	PPGMPAGE pPageNext = &pRam->aPages[iPageClean];
1264	if ( RT_UNLIKELY(PGM_PAGE_GET_TYPE(pPageNext) != PGMPAGETYPE_RAM)
1265	\|\| PGM_PAGE_GET_STATE(pPageNext) != PGM_PAGE_STATE_ALLOCATED
1266	\|\| PGM_PAGE_IS_WRITTEN_TO(pPageNext)
1267	\|\| !PGM_PAGE_IS_FT_DIRTY(pPageNext)
1268	/* Crossing a chunk boundary? */
1269	\|\| (GCPhysPage & GMM_PAGEID_IDX_MASK) != ((GCPhysPage + cbPageRange) & GMM_PAGEID_IDX_MASK)
1270	)
1271	break;
1272
1273	cbPageRange += PAGE_SIZE;
1274	}
1275
1276	rc = PGMPhysGCPhys2CCPtrReadOnly(pVM, GCPhysPage, (const void **)&pu8Page, &Lock);
1277	if (RT_SUCCESS(rc))
1278	{
1279	/** @todo this is risky; the range might be changed, but little choice as the sync
1280	* costs a lot of time. */
1281	pgmUnlock(pVM);
1282	pfnEnum(pVM, GCPhysPage, pu8Page, cbPageRange, pvUser);
1283	pgmLock(pVM);
1284	PGMPhysReleasePageMappingLock(pVM, &Lock);
1285	}
1286
1287	for (uint32_t iTmp = iPage; iTmp < iPageClean; iTmp++)
1288	PGM_PAGE_CLEAR_FT_DIRTY(&pRam->aPages[iTmp]);
1289	}
1290	break;
1291	}
1292	}
1293	}
1294	}
1295	pgmUnlock(pVM);
1296	return rc;
1297	}
1298
1299
1300	/**
1301	* Gets the number of ram ranges.
1302	*
1303	* @returns Number of ram ranges. Returns UINT32_MAX if @a pVM is invalid.
1304	* @param pVM The cross context VM structure.
1305	*/
1306	VMMR3DECL(uint32_t) PGMR3PhysGetRamRangeCount(PVM pVM)
1307	{
1308	VM_ASSERT_VALID_EXT_RETURN(pVM, UINT32_MAX);
1309
1310	pgmLock(pVM);
1311	uint32_t cRamRanges = 0;
1312	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext))
1313	cRamRanges++;
1314	pgmUnlock(pVM);
1315	return cRamRanges;
1316	}
1317
1318
1319	/**
1320	* Get information about a range.
1321	*
1322	* @returns VINF_SUCCESS or VERR_OUT_OF_RANGE.
1323	* @param pVM The cross context VM structure.
1324	* @param iRange The ordinal of the range.
1325	* @param pGCPhysStart Where to return the start of the range. Optional.
1326	* @param pGCPhysLast Where to return the address of the last byte in the
1327	* range. Optional.
1328	* @param ppszDesc Where to return the range description. Optional.
1329	* @param pfIsMmio Where to indicate that this is a pure MMIO range.
1330	* Optional.
1331	*/
1332	VMMR3DECL(int) PGMR3PhysGetRange(PVM pVM, uint32_t iRange, PRTGCPHYS pGCPhysStart, PRTGCPHYS pGCPhysLast,
1333	const char *ppszDesc, bool pfIsMmio)
1334	{
1335	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1336
1337	pgmLock(pVM);
1338	uint32_t iCurRange = 0;
1339	for (PPGMRAMRANGE pCur = pVM->pgm.s.CTX_SUFF(pRamRangesX); pCur; pCur = pCur->CTX_SUFF(pNext), iCurRange++)
1340	if (iCurRange == iRange)
1341	{
1342	if (pGCPhysStart)
1343	*pGCPhysStart = pCur->GCPhys;
1344	if (pGCPhysLast)
1345	*pGCPhysLast = pCur->GCPhysLast;
1346	if (ppszDesc)
1347	*ppszDesc = pCur->pszDesc;
1348	if (pfIsMmio)
1349	*pfIsMmio = !!(pCur->fFlags & PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO);
1350
1351	pgmUnlock(pVM);
1352	return VINF_SUCCESS;
1353	}
1354	pgmUnlock(pVM);
1355	return VERR_OUT_OF_RANGE;
1356	}
1357
1358
1359	/**
1360	* Query the amount of free memory inside VMMR0
1361	*
1362	* @returns VBox status code.
1363	* @param pUVM The user mode VM handle.
1364	* @param pcbAllocMem Where to return the amount of memory allocated
1365	* by VMs.
1366	* @param pcbFreeMem Where to return the amount of memory that is
1367	* allocated from the host but not currently used
1368	* by any VMs.
1369	* @param pcbBallonedMem Where to return the sum of memory that is
1370	* currently ballooned by the VMs.
1371	* @param pcbSharedMem Where to return the amount of memory that is
1372	* currently shared.
1373	*/
1374	VMMR3DECL(int) PGMR3QueryGlobalMemoryStats(PUVM pUVM, uint64_t pcbAllocMem, uint64_t pcbFreeMem,
1375	uint64_t pcbBallonedMem, uint64_t pcbSharedMem)
1376	{
1377	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1378	VM_ASSERT_VALID_EXT_RETURN(pUVM->pVM, VERR_INVALID_VM_HANDLE);
1379
1380	uint64_t cAllocPages = 0;
1381	uint64_t cFreePages = 0;
1382	uint64_t cBalloonPages = 0;
1383	uint64_t cSharedPages = 0;
1384	int rc = GMMR3QueryHypervisorMemoryStats(pUVM->pVM, &cAllocPages, &cFreePages, &cBalloonPages, &cSharedPages);
1385	AssertRCReturn(rc, rc);
1386
1387	if (pcbAllocMem)
1388	pcbAllocMem = cAllocPages _4K;
1389
1390	if (pcbFreeMem)
1391	pcbFreeMem = cFreePages _4K;
1392
1393	if (pcbBallonedMem)
1394	pcbBallonedMem = cBalloonPages _4K;
1395
1396	if (pcbSharedMem)
1397	pcbSharedMem = cSharedPages _4K;
1398
1399	Log(("PGMR3QueryVMMMemoryStats: all=%llx free=%llx ballooned=%llx shared=%llx\n",
1400	cAllocPages, cFreePages, cBalloonPages, cSharedPages));
1401	return VINF_SUCCESS;
1402	}
1403
1404
1405	/**
1406	* Query memory stats for the VM.
1407	*
1408	* @returns VBox status code.
1409	* @param pUVM The user mode VM handle.
1410	* @param pcbTotalMem Where to return total amount memory the VM may
1411	* possibly use.
1412	* @param pcbPrivateMem Where to return the amount of private memory
1413	* currently allocated.
1414	* @param pcbSharedMem Where to return the amount of actually shared
1415	* memory currently used by the VM.
1416	* @param pcbZeroMem Where to return the amount of memory backed by
1417	* zero pages.
1418	*
1419	* @remarks The total mem is normally larger than the sum of the three
1420	* components. There are two reasons for this, first the amount of
1421	* shared memory is what we're sure is shared instead of what could
1422	* possibly be shared with someone. Secondly, because the total may
1423	* include some pure MMIO pages that doesn't go into any of the three
1424	* sub-counts.
1425	*
1426	* @todo Why do we return reused shared pages instead of anything that could
1427	* potentially be shared? Doesn't this mean the first VM gets a much
1428	* lower number of shared pages?
1429	*/
1430	VMMR3DECL(int) PGMR3QueryMemoryStats(PUVM pUVM, uint64_t pcbTotalMem, uint64_t pcbPrivateMem,
1431	uint64_t pcbSharedMem, uint64_t pcbZeroMem)
1432	{
1433	UVM_ASSERT_VALID_EXT_RETURN(pUVM, VERR_INVALID_VM_HANDLE);
1434	PVM pVM = pUVM->pVM;
1435	VM_ASSERT_VALID_EXT_RETURN(pVM, VERR_INVALID_VM_HANDLE);
1436
1437	if (pcbTotalMem)
1438	pcbTotalMem = (uint64_t)pVM->pgm.s.cAllPages PAGE_SIZE;
1439
1440	if (pcbPrivateMem)
1441	pcbPrivateMem = (uint64_t)pVM->pgm.s.cPrivatePages PAGE_SIZE;
1442
1443	if (pcbSharedMem)
1444	pcbSharedMem = (uint64_t)pVM->pgm.s.cReusedSharedPages PAGE_SIZE;
1445
1446	if (pcbZeroMem)
1447	pcbZeroMem = (uint64_t)pVM->pgm.s.cZeroPages PAGE_SIZE;
1448
1449	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));
1450	return VINF_SUCCESS;
1451	}
1452
1453
1454	/**
1455	* PGMR3PhysRegisterRam worker that initializes and links a RAM range.
1456	*
1457	* @param pVM The cross context VM structure.
1458	* @param pNew The new RAM range.
1459	* @param GCPhys The address of the RAM range.
1460	* @param GCPhysLast The last address of the RAM range.
1461	* @param RCPtrNew The RC address if the range is floating. NIL_RTRCPTR
1462	* if in HMA.
1463	* @param R0PtrNew Ditto for R0.
1464	* @param pszDesc The description.
1465	* @param pPrev The previous RAM range (for linking).
1466	*/
1467	static void pgmR3PhysInitAndLinkRamRange(PVM pVM, PPGMRAMRANGE pNew, RTGCPHYS GCPhys, RTGCPHYS GCPhysLast,
1468	RTRCPTR RCPtrNew, RTR0PTR R0PtrNew, const char *pszDesc, PPGMRAMRANGE pPrev)
1469	{
1470	/*
1471	* Initialize the range.
1472	*/
1473	pNew->pSelfR0 = R0PtrNew != NIL_RTR0PTR ? R0PtrNew : MMHyperCCToR0(pVM, pNew);
1474	pNew->pSelfRC = RCPtrNew != NIL_RTRCPTR ? RCPtrNew : MMHyperCCToRC(pVM, pNew);
1475	pNew->GCPhys = GCPhys;
1476	pNew->GCPhysLast = GCPhysLast;
1477	pNew->cb = GCPhysLast - GCPhys + 1;
1478	pNew->pszDesc = pszDesc;
1479	pNew->fFlags = RCPtrNew != NIL_RTRCPTR ? PGM_RAM_RANGE_FLAGS_FLOATING : 0;
1480	pNew->pvR3 = NULL;
1481	pNew->paLSPages = NULL;
1482
1483	uint32_t const cPages = pNew->cb >> PAGE_SHIFT;
1484	RTGCPHYS iPage = cPages;
1485	while (iPage-- > 0)
1486	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_RAM);
1487
1488	/* Update the page count stats. */
1489	pVM->pgm.s.cZeroPages += cPages;
1490	pVM->pgm.s.cAllPages += cPages;
1491
1492	/*
1493	* Link it.
1494	*/
1495	pgmR3PhysLinkRamRange(pVM, pNew, pPrev);
1496	}
1497
1498
1499	/**
1500	* Relocate a floating RAM range.
1501	*
1502	* @copydoc FNPGMRELOCATE
1503	*/
1504	static DECLCALLBACK(bool) pgmR3PhysRamRangeRelocate(PVM pVM, RTGCPTR GCPtrOld, RTGCPTR GCPtrNew,
1505	PGMRELOCATECALL enmMode, void *pvUser)
1506	{
1507	PPGMRAMRANGE pRam = (PPGMRAMRANGE)pvUser;
1508	Assert(pRam->fFlags & PGM_RAM_RANGE_FLAGS_FLOATING);
1509	Assert(pRam->pSelfRC == GCPtrOld + PAGE_SIZE); RT_NOREF_PV(GCPtrOld);
1510
1511	switch (enmMode)
1512	{
1513	case PGMRELOCATECALL_SUGGEST:
1514	return true;
1515
1516	case PGMRELOCATECALL_RELOCATE:
1517	{
1518	/*
1519	* Update myself, then relink all the ranges and flush the RC TLB.
1520	*/
1521	pgmLock(pVM);
1522
1523	pRam->pSelfRC = (RTRCPTR)(GCPtrNew + PAGE_SIZE);
1524
1525	pgmR3PhysRelinkRamRanges(pVM);
1526	for (unsigned i = 0; i < PGM_RAMRANGE_TLB_ENTRIES; i++)
1527	pVM->pgm.s.apRamRangesTlbRC[i] = NIL_RTRCPTR;
1528
1529	pgmUnlock(pVM);
1530	return true;
1531	}
1532
1533	default:
1534	AssertFailedReturn(false);
1535	}
1536	}
1537
1538
1539	/**
1540	* PGMR3PhysRegisterRam worker that registers a high chunk.
1541	*
1542	* @returns VBox status code.
1543	* @param pVM The cross context VM structure.
1544	* @param GCPhys The address of the RAM.
1545	* @param cRamPages The number of RAM pages to register.
1546	* @param cbChunk The size of the PGMRAMRANGE guest mapping.
1547	* @param iChunk The chunk number.
1548	* @param pszDesc The RAM range description.
1549	* @param ppPrev Previous RAM range pointer. In/Out.
1550	*/
1551	static int pgmR3PhysRegisterHighRamChunk(PVM pVM, RTGCPHYS GCPhys, uint32_t cRamPages,
1552	uint32_t cbChunk, uint32_t iChunk, const char *pszDesc,
1553	PPGMRAMRANGE *ppPrev)
1554	{
1555	const char *pszDescChunk = iChunk == 0
1556	? pszDesc
1557	: MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s (#%u)", pszDesc, iChunk + 1);
1558	AssertReturn(pszDescChunk, VERR_NO_MEMORY);
1559
1560	/*
1561	* Allocate memory for the new chunk.
1562	*/
1563	size_t const cChunkPages = RT_ALIGN_Z(RT_UOFFSETOF(PGMRAMRANGE, aPages[cRamPages]), PAGE_SIZE) >> PAGE_SHIFT;
1564	PSUPPAGE paChunkPages = (PSUPPAGE)RTMemTmpAllocZ(sizeof(SUPPAGE) * cChunkPages);
1565	AssertReturn(paChunkPages, VERR_NO_TMP_MEMORY);
1566	RTR0PTR R0PtrChunk = NIL_RTR0PTR;
1567	void *pvChunk = NULL;
1568	int rc = SUPR3PageAllocEx(cChunkPages, 0 /fFlags/, &pvChunk,
1569	#if defined(VBOX_WITH_MORE_RING0_MEM_MAPPINGS)
1570	&R0PtrChunk,
1571	#elif defined(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	#if defined(VBOX_WITH_MORE_RING0_MEM_MAPPINGS)
1580	Assert(R0PtrChunk != NIL_RTR0PTR);
1581	#elif defined(VBOX_WITH_2X_4GB_ADDR_SPACE)
1582	if (!HMIsEnabled(pVM))
1583	R0PtrChunk = NIL_RTR0PTR;
1584	#else
1585	R0PtrChunk = (uintptr_t)pvChunk;
1586	#endif
1587	memset(pvChunk, 0, cChunkPages << PAGE_SHIFT);
1588
1589	PPGMRAMRANGE pNew = (PPGMRAMRANGE)pvChunk;
1590
1591	/*
1592	* Create a mapping and map the pages into it.
1593	* We push these in below the HMA.
1594	*/
1595	RTGCPTR GCPtrChunkMap = pVM->pgm.s.GCPtrPrevRamRangeMapping - cbChunk;
1596	rc = PGMR3MapPT(pVM, GCPtrChunkMap, cbChunk, 0 /fFlags/, pgmR3PhysRamRangeRelocate, pNew, pszDescChunk);
1597	if (RT_SUCCESS(rc))
1598	{
1599	pVM->pgm.s.GCPtrPrevRamRangeMapping = GCPtrChunkMap;
1600
1601	RTGCPTR const GCPtrChunk = GCPtrChunkMap + PAGE_SIZE;
1602	RTGCPTR GCPtrPage = GCPtrChunk;
1603	for (uint32_t iPage = 0; iPage < cChunkPages && RT_SUCCESS(rc); iPage++, GCPtrPage += PAGE_SIZE)
1604	rc = PGMMap(pVM, GCPtrPage, paChunkPages[iPage].Phys, PAGE_SIZE, 0);
1605	if (RT_SUCCESS(rc))
1606	{
1607	/*
1608	* Ok, init and link the range.
1609	*/
1610	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhys + ((RTGCPHYS)cRamPages << PAGE_SHIFT) - 1,
1611	(RTRCPTR)GCPtrChunk, R0PtrChunk, pszDescChunk, *ppPrev);
1612	*ppPrev = pNew;
1613	}
1614	}
1615
1616	if (RT_FAILURE(rc))
1617	SUPR3PageFreeEx(pvChunk, cChunkPages);
1618	}
1619
1620	RTMemTmpFree(paChunkPages);
1621	return rc;
1622	}
1623
1624
1625	/**
1626	* Sets up a range RAM.
1627	*
1628	* This will check for conflicting registrations, make a resource
1629	* reservation for the memory (with GMM), and setup the per-page
1630	* tracking structures (PGMPAGE).
1631	*
1632	* @returns VBox status code.
1633	* @param pVM The cross context VM structure.
1634	* @param GCPhys The physical address of the RAM.
1635	* @param cb The size of the RAM.
1636	* @param pszDesc The description - not copied, so, don't free or change it.
1637	*/
1638	VMMR3DECL(int) PGMR3PhysRegisterRam(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, const char *pszDesc)
1639	{
1640	/*
1641	* Validate input.
1642	*/
1643	Log(("PGMR3PhysRegisterRam: GCPhys=%RGp cb=%RGp pszDesc=%s\n", GCPhys, cb, pszDesc));
1644	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
1645	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
1646	AssertReturn(cb > 0, VERR_INVALID_PARAMETER);
1647	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
1648	AssertMsgReturn(GCPhysLast > GCPhys, ("The range wraps! GCPhys=%RGp cb=%RGp\n", GCPhys, cb), VERR_INVALID_PARAMETER);
1649	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
1650	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
1651
1652	pgmLock(pVM);
1653
1654	/*
1655	* Find range location and check for conflicts.
1656	* (We don't lock here because the locking by EMT is only required on update.)
1657	*/
1658	PPGMRAMRANGE pPrev = NULL;
1659	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
1660	while (pRam && GCPhysLast >= pRam->GCPhys)
1661	{
1662	if ( GCPhysLast >= pRam->GCPhys
1663	&& GCPhys <= pRam->GCPhysLast)
1664	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
1665	GCPhys, GCPhysLast, pszDesc,
1666	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
1667	VERR_PGM_RAM_CONFLICT);
1668
1669	/* next */
1670	pPrev = pRam;
1671	pRam = pRam->pNextR3;
1672	}
1673
1674	/*
1675	* Register it with GMM (the API bitches).
1676	*/
1677	const RTGCPHYS cPages = cb >> PAGE_SHIFT;
1678	int rc = MMR3IncreaseBaseReservation(pVM, cPages);
1679	if (RT_FAILURE(rc))
1680	{
1681	pgmUnlock(pVM);
1682	return rc;
1683	}
1684
1685	if ( GCPhys >= _4G
1686	&& cPages > 256)
1687	{
1688	/*
1689	* The PGMRAMRANGE structures for the high memory can get very big.
1690	* In order to avoid SUPR3PageAllocEx allocation failures due to the
1691	* allocation size limit there and also to avoid being unable to find
1692	* guest mapping space for them, we split this memory up into 4MB in
1693	* (potential) raw-mode configs and 16MB chunks in forced AMD-V/VT-x
1694	* mode.
1695	*
1696	* The first and last page of each mapping are guard pages and marked
1697	* not-present. So, we've got 4186112 and 16769024 bytes available for
1698	* the PGMRAMRANGE structure.
1699	*
1700	* Note! The sizes used here will influence the saved state.
1701	*/
1702	uint32_t cbChunk;
1703	uint32_t cPagesPerChunk;
1704	if (HMIsEnabled(pVM))
1705	{
1706	cbChunk = 16U*_1M;
1707	cPagesPerChunk = 1048048; /* max ~1048059 */
1708	AssertCompile(sizeof(PGMRAMRANGE) + sizeof(PGMPAGE) * 1048048 < 16U_1M - PAGE_SIZE 2);
1709	}
1710	else
1711	{
1712	cbChunk = 4U*_1M;
1713	cPagesPerChunk = 261616; /* max ~261627 */
1714	AssertCompile(sizeof(PGMRAMRANGE) + sizeof(PGMPAGE) * 261616 < 4U_1M - PAGE_SIZE 2);
1715	}
1716	AssertRelease(RT_UOFFSETOF(PGMRAMRANGE, aPages[cPagesPerChunk]) + PAGE_SIZE * 2 <= cbChunk);
1717
1718	RTGCPHYS cPagesLeft = cPages;
1719	RTGCPHYS GCPhysChunk = GCPhys;
1720	uint32_t iChunk = 0;
1721	while (cPagesLeft > 0)
1722	{
1723	uint32_t cPagesInChunk = cPagesLeft;
1724	if (cPagesInChunk > cPagesPerChunk)
1725	cPagesInChunk = cPagesPerChunk;
1726
1727	rc = pgmR3PhysRegisterHighRamChunk(pVM, GCPhysChunk, cPagesInChunk, cbChunk, iChunk, pszDesc, &pPrev);
1728	AssertRCReturn(rc, rc);
1729
1730	/* advance */
1731	GCPhysChunk += (RTGCPHYS)cPagesInChunk << PAGE_SHIFT;
1732	cPagesLeft -= cPagesInChunk;
1733	iChunk++;
1734	}
1735	}
1736	else
1737	{
1738	/*
1739	* Allocate, initialize and link the new RAM range.
1740	*/
1741	const size_t cbRamRange = RT_OFFSETOF(PGMRAMRANGE, aPages[cPages]);
1742	PPGMRAMRANGE pNew;
1743	rc = MMR3HyperAllocOnceNoRel(pVM, cbRamRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
1744	AssertLogRelMsgRCReturn(rc, ("cbRamRange=%zu\n", cbRamRange), rc);
1745
1746	pgmR3PhysInitAndLinkRamRange(pVM, pNew, GCPhys, GCPhysLast, NIL_RTRCPTR, NIL_RTR0PTR, pszDesc, pPrev);
1747	}
1748	pgmPhysInvalidatePageMapTLB(pVM);
1749	pgmUnlock(pVM);
1750
1751	#ifdef VBOX_WITH_REM
1752	/*
1753	* Notify REM.
1754	*/
1755	REMR3NotifyPhysRamRegister(pVM, GCPhys, cb, REM_NOTIFY_PHYS_RAM_FLAGS_RAM);
1756	#endif
1757
1758	return VINF_SUCCESS;
1759	}
1760
1761
1762	/**
1763	* Worker called by PGMR3InitFinalize if we're configured to pre-allocate RAM.
1764	*
1765	* We do this late in the init process so that all the ROM and MMIO ranges have
1766	* been registered already and we don't go wasting memory on them.
1767	*
1768	* @returns VBox status code.
1769	*
1770	* @param pVM The cross context VM structure.
1771	*/
1772	int pgmR3PhysRamPreAllocate(PVM pVM)
1773	{
1774	Assert(pVM->pgm.s.fRamPreAlloc);
1775	Log(("pgmR3PhysRamPreAllocate: enter\n"));
1776
1777	/*
1778	* Walk the RAM ranges and allocate all RAM pages, halt at
1779	* the first allocation error.
1780	*/
1781	uint64_t cPages = 0;
1782	uint64_t NanoTS = RTTimeNanoTS();
1783	pgmLock(pVM);
1784	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1785	{
1786	PPGMPAGE pPage = &pRam->aPages[0];
1787	RTGCPHYS GCPhys = pRam->GCPhys;
1788	uint32_t cLeft = pRam->cb >> PAGE_SHIFT;
1789	while (cLeft-- > 0)
1790	{
1791	if (PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM)
1792	{
1793	switch (PGM_PAGE_GET_STATE(pPage))
1794	{
1795	case PGM_PAGE_STATE_ZERO:
1796	{
1797	int rc = pgmPhysAllocPage(pVM, pPage, GCPhys);
1798	if (RT_FAILURE(rc))
1799	{
1800	LogRel(("PGM: RAM Pre-allocation failed at %RGp (in %s) with rc=%Rrc\n", GCPhys, pRam->pszDesc, rc));
1801	pgmUnlock(pVM);
1802	return rc;
1803	}
1804	cPages++;
1805	break;
1806	}
1807
1808	case PGM_PAGE_STATE_BALLOONED:
1809	case PGM_PAGE_STATE_ALLOCATED:
1810	case PGM_PAGE_STATE_WRITE_MONITORED:
1811	case PGM_PAGE_STATE_SHARED:
1812	/* nothing to do here. */
1813	break;
1814	}
1815	}
1816
1817	/* next */
1818	pPage++;
1819	GCPhys += PAGE_SIZE;
1820	}
1821	}
1822	pgmUnlock(pVM);
1823	NanoTS = RTTimeNanoTS() - NanoTS;
1824
1825	LogRel(("PGM: Pre-allocated %llu pages in %llu ms\n", cPages, NanoTS / 1000000));
1826	Log(("pgmR3PhysRamPreAllocate: returns VINF_SUCCESS\n"));
1827	return VINF_SUCCESS;
1828	}
1829
1830
1831	/**
1832	* Checks shared page checksums.
1833	*
1834	* @param pVM The cross context VM structure.
1835	*/
1836	void pgmR3PhysAssertSharedPageChecksums(PVM pVM)
1837	{
1838	#ifdef VBOX_STRICT
1839	pgmLock(pVM);
1840
1841	if (pVM->pgm.s.cSharedPages > 0)
1842	{
1843	/*
1844	* Walk the ram ranges.
1845	*/
1846	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1847	{
1848	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
1849	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
1850
1851	while (iPage-- > 0)
1852	{
1853	PPGMPAGE pPage = &pRam->aPages[iPage];
1854	if (PGM_PAGE_IS_SHARED(pPage))
1855	{
1856	uint32_t u32Checksum = pPage->s.u2Unused0 \| ((uint32_t)pPage->s.u2Unused1 << 8);
1857	if (!u32Checksum)
1858	{
1859	RTGCPHYS GCPhysPage = pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT);
1860	void const *pvPage;
1861	int rc = pgmPhysPageMapReadOnly(pVM, pPage, GCPhysPage, &pvPage);
1862	if (RT_SUCCESS(rc))
1863	{
1864	uint32_t u32Checksum2 = RTCrc32(pvPage, PAGE_SIZE);
1865	# if 0
1866	AssertMsg((u32Checksum2 & UINT32_C(0x00000303)) == u32Checksum, ("GCPhysPage=%RGp\n", GCPhysPage));
1867	# else
1868	if ((u32Checksum2 & UINT32_C(0x00000303)) == u32Checksum)
1869	LogFlow(("shpg %#x @ %RGp %#x [OK]\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
1870	else
1871	AssertMsgFailed(("shpg %#x @ %RGp %#x\n", PGM_PAGE_GET_PAGEID(pPage), GCPhysPage, u32Checksum2));
1872	# endif
1873	}
1874	else
1875	AssertRC(rc);
1876	}
1877	}
1878
1879	} /* for each page */
1880
1881	} /* for each ram range */
1882	}
1883
1884	pgmUnlock(pVM);
1885	#endif /* VBOX_STRICT */
1886	NOREF(pVM);
1887	}
1888
1889
1890	/**
1891	* Resets the physical memory state.
1892	*
1893	* ASSUMES that the caller owns the PGM lock.
1894	*
1895	* @returns VBox status code.
1896	* @param pVM The cross context VM structure.
1897	*/
1898	int pgmR3PhysRamReset(PVM pVM)
1899	{
1900	PGM_LOCK_ASSERT_OWNER(pVM);
1901
1902	/* Reset the memory balloon. */
1903	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
1904	AssertRC(rc);
1905
1906	#ifdef VBOX_WITH_PAGE_SHARING
1907	/* Clear all registered shared modules. */
1908	pgmR3PhysAssertSharedPageChecksums(pVM);
1909	rc = GMMR3ResetSharedModules(pVM);
1910	AssertRC(rc);
1911	#endif
1912	/* Reset counters. */
1913	pVM->pgm.s.cReusedSharedPages = 0;
1914	pVM->pgm.s.cBalloonedPages = 0;
1915
1916	return VINF_SUCCESS;
1917	}
1918
1919
1920	/**
1921	* Resets (zeros) the RAM after all devices and components have been reset.
1922	*
1923	* ASSUMES that the caller owns the PGM lock.
1924	*
1925	* @returns VBox status code.
1926	* @param pVM The cross context VM structure.
1927	*/
1928	int pgmR3PhysRamZeroAll(PVM pVM)
1929	{
1930	PGM_LOCK_ASSERT_OWNER(pVM);
1931
1932	/*
1933	* We batch up pages that should be freed instead of calling GMM for
1934	* each and every one of them.
1935	*/
1936	uint32_t cPendingPages = 0;
1937	PGMMFREEPAGESREQ pReq;
1938	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
1939	AssertLogRelRCReturn(rc, rc);
1940
1941	/*
1942	* Walk the ram ranges.
1943	*/
1944	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
1945	{
1946	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
1947	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
1948
1949	if ( !pVM->pgm.s.fRamPreAlloc
1950	&& pVM->pgm.s.fZeroRamPagesOnReset)
1951	{
1952	/* Replace all RAM pages by ZERO pages. */
1953	while (iPage-- > 0)
1954	{
1955	PPGMPAGE pPage = &pRam->aPages[iPage];
1956	switch (PGM_PAGE_GET_TYPE(pPage))
1957	{
1958	case PGMPAGETYPE_RAM:
1959	/* Do not replace pages part of a 2 MB continuous range
1960	with zero pages, but zero them instead. */
1961	if ( PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE
1962	\|\| PGM_PAGE_GET_PDE_TYPE(pPage) == PGM_PAGE_PDE_TYPE_PDE_DISABLED)
1963	{
1964	void *pvPage;
1965	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
1966	AssertLogRelRCReturn(rc, rc);
1967	ASMMemZeroPage(pvPage);
1968	}
1969	else if (PGM_PAGE_IS_BALLOONED(pPage))
1970	{
1971	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
1972	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
1973	}
1974	else if (!PGM_PAGE_IS_ZERO(pPage))
1975	{
1976	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
1977	AssertLogRelRCReturn(rc, rc);
1978	}
1979	break;
1980
1981	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
1982	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
1983	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
1984	true /fDoAccounting/);
1985	break;
1986
1987	case PGMPAGETYPE_MMIO2:
1988	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
1989	case PGMPAGETYPE_ROM:
1990	case PGMPAGETYPE_MMIO:
1991	break;
1992	default:
1993	AssertFailed();
1994	}
1995	} /* for each page */
1996	}
1997	else
1998	{
1999	/* Zero the memory. */
2000	while (iPage-- > 0)
2001	{
2002	PPGMPAGE pPage = &pRam->aPages[iPage];
2003	switch (PGM_PAGE_GET_TYPE(pPage))
2004	{
2005	case PGMPAGETYPE_RAM:
2006	switch (PGM_PAGE_GET_STATE(pPage))
2007	{
2008	case PGM_PAGE_STATE_ZERO:
2009	break;
2010
2011	case PGM_PAGE_STATE_BALLOONED:
2012	/* Turn into a zero page; the balloon status is lost when the VM reboots. */
2013	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
2014	break;
2015
2016	case PGM_PAGE_STATE_SHARED:
2017	case PGM_PAGE_STATE_WRITE_MONITORED:
2018	rc = pgmPhysPageMakeWritable(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2019	AssertLogRelRCReturn(rc, rc);
2020	/* no break */
2021
2022	case PGM_PAGE_STATE_ALLOCATED:
2023	if (pVM->pgm.s.fZeroRamPagesOnReset)
2024	{
2025	void *pvPage;
2026	rc = pgmPhysPageMap(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pvPage);
2027	AssertLogRelRCReturn(rc, rc);
2028	ASMMemZeroPage(pvPage);
2029	}
2030	break;
2031	}
2032	break;
2033
2034	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2035	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO: /** @todo perhaps leave the special page alone? I don't think VT-x copes with this code. */
2036	pgmHandlerPhysicalResetAliasedPage(pVM, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT),
2037	true /fDoAccounting/);
2038	break;
2039
2040	case PGMPAGETYPE_MMIO2:
2041	case PGMPAGETYPE_ROM_SHADOW:
2042	case PGMPAGETYPE_ROM:
2043	case PGMPAGETYPE_MMIO:
2044	break;
2045	default:
2046	AssertFailed();
2047
2048	}
2049	} /* for each page */
2050	}
2051
2052	}
2053
2054	/*
2055	* Finish off any pages pending freeing.
2056	*/
2057	if (cPendingPages)
2058	{
2059	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2060	AssertLogRelRCReturn(rc, rc);
2061	}
2062	GMMR3FreePagesCleanup(pReq);
2063	return VINF_SUCCESS;
2064	}
2065
2066
2067	/**
2068	* Frees all RAM during VM termination
2069	*
2070	* ASSUMES that the caller owns the PGM lock.
2071	*
2072	* @returns VBox status code.
2073	* @param pVM The cross context VM structure.
2074	*/
2075	int pgmR3PhysRamTerm(PVM pVM)
2076	{
2077	PGM_LOCK_ASSERT_OWNER(pVM);
2078
2079	/* Reset the memory balloon. */
2080	int rc = GMMR3BalloonedPages(pVM, GMMBALLOONACTION_RESET, 0);
2081	AssertRC(rc);
2082
2083	#ifdef VBOX_WITH_PAGE_SHARING
2084	/*
2085	* Clear all registered shared modules.
2086	*/
2087	pgmR3PhysAssertSharedPageChecksums(pVM);
2088	rc = GMMR3ResetSharedModules(pVM);
2089	AssertRC(rc);
2090
2091	/*
2092	* Flush the handy pages updates to make sure no shared pages are hiding
2093	* in there. (No unlikely if the VM shuts down, apparently.)
2094	*/
2095	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_FLUSH_HANDY_PAGES, 0, NULL);
2096	#endif
2097
2098	/*
2099	* We batch up pages that should be freed instead of calling GMM for
2100	* each and every one of them.
2101	*/
2102	uint32_t cPendingPages = 0;
2103	PGMMFREEPAGESREQ pReq;
2104	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2105	AssertLogRelRCReturn(rc, rc);
2106
2107	/*
2108	* Walk the ram ranges.
2109	*/
2110	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3; pRam; pRam = pRam->pNextR3)
2111	{
2112	uint32_t iPage = pRam->cb >> PAGE_SHIFT;
2113	AssertMsg(((RTGCPHYS)iPage << PAGE_SHIFT) == pRam->cb, ("%RGp %RGp\n", (RTGCPHYS)iPage << PAGE_SHIFT, pRam->cb));
2114
2115	while (iPage-- > 0)
2116	{
2117	PPGMPAGE pPage = &pRam->aPages[iPage];
2118	switch (PGM_PAGE_GET_TYPE(pPage))
2119	{
2120	case PGMPAGETYPE_RAM:
2121	/* Free all shared pages. Private pages are automatically freed during GMM VM cleanup. */
2122	/** @todo change this to explicitly free private pages here. */
2123	if (PGM_PAGE_IS_SHARED(pPage))
2124	{
2125	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPage, pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT));
2126	AssertLogRelRCReturn(rc, rc);
2127	}
2128	break;
2129
2130	case PGMPAGETYPE_MMIO2_ALIAS_MMIO:
2131	case PGMPAGETYPE_SPECIAL_ALIAS_MMIO:
2132	case PGMPAGETYPE_MMIO2:
2133	case PGMPAGETYPE_ROM_SHADOW: /* handled by pgmR3PhysRomReset. */
2134	case PGMPAGETYPE_ROM:
2135	case PGMPAGETYPE_MMIO:
2136	break;
2137	default:
2138	AssertFailed();
2139	}
2140	} /* for each page */
2141	}
2142
2143	/*
2144	* Finish off any pages pending freeing.
2145	*/
2146	if (cPendingPages)
2147	{
2148	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2149	AssertLogRelRCReturn(rc, rc);
2150	}
2151	GMMR3FreePagesCleanup(pReq);
2152	return VINF_SUCCESS;
2153	}
2154
2155
2156	/**
2157	* This is the interface IOM is using to register an MMIO region.
2158	*
2159	* It will check for conflicts and ensure that a RAM range structure
2160	* is present before calling the PGMR3HandlerPhysicalRegister API to
2161	* register the callbacks.
2162	*
2163	* @returns VBox status code.
2164	*
2165	* @param pVM The cross context VM structure.
2166	* @param GCPhys The start of the MMIO region.
2167	* @param cb The size of the MMIO region.
2168	* @param hType The physical access handler type registration.
2169	* @param pvUserR3 The user argument for R3.
2170	* @param pvUserR0 The user argument for R0.
2171	* @param pvUserRC The user argument for RC.
2172	* @param pszDesc The description of the MMIO region.
2173	*/
2174	VMMR3DECL(int) PGMR3PhysMMIORegister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMPHYSHANDLERTYPE hType,
2175	RTR3PTR pvUserR3, RTR0PTR pvUserR0, RTRCPTR pvUserRC, const char *pszDesc)
2176	{
2177	/*
2178	* Assert on some assumption.
2179	*/
2180	VM_ASSERT_EMT(pVM);
2181	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2182	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2183	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2184	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2185	Assert(((PPGMPHYSHANDLERTYPEINT)MMHyperHeapOffsetToPtr(pVM, hType))->enmKind == PGMPHYSHANDLERKIND_MMIO);
2186
2187	int rc = pgmLock(pVM);
2188	AssertRCReturn(rc, rc);
2189
2190	/*
2191	* Make sure there's a RAM range structure for the region.
2192	*/
2193	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2194	bool fRamExists = false;
2195	PPGMRAMRANGE pRamPrev = NULL;
2196	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2197	while (pRam && GCPhysLast >= pRam->GCPhys)
2198	{
2199	if ( GCPhysLast >= pRam->GCPhys
2200	&& GCPhys <= pRam->GCPhysLast)
2201	{
2202	/* Simplification: all within the same range. */
2203	AssertLogRelMsgReturnStmt( GCPhys >= pRam->GCPhys
2204	&& GCPhysLast <= pRam->GCPhysLast,
2205	("%RGp-%RGp (MMIO/%s) falls partly outside %RGp-%RGp (%s)\n",
2206	GCPhys, GCPhysLast, pszDesc,
2207	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2208	pgmUnlock(pVM),
2209	VERR_PGM_RAM_CONFLICT);
2210
2211	/* Check that it's all RAM or MMIO pages. */
2212	PCPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2213	uint32_t cLeft = cb >> PAGE_SHIFT;
2214	while (cLeft-- > 0)
2215	{
2216	AssertLogRelMsgReturnStmt( PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM
2217	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO,
2218	("%RGp-%RGp (MMIO/%s): %RGp is not a RAM or MMIO page - type=%d desc=%s\n",
2219	GCPhys, GCPhysLast, pszDesc, pRam->GCPhys, PGM_PAGE_GET_TYPE(pPage), pRam->pszDesc),
2220	pgmUnlock(pVM),
2221	VERR_PGM_RAM_CONFLICT);
2222	pPage++;
2223	}
2224
2225	/* Looks good. */
2226	fRamExists = true;
2227	break;
2228	}
2229
2230	/* next */
2231	pRamPrev = pRam;
2232	pRam = pRam->pNextR3;
2233	}
2234	PPGMRAMRANGE pNew;
2235	if (fRamExists)
2236	{
2237	pNew = NULL;
2238
2239	/*
2240	* Make all the pages in the range MMIO/ZERO pages, freeing any
2241	* RAM pages currently mapped here. This might not be 100% correct
2242	* for PCI memory, but we're doing the same thing for MMIO2 pages.
2243	*/
2244	rc = pgmR3PhysFreePageRange(pVM, pRam, GCPhys, GCPhysLast, PGMPAGETYPE_MMIO);
2245	AssertRCReturnStmt(rc, pgmUnlock(pVM), rc);
2246
2247	/* Force a PGM pool flush as guest ram references have been changed. */
2248	/** @todo not entirely SMP safe; assuming for now the guest takes
2249	* care of this internally (not touch mapped mmio while changing the
2250	* mapping). */
2251	PVMCPU pVCpu = VMMGetCpu(pVM);
2252	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2253	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2254	}
2255	else
2256	{
2257
2258	/*
2259	* No RAM range, insert an ad hoc one.
2260	*
2261	* Note that we don't have to tell REM about this range because
2262	* PGMHandlerPhysicalRegisterEx will do that for us.
2263	*/
2264	Log(("PGMR3PhysMMIORegister: Adding ad hoc MMIO range for %RGp-%RGp %s\n", GCPhys, GCPhysLast, pszDesc));
2265
2266	const uint32_t cPages = cb >> PAGE_SHIFT;
2267	const size_t cbRamRange = RT_OFFSETOF(PGMRAMRANGE, aPages[cPages]);
2268	rc = MMHyperAlloc(pVM, RT_OFFSETOF(PGMRAMRANGE, aPages[cPages]), 16, MM_TAG_PGM_PHYS, (void **)&pNew);
2269	AssertLogRelMsgRCReturnStmt(rc, ("cbRamRange=%zu\n", cbRamRange), pgmUnlock(pVM), rc);
2270
2271	/* Initialize the range. */
2272	pNew->pSelfR0 = MMHyperCCToR0(pVM, pNew);
2273	pNew->pSelfRC = MMHyperCCToRC(pVM, pNew);
2274	pNew->GCPhys = GCPhys;
2275	pNew->GCPhysLast = GCPhysLast;
2276	pNew->cb = cb;
2277	pNew->pszDesc = pszDesc;
2278	pNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO;
2279	pNew->pvR3 = NULL;
2280	pNew->paLSPages = NULL;
2281
2282	uint32_t iPage = cPages;
2283	while (iPage-- > 0)
2284	PGM_PAGE_INIT_ZERO(&pNew->aPages[iPage], pVM, PGMPAGETYPE_MMIO);
2285	Assert(PGM_PAGE_GET_TYPE(&pNew->aPages[0]) == PGMPAGETYPE_MMIO);
2286
2287	/* update the page count stats. */
2288	pVM->pgm.s.cPureMmioPages += cPages;
2289	pVM->pgm.s.cAllPages += cPages;
2290
2291	/* link it */
2292	pgmR3PhysLinkRamRange(pVM, pNew, pRamPrev);
2293	}
2294
2295	/*
2296	* Register the access handler.
2297	*/
2298	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, hType, pvUserR3, pvUserR0, pvUserRC, pszDesc);
2299	if ( RT_FAILURE(rc)
2300	&& !fRamExists)
2301	{
2302	pVM->pgm.s.cPureMmioPages -= cb >> PAGE_SHIFT;
2303	pVM->pgm.s.cAllPages -= cb >> PAGE_SHIFT;
2304
2305	/* remove the ad hoc range. */
2306	pgmR3PhysUnlinkRamRange2(pVM, pNew, pRamPrev);
2307	pNew->cb = pNew->GCPhys = pNew->GCPhysLast = NIL_RTGCPHYS;
2308	MMHyperFree(pVM, pRam);
2309	}
2310	pgmPhysInvalidatePageMapTLB(pVM);
2311
2312	pgmUnlock(pVM);
2313	return rc;
2314	}
2315
2316
2317	/**
2318	* This is the interface IOM is using to register an MMIO region.
2319	*
2320	* It will take care of calling PGMHandlerPhysicalDeregister and clean up
2321	* any ad hoc PGMRAMRANGE left behind.
2322	*
2323	* @returns VBox status code.
2324	* @param pVM The cross context VM structure.
2325	* @param GCPhys The start of the MMIO region.
2326	* @param cb The size of the MMIO region.
2327	*/
2328	VMMR3DECL(int) PGMR3PhysMMIODeregister(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb)
2329	{
2330	VM_ASSERT_EMT(pVM);
2331
2332	int rc = pgmLock(pVM);
2333	AssertRCReturn(rc, rc);
2334
2335	/*
2336	* First deregister the handler, then check if we should remove the ram range.
2337	*/
2338	rc = PGMHandlerPhysicalDeregister(pVM, GCPhys);
2339	if (RT_SUCCESS(rc))
2340	{
2341	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
2342	PPGMRAMRANGE pRamPrev = NULL;
2343	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2344	while (pRam && GCPhysLast >= pRam->GCPhys)
2345	{
2346	/** @todo We're being a bit too careful here. rewrite. */
2347	if ( GCPhysLast == pRam->GCPhysLast
2348	&& GCPhys == pRam->GCPhys)
2349	{
2350	Assert(pRam->cb == cb);
2351
2352	/*
2353	* See if all the pages are dead MMIO pages.
2354	*/
2355	uint32_t const cPages = cb >> PAGE_SHIFT;
2356	bool fAllMMIO = true;
2357	uint32_t iPage = 0;
2358	uint32_t cLeft = cPages;
2359	while (cLeft-- > 0)
2360	{
2361	PPGMPAGE pPage = &pRam->aPages[iPage];
2362	if ( !PGM_PAGE_IS_MMIO_OR_ALIAS(pPage)
2363	/\|\| not-out-of-action later /)
2364	{
2365	fAllMMIO = false;
2366	AssertMsgFailed(("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2367	break;
2368	}
2369	Assert( PGM_PAGE_IS_ZERO(pPage)
2370	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2371	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO);
2372	pPage++;
2373	}
2374	if (fAllMMIO)
2375	{
2376	/*
2377	* Ad-hoc range, unlink and free it.
2378	*/
2379	Log(("PGMR3PhysMMIODeregister: Freeing ad hoc MMIO range for %RGp-%RGp %s\n",
2380	GCPhys, GCPhysLast, pRam->pszDesc));
2381
2382	pVM->pgm.s.cAllPages -= cPages;
2383	pVM->pgm.s.cPureMmioPages -= cPages;
2384
2385	pgmR3PhysUnlinkRamRange2(pVM, pRam, pRamPrev);
2386	pRam->cb = pRam->GCPhys = pRam->GCPhysLast = NIL_RTGCPHYS;
2387	MMHyperFree(pVM, pRam);
2388	break;
2389	}
2390	}
2391
2392	/*
2393	* Range match? It will all be within one range (see PGMAllHandler.cpp).
2394	*/
2395	if ( GCPhysLast >= pRam->GCPhys
2396	&& GCPhys <= pRam->GCPhysLast)
2397	{
2398	Assert(GCPhys >= pRam->GCPhys);
2399	Assert(GCPhysLast <= pRam->GCPhysLast);
2400
2401	/*
2402	* Turn the pages back into RAM pages.
2403	*/
2404	uint32_t iPage = (GCPhys - pRam->GCPhys) >> PAGE_SHIFT;
2405	uint32_t cLeft = cb >> PAGE_SHIFT;
2406	while (cLeft--)
2407	{
2408	PPGMPAGE pPage = &pRam->aPages[iPage];
2409	AssertMsg( (PGM_PAGE_IS_MMIO(pPage) && PGM_PAGE_IS_ZERO(pPage))
2410	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_MMIO2_ALIAS_MMIO
2411	\|\| PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_SPECIAL_ALIAS_MMIO,
2412	("%RGp %R[pgmpage]\n", pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), pPage));
2413	if (PGM_PAGE_IS_MMIO_OR_ALIAS(pPage))
2414	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_RAM);
2415	}
2416	break;
2417	}
2418
2419	/* next */
2420	pRamPrev = pRam;
2421	pRam = pRam->pNextR3;
2422	}
2423	}
2424
2425	/* Force a PGM pool flush as guest ram references have been changed. */
2426	/** @todo Not entirely SMP safe; assuming for now the guest takes care of
2427	* this internally (not touch mapped mmio while changing the mapping). */
2428	PVMCPU pVCpu = VMMGetCpu(pVM);
2429	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2430	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2431
2432	pgmPhysInvalidatePageMapTLB(pVM);
2433	pgmPhysInvalidRamRangeTlbs(pVM);
2434	pgmUnlock(pVM);
2435	return rc;
2436	}
2437
2438
2439	/**
2440	* Locate a MMIO2 range.
2441	*
2442	* @returns Pointer to the MMIO2 range.
2443	* @param pVM The cross context VM structure.
2444	* @param pDevIns The device instance owning the region.
2445	* @param iRegion The region.
2446	*/
2447	DECLINLINE(PPGMMMIO2RANGE) pgmR3PhysMMIO2Find(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion)
2448	{
2449	/*
2450	* Search the list.
2451	*/
2452	for (PPGMMMIO2RANGE pCur = pVM->pgm.s.pMmio2RangesR3; pCur; pCur = pCur->pNextR3)
2453	if ( pCur->pDevInsR3 == pDevIns
2454	&& pCur->iRegion == iRegion)
2455	return pCur;
2456	return NULL;
2457	}
2458
2459
2460	/**
2461	* Allocate and register an MMIO2 region.
2462	*
2463	* As mentioned elsewhere, MMIO2 is just RAM spelled differently. It's RAM
2464	* associated with a device. It is also non-shared memory with a permanent
2465	* ring-3 mapping and page backing (presently).
2466	*
2467	* A MMIO2 range may overlap with base memory if a lot of RAM is configured for
2468	* the VM, in which case we'll drop the base memory pages. Presently we will
2469	* make no attempt to preserve anything that happens to be present in the base
2470	* memory that is replaced, this is of course incorrect but it's too much
2471	* effort.
2472	*
2473	* @returns VBox status code.
2474	* @retval VINF_SUCCESS on success, *ppv pointing to the R3 mapping of the
2475	* memory.
2476	* @retval VERR_ALREADY_EXISTS if the region already exists.
2477	*
2478	* @param pVM The cross context VM structure.
2479	* @param pDevIns The device instance owning the region.
2480	* @param iRegion The region number. If the MMIO2 memory is a PCI
2481	* I/O region this number has to be the number of that
2482	* region. Otherwise it can be any number safe
2483	* UINT8_MAX.
2484	* @param cb The size of the region. Must be page aligned.
2485	* @param fFlags Reserved for future use, must be zero.
2486	* @param ppv Where to store the pointer to the ring-3 mapping of
2487	* the memory.
2488	* @param pszDesc The description.
2489	*/
2490	VMMR3DECL(int) PGMR3PhysMMIO2Register(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS cb, uint32_t fFlags,
2491	void *ppv, const char pszDesc)
2492	{
2493	/*
2494	* Validate input.
2495	*/
2496	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2497	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2498	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
2499	AssertPtrReturn(ppv, VERR_INVALID_POINTER);
2500	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
2501	AssertReturn(*pszDesc, VERR_INVALID_PARAMETER);
2502	AssertReturn(pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion) == NULL, VERR_ALREADY_EXISTS);
2503	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2504	AssertReturn(cb, VERR_INVALID_PARAMETER);
2505	AssertReturn(!fFlags, VERR_INVALID_PARAMETER);
2506
2507	const uint32_t cPages = cb >> PAGE_SHIFT;
2508	AssertLogRelReturn(((RTGCPHYS)cPages << PAGE_SHIFT) == cb, VERR_INVALID_PARAMETER);
2509	AssertLogRelReturn(cPages <= PGM_MMIO2_MAX_PAGE_COUNT, VERR_NO_MEMORY);
2510
2511	/*
2512	* For the 2nd+ instance, mangle the description string so it's unique.
2513	*/
2514	if (pDevIns->iInstance > 0) /** @todo Move to PDMDevHlp.cpp and use a real string cache. */
2515	{
2516	pszDesc = MMR3HeapAPrintf(pVM, MM_TAG_PGM_PHYS, "%s [%u]", pszDesc, pDevIns->iInstance);
2517	if (!pszDesc)
2518	return VERR_NO_MEMORY;
2519	}
2520
2521	/*
2522	* Allocate an MMIO2 range ID (not freed on failure).
2523	* The zero ID is not used as it could be confused with NIL_GMM_PAGEID.
2524	*/
2525	pgmLock(pVM);
2526	uint8_t idMmio2 = pVM->pgm.s.cMmio2Regions + 1;
2527	if (idMmio2 > PGM_MMIO2_MAX_RANGES)
2528	{
2529	pgmUnlock(pVM);
2530	AssertLogRelFailedReturn(VERR_PGM_TOO_MANY_MMIO2_RANGES);
2531	}
2532	pVM->pgm.s.cMmio2Regions = idMmio2;
2533	pgmUnlock(pVM);
2534
2535	/*
2536	* Try reserve and allocate the backing memory first as this is what is
2537	* most likely to fail.
2538	*/
2539	int rc = MMR3AdjustFixedReservation(pVM, cPages, pszDesc);
2540	if (RT_SUCCESS(rc))
2541	{
2542	PSUPPAGE paPages = (PSUPPAGE)RTMemTmpAlloc(cPages * sizeof(SUPPAGE));
2543	if (RT_SUCCESS(rc))
2544	{
2545	void *pvPages;
2546	rc = SUPR3PageAllocEx(cPages, 0 /fFlags/, &pvPages, NULL /pR0Ptr/, paPages);
2547	if (RT_SUCCESS(rc))
2548	{
2549	memset(pvPages, 0, cPages * PAGE_SIZE);
2550
2551	/*
2552	* Create the MMIO2 range record for it.
2553	*/
2554	const size_t cbRange = RT_OFFSETOF(PGMMMIO2RANGE, RamRange.aPages[cPages]);
2555	PPGMMMIO2RANGE pNew;
2556	rc = MMR3HyperAllocOnceNoRel(pVM, cbRange, 0, MM_TAG_PGM_PHYS, (void **)&pNew);
2557	AssertLogRelMsgRC(rc, ("cbRamRange=%zu\n", cbRange));
2558	if (RT_SUCCESS(rc))
2559	{
2560	pNew->pDevInsR3 = pDevIns;
2561	pNew->pvR3 = pvPages;
2562	//pNew->pNext = NULL;
2563	//pNew->fMapped = false;
2564	//pNew->fOverlapping = false;
2565	pNew->iRegion = iRegion;
2566	pNew->idSavedState = UINT8_MAX;
2567	pNew->idMmio2 = idMmio2;
2568	pNew->RamRange.pSelfR0 = MMHyperCCToR0(pVM, &pNew->RamRange);
2569	pNew->RamRange.pSelfRC = MMHyperCCToRC(pVM, &pNew->RamRange);
2570	pNew->RamRange.GCPhys = NIL_RTGCPHYS;
2571	pNew->RamRange.GCPhysLast = NIL_RTGCPHYS;
2572	pNew->RamRange.pszDesc = pszDesc;
2573	pNew->RamRange.cb = cb;
2574	pNew->RamRange.fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_MMIO2;
2575	pNew->RamRange.pvR3 = pvPages;
2576	//pNew->RamRange.paLSPages = NULL;
2577
2578	uint32_t iPage = cPages;
2579	while (iPage-- > 0)
2580	{
2581	PGM_PAGE_INIT(&pNew->RamRange.aPages[iPage],
2582	paPages[iPage].Phys,
2583	PGM_MMIO2_PAGEID_MAKE(idMmio2, iPage),
2584	PGMPAGETYPE_MMIO2, PGM_PAGE_STATE_ALLOCATED);
2585	}
2586
2587	/* update page count stats */
2588	pVM->pgm.s.cAllPages += cPages;
2589	pVM->pgm.s.cPrivatePages += cPages;
2590
2591	/*
2592	* Link it into the list.
2593	* Since there is no particular order, just push it.
2594	*/
2595	/** @todo we can save us the linked list now, just search the lookup table... */
2596	pgmLock(pVM);
2597	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == NULL);
2598	Assert(pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] == NIL_RTR0PTR);
2599	pNew->pNextR3 = pVM->pgm.s.pMmio2RangesR3;
2600	pVM->pgm.s.pMmio2RangesR3 = pNew;
2601	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = pNew;
2602	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = MMHyperCCToR0(pVM, pNew);
2603	pgmUnlock(pVM);
2604
2605	*ppv = pvPages;
2606	RTMemTmpFree(paPages);
2607	pgmPhysInvalidatePageMapTLB(pVM);
2608	return VINF_SUCCESS;
2609	}
2610
2611	SUPR3PageFreeEx(pvPages, cPages);
2612	}
2613	}
2614	RTMemTmpFree(paPages);
2615	MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pszDesc);
2616	}
2617	if (pDevIns->iInstance > 0)
2618	MMR3HeapFree((void *)pszDesc);
2619	return rc;
2620	}
2621
2622
2623	/**
2624	* Deregisters and frees an MMIO2 region.
2625	*
2626	* Any physical (and virtual) access handlers registered for the region must
2627	* be deregistered before calling this function.
2628	*
2629	* @returns VBox status code.
2630	* @param pVM The cross context VM structure.
2631	* @param pDevIns The device instance owning the region.
2632	* @param iRegion The region. If it's UINT32_MAX it'll be a wildcard match.
2633	*/
2634	VMMR3DECL(int) PGMR3PhysMMIO2Deregister(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion)
2635	{
2636	/*
2637	* Validate input.
2638	*/
2639	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2640	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2641	AssertReturn(iRegion <= UINT8_MAX \|\| iRegion == UINT32_MAX, VERR_INVALID_PARAMETER);
2642
2643	pgmLock(pVM);
2644	int rc = VINF_SUCCESS;
2645	unsigned cFound = 0;
2646	PPGMMMIO2RANGE pPrev = NULL;
2647	PPGMMMIO2RANGE pCur = pVM->pgm.s.pMmio2RangesR3;
2648	while (pCur)
2649	{
2650	if ( pCur->pDevInsR3 == pDevIns
2651	&& ( iRegion == UINT32_MAX
2652	\|\| pCur->iRegion == iRegion))
2653	{
2654	cFound++;
2655
2656	/*
2657	* Unmap it if it's mapped.
2658	*/
2659	if (pCur->fMapped)
2660	{
2661	int rc2 = PGMR3PhysMMIO2Unmap(pVM, pCur->pDevInsR3, pCur->iRegion, pCur->RamRange.GCPhys);
2662	AssertRC(rc2);
2663	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
2664	rc = rc2;
2665	}
2666
2667	/*
2668	* Unlink it
2669	*/
2670	PPGMMMIO2RANGE pNext = pCur->pNextR3;
2671	if (pPrev)
2672	pPrev->pNextR3 = pNext;
2673	else
2674	pVM->pgm.s.pMmio2RangesR3 = pNext;
2675	pCur->pNextR3 = NULL;
2676
2677	uint8_t idMmio2 = pCur->idMmio2;
2678	Assert(pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] == pCur);
2679	pVM->pgm.s.apMmio2RangesR3[idMmio2 - 1] = NULL;
2680	pVM->pgm.s.apMmio2RangesR0[idMmio2 - 1] = NIL_RTR0PTR;
2681
2682	/*
2683	* Free the memory.
2684	*/
2685	int rc2 = SUPR3PageFreeEx(pCur->pvR3, pCur->RamRange.cb >> PAGE_SHIFT);
2686	AssertRC(rc2);
2687	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
2688	rc = rc2;
2689
2690	uint32_t const cPages = pCur->RamRange.cb >> PAGE_SHIFT;
2691	rc2 = MMR3AdjustFixedReservation(pVM, -(int32_t)cPages, pCur->RamRange.pszDesc);
2692	AssertRC(rc2);
2693	if (RT_FAILURE(rc2) && RT_SUCCESS(rc))
2694	rc = rc2;
2695
2696	/* we're leaking hyper memory here if done at runtime. */
2697	#ifdef VBOX_STRICT
2698	VMSTATE const enmState = VMR3GetState(pVM);
2699	AssertMsg( enmState == VMSTATE_POWERING_OFF
2700	\|\| enmState == VMSTATE_POWERING_OFF_LS
2701	\|\| enmState == VMSTATE_OFF
2702	\|\| enmState == VMSTATE_OFF_LS
2703	\|\| enmState == VMSTATE_DESTROYING
2704	\|\| enmState == VMSTATE_TERMINATED
2705	\|\| enmState == VMSTATE_CREATING
2706	, ("%s\n", VMR3GetStateName(enmState)));
2707	#endif
2708	/*rc = MMHyperFree(pVM, pCur);
2709	AssertRCReturn(rc, rc); - not safe, see the alloc call. */
2710
2711
2712	/* update page count stats */
2713	pVM->pgm.s.cAllPages -= cPages;
2714	pVM->pgm.s.cPrivatePages -= cPages;
2715
2716	/* next */
2717	pCur = pNext;
2718	}
2719	else
2720	{
2721	pPrev = pCur;
2722	pCur = pCur->pNextR3;
2723	}
2724	}
2725	pgmPhysInvalidatePageMapTLB(pVM);
2726	pgmUnlock(pVM);
2727	return !cFound && iRegion != UINT32_MAX ? VERR_NOT_FOUND : rc;
2728	}
2729
2730
2731	/**
2732	* Maps a MMIO2 region.
2733	*
2734	* This is done when a guest / the bios / state loading changes the
2735	* PCI config. The replacing of base memory has the same restrictions
2736	* as during registration, of course.
2737	*
2738	* @returns VBox status code.
2739	*
2740	* @param pVM The cross context VM structure.
2741	* @param pDevIns The device instance owning the region.
2742	* @param iRegion The index of the registered region.
2743	* @param GCPhys The guest-physical address to be remapped.
2744	*/
2745	VMMR3DECL(int) PGMR3PhysMMIO2Map(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS GCPhys)
2746	{
2747	/*
2748	* Validate input
2749	*/
2750	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2751	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2752	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
2753	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
2754	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
2755	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2756
2757	PPGMMMIO2RANGE pCur = pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion);
2758	AssertReturn(pCur, VERR_NOT_FOUND);
2759	AssertReturn(!pCur->fMapped, VERR_WRONG_ORDER);
2760	Assert(pCur->RamRange.GCPhys == NIL_RTGCPHYS);
2761	Assert(pCur->RamRange.GCPhysLast == NIL_RTGCPHYS);
2762
2763	const RTGCPHYS GCPhysLast = GCPhys + pCur->RamRange.cb - 1;
2764	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
2765
2766	/*
2767	* Find our location in the ram range list, checking for
2768	* restriction we don't bother implementing yet (partially overlapping).
2769	*/
2770	bool fRamExists = false;
2771	PPGMRAMRANGE pRamPrev = NULL;
2772	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2773	while (pRam && GCPhysLast >= pRam->GCPhys)
2774	{
2775	if ( GCPhys <= pRam->GCPhysLast
2776	&& GCPhysLast >= pRam->GCPhys)
2777	{
2778	/* completely within? */
2779	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
2780	&& GCPhysLast <= pRam->GCPhysLast,
2781	("%RGp-%RGp (MMIO2/%s) falls partly outside %RGp-%RGp (%s)\n",
2782	GCPhys, GCPhysLast, pCur->RamRange.pszDesc,
2783	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
2784	VERR_PGM_RAM_CONFLICT);
2785	fRamExists = true;
2786	break;
2787	}
2788
2789	/* next */
2790	pRamPrev = pRam;
2791	pRam = pRam->pNextR3;
2792	}
2793	if (fRamExists)
2794	{
2795	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2796	uint32_t cPagesLeft = pCur->RamRange.cb >> PAGE_SHIFT;
2797	while (cPagesLeft-- > 0)
2798	{
2799	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
2800	("%RGp isn't a RAM page (%d) - mapping %RGp-%RGp (MMIO2/%s).\n",
2801	GCPhys, PGM_PAGE_GET_TYPE(pPage), GCPhys, GCPhysLast, pCur->RamRange.pszDesc),
2802	VERR_PGM_RAM_CONFLICT);
2803	pPage++;
2804	}
2805	}
2806	Log(("PGMR3PhysMMIO2Map: %RGp-%RGp fRamExists=%RTbool %s\n",
2807	GCPhys, GCPhysLast, fRamExists, pCur->RamRange.pszDesc));
2808
2809	/*
2810	* Make the changes.
2811	*/
2812	pgmLock(pVM);
2813
2814	pCur->RamRange.GCPhys = GCPhys;
2815	pCur->RamRange.GCPhysLast = GCPhysLast;
2816	pCur->fMapped = true;
2817	pCur->fOverlapping = fRamExists;
2818
2819	if (fRamExists)
2820	{
2821	/** @todo use pgmR3PhysFreePageRange here. */
2822	uint32_t cPendingPages = 0;
2823	PGMMFREEPAGESREQ pReq;
2824	int rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
2825	AssertLogRelRCReturn(rc, rc);
2826
2827	/* replace the pages, freeing all present RAM pages. */
2828	PPGMPAGE pPageSrc = &pCur->RamRange.aPages[0];
2829	PPGMPAGE pPageDst = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2830	uint32_t cPagesLeft = pCur->RamRange.cb >> PAGE_SHIFT;
2831	while (cPagesLeft-- > 0)
2832	{
2833	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, pPageDst, GCPhys);
2834	AssertLogRelRCReturn(rc, rc); /* We're done for if this goes wrong. */
2835
2836	RTHCPHYS const HCPhys = PGM_PAGE_GET_HCPHYS(pPageSrc);
2837	uint32_t const idPage = PGM_PAGE_GET_PAGEID(pPageSrc);
2838	PGM_PAGE_SET_PAGEID(pVM, pPageDst, idPage);
2839	PGM_PAGE_SET_HCPHYS(pVM, pPageDst, HCPhys);
2840	PGM_PAGE_SET_TYPE(pVM, pPageDst, PGMPAGETYPE_MMIO2);
2841	PGM_PAGE_SET_STATE(pVM, pPageDst, PGM_PAGE_STATE_ALLOCATED);
2842	PGM_PAGE_SET_PDE_TYPE(pVM, pPageDst, PGM_PAGE_PDE_TYPE_DONTCARE);
2843	PGM_PAGE_SET_PTE_INDEX(pVM, pPageDst, 0);
2844	PGM_PAGE_SET_TRACKING(pVM, pPageDst, 0);
2845
2846	pVM->pgm.s.cZeroPages--;
2847	GCPhys += PAGE_SIZE;
2848	pPageSrc++;
2849	pPageDst++;
2850	}
2851
2852	/* Flush physical page map TLB. */
2853	pgmPhysInvalidatePageMapTLB(pVM);
2854
2855	if (cPendingPages)
2856	{
2857	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
2858	AssertLogRelRCReturn(rc, rc);
2859	}
2860	GMMR3FreePagesCleanup(pReq);
2861
2862	/* Force a PGM pool flush as guest ram references have been changed. */
2863	/** @todo not entirely SMP safe; assuming for now the guest takes care of
2864	* this internally (not touch mapped mmio while changing the mapping). */
2865	PVMCPU pVCpu = VMMGetCpu(pVM);
2866	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2867	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2868
2869	pgmUnlock(pVM);
2870	}
2871	else
2872	{
2873	#ifdef VBOX_WITH_REM
2874	RTGCPHYS cb = pCur->RamRange.cb;
2875	#endif
2876
2877	/* Clear the tracking data of pages we're going to reactivate. */
2878	PPGMPAGE pPageSrc = &pCur->RamRange.aPages[0];
2879	uint32_t cPagesLeft = pCur->RamRange.cb >> PAGE_SHIFT;
2880	while (cPagesLeft-- > 0)
2881	{
2882	PGM_PAGE_SET_TRACKING(pVM, pPageSrc, 0);
2883	PGM_PAGE_SET_PTE_INDEX(pVM, pPageSrc, 0);
2884	pPageSrc++;
2885	}
2886
2887	/* link in the ram range */
2888	pgmR3PhysLinkRamRange(pVM, &pCur->RamRange, pRamPrev);
2889	pgmUnlock(pVM);
2890
2891	#ifdef VBOX_WITH_REM
2892	REMR3NotifyPhysRamRegister(pVM, GCPhys, cb, REM_NOTIFY_PHYS_RAM_FLAGS_MMIO2);
2893	#endif
2894	}
2895
2896	pgmPhysInvalidatePageMapTLB(pVM);
2897	return VINF_SUCCESS;
2898	}
2899
2900
2901	/**
2902	* Unmaps a MMIO2 region.
2903	*
2904	* This is done when a guest / the bios / state loading changes the
2905	* PCI config. The replacing of base memory has the same restrictions
2906	* as during registration, of course.
2907	*/
2908	VMMR3DECL(int) PGMR3PhysMMIO2Unmap(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS GCPhys)
2909	{
2910	/*
2911	* Validate input
2912	*/
2913	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
2914	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
2915	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
2916	AssertReturn(GCPhys != NIL_RTGCPHYS, VERR_INVALID_PARAMETER);
2917	AssertReturn(GCPhys != 0, VERR_INVALID_PARAMETER);
2918	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
2919
2920	PPGMMMIO2RANGE pCur = pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion);
2921	AssertReturn(pCur, VERR_NOT_FOUND);
2922	AssertReturn(pCur->fMapped, VERR_WRONG_ORDER);
2923	AssertReturn(pCur->RamRange.GCPhys == GCPhys, VERR_INVALID_PARAMETER);
2924	Assert(pCur->RamRange.GCPhysLast != NIL_RTGCPHYS);
2925
2926	Log(("PGMR3PhysMMIO2Unmap: %RGp-%RGp %s\n",
2927	pCur->RamRange.GCPhys, pCur->RamRange.GCPhysLast, pCur->RamRange.pszDesc));
2928
2929	/*
2930	* Unmap it.
2931	*/
2932	pgmLock(pVM);
2933
2934	#ifdef VBOX_WITH_REM
2935	RTGCPHYS GCPhysRangeREM;
2936	RTGCPHYS cbRangeREM;
2937	bool fInformREM;
2938	#endif
2939	if (pCur->fOverlapping)
2940	{
2941	/* Restore the RAM pages we've replaced. */
2942	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
2943	while (pRam->GCPhys > pCur->RamRange.GCPhysLast)
2944	pRam = pRam->pNextR3;
2945
2946	PPGMPAGE pPageDst = &pRam->aPages[(pCur->RamRange.GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
2947	uint32_t cPagesLeft = pCur->RamRange.cb >> PAGE_SHIFT;
2948	while (cPagesLeft-- > 0)
2949	{
2950	PGM_PAGE_INIT_ZERO(pPageDst, pVM, PGMPAGETYPE_RAM);
2951	pVM->pgm.s.cZeroPages++;
2952	pPageDst++;
2953	}
2954
2955	/* Flush physical page map TLB. */
2956	pgmPhysInvalidatePageMapTLB(pVM);
2957	#ifdef VBOX_WITH_REM
2958	GCPhysRangeREM = NIL_RTGCPHYS; /* shuts up gcc */
2959	cbRangeREM = RTGCPHYS_MAX; /* ditto */
2960	fInformREM = false;
2961	#endif
2962	}
2963	else
2964	{
2965	#ifdef VBOX_WITH_REM
2966	GCPhysRangeREM = pCur->RamRange.GCPhys;
2967	cbRangeREM = pCur->RamRange.cb;
2968	fInformREM = true;
2969	#endif
2970	pgmR3PhysUnlinkRamRange(pVM, &pCur->RamRange);
2971	}
2972
2973	pCur->RamRange.GCPhys = NIL_RTGCPHYS;
2974	pCur->RamRange.GCPhysLast = NIL_RTGCPHYS;
2975	pCur->fOverlapping = false;
2976	pCur->fMapped = false;
2977
2978	/* Force a PGM pool flush as guest ram references have been changed. */
2979	/** @todo not entirely SMP safe; assuming for now the guest takes care
2980	* of this internally (not touch mapped mmio while changing the
2981	* mapping). */
2982	PVMCPU pVCpu = VMMGetCpu(pVM);
2983	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_CLEAR_PGM_POOL;
2984	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
2985
2986	pgmPhysInvalidatePageMapTLB(pVM);
2987	pgmPhysInvalidRamRangeTlbs(pVM);
2988	pgmUnlock(pVM);
2989
2990	#ifdef VBOX_WITH_REM
2991	if (fInformREM)
2992	REMR3NotifyPhysRamDeregister(pVM, GCPhysRangeREM, cbRangeREM);
2993	#endif
2994
2995	return VINF_SUCCESS;
2996	}
2997
2998
2999	/**
3000	* Checks if the given address is an MMIO2 base address or not.
3001	*
3002	* @returns true/false accordingly.
3003	* @param pVM The cross context VM structure.
3004	* @param pDevIns The owner of the memory, optional.
3005	* @param GCPhys The address to check.
3006	*/
3007	VMMR3DECL(bool) PGMR3PhysMMIO2IsBase(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys)
3008	{
3009	/*
3010	* Validate input
3011	*/
3012	VM_ASSERT_EMT_RETURN(pVM, false);
3013	AssertPtrReturn(pDevIns, false);
3014	AssertReturn(GCPhys != NIL_RTGCPHYS, false);
3015	AssertReturn(GCPhys != 0, false);
3016	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), false);
3017
3018	/*
3019	* Search the list.
3020	*/
3021	pgmLock(pVM);
3022	for (PPGMMMIO2RANGE pCur = pVM->pgm.s.pMmio2RangesR3; pCur; pCur = pCur->pNextR3)
3023	if (pCur->RamRange.GCPhys == GCPhys)
3024	{
3025	Assert(pCur->fMapped);
3026	pgmUnlock(pVM);
3027	return true;
3028	}
3029	pgmUnlock(pVM);
3030	return false;
3031	}
3032
3033
3034	/**
3035	* Gets the HC physical address of a page in the MMIO2 region.
3036	*
3037	* This is API is intended for MMHyper and shouldn't be called
3038	* by anyone else...
3039	*
3040	* @returns VBox status code.
3041	* @param pVM The cross context VM structure.
3042	* @param pDevIns The owner of the memory, optional.
3043	* @param iRegion The region.
3044	* @param off The page expressed an offset into the MMIO2 region.
3045	* @param pHCPhys Where to store the result.
3046	*/
3047	VMMR3DECL(int) PGMR3PhysMMIO2GetHCPhys(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS off, PRTHCPHYS pHCPhys)
3048	{
3049	/*
3050	* Validate input
3051	*/
3052	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3053	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3054	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3055
3056	pgmLock(pVM);
3057	PPGMMMIO2RANGE pCur = pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion);
3058	AssertReturn(pCur, VERR_NOT_FOUND);
3059	AssertReturn(off < pCur->RamRange.cb, VERR_INVALID_PARAMETER);
3060
3061	PCPGMPAGE pPage = &pCur->RamRange.aPages[off >> PAGE_SHIFT];
3062	*pHCPhys = PGM_PAGE_GET_HCPHYS(pPage);
3063	pgmUnlock(pVM);
3064	return VINF_SUCCESS;
3065	}
3066
3067
3068	/**
3069	* Maps a portion of an MMIO2 region into kernel space (host).
3070	*
3071	* The kernel mapping will become invalid when the MMIO2 memory is deregistered
3072	* or the VM is terminated.
3073	*
3074	* @return VBox status code.
3075	*
3076	* @param pVM The cross context VM structure.
3077	* @param pDevIns The device owning the MMIO2 memory.
3078	* @param iRegion The region.
3079	* @param off The offset into the region. Must be page aligned.
3080	* @param cb The number of bytes to map. Must be page aligned.
3081	* @param pszDesc Mapping description.
3082	* @param pR0Ptr Where to store the R0 address.
3083	*/
3084	VMMR3DECL(int) PGMR3PhysMMIO2MapKernel(PVM pVM, PPDMDEVINS pDevIns, uint32_t iRegion, RTGCPHYS off, RTGCPHYS cb,
3085	const char *pszDesc, PRTR0PTR pR0Ptr)
3086	{
3087	/*
3088	* Validate input.
3089	*/
3090	VM_ASSERT_EMT_RETURN(pVM, VERR_VM_THREAD_NOT_EMT);
3091	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3092	AssertReturn(iRegion <= UINT8_MAX, VERR_INVALID_PARAMETER);
3093
3094	PPGMMMIO2RANGE pCur = pgmR3PhysMMIO2Find(pVM, pDevIns, iRegion);
3095	AssertReturn(pCur, VERR_NOT_FOUND);
3096	AssertReturn(off < pCur->RamRange.cb, VERR_INVALID_PARAMETER);
3097	AssertReturn(cb <= pCur->RamRange.cb, VERR_INVALID_PARAMETER);
3098	AssertReturn(off + cb <= pCur->RamRange.cb, VERR_INVALID_PARAMETER);
3099	NOREF(pszDesc);
3100
3101	/*
3102	* Pass the request on to the support library/driver.
3103	*/
3104	int rc = SUPR3PageMapKernel(pCur->pvR3, off, cb, 0, pR0Ptr);
3105
3106	return rc;
3107	}
3108
3109
3110	/**
3111	* Worker for PGMR3PhysRomRegister.
3112	*
3113	* This is here to simplify lock management, i.e. the caller does all the
3114	* locking and we can simply return without needing to remember to unlock
3115	* anything first.
3116	*
3117	* @returns VBox status code.
3118	* @param pVM The cross context VM structure.
3119	* @param pDevIns The device instance owning the ROM.
3120	* @param GCPhys First physical address in the range.
3121	* Must be page aligned!
3122	* @param cb The size of the range (in bytes).
3123	* Must be page aligned!
3124	* @param pvBinary Pointer to the binary data backing the ROM image.
3125	* @param cbBinary The size of the binary data pvBinary points to.
3126	* This must be less or equal to @a cb.
3127	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
3128	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
3129	* @param pszDesc Pointer to description string. This must not be freed.
3130	*/
3131	static int pgmR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
3132	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
3133	{
3134	/*
3135	* Validate input.
3136	*/
3137	AssertPtrReturn(pDevIns, VERR_INVALID_PARAMETER);
3138	AssertReturn(RT_ALIGN_T(GCPhys, PAGE_SIZE, RTGCPHYS) == GCPhys, VERR_INVALID_PARAMETER);
3139	AssertReturn(RT_ALIGN_T(cb, PAGE_SIZE, RTGCPHYS) == cb, VERR_INVALID_PARAMETER);
3140	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
3141	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3142	AssertPtrReturn(pvBinary, VERR_INVALID_PARAMETER);
3143	AssertPtrReturn(pszDesc, VERR_INVALID_POINTER);
3144	AssertReturn(!(fFlags & ~(PGMPHYS_ROM_FLAGS_SHADOWED \| PGMPHYS_ROM_FLAGS_PERMANENT_BINARY)), VERR_INVALID_PARAMETER);
3145	VM_ASSERT_STATE_RETURN(pVM, VMSTATE_CREATING, VERR_VM_INVALID_VM_STATE);
3146
3147	const uint32_t cPages = cb >> PAGE_SHIFT;
3148
3149	/*
3150	* Find the ROM location in the ROM list first.
3151	*/
3152	PPGMROMRANGE pRomPrev = NULL;
3153	PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3;
3154	while (pRom && GCPhysLast >= pRom->GCPhys)
3155	{
3156	if ( GCPhys <= pRom->GCPhysLast
3157	&& GCPhysLast >= pRom->GCPhys)
3158	AssertLogRelMsgFailedReturn(("%RGp-%RGp (%s) conflicts with existing %RGp-%RGp (%s)\n",
3159	GCPhys, GCPhysLast, pszDesc,
3160	pRom->GCPhys, pRom->GCPhysLast, pRom->pszDesc),
3161	VERR_PGM_RAM_CONFLICT);
3162	/* next */
3163	pRomPrev = pRom;
3164	pRom = pRom->pNextR3;
3165	}
3166
3167	/*
3168	* Find the RAM location and check for conflicts.
3169	*
3170	* Conflict detection is a bit different than for RAM
3171	* registration since a ROM can be located within a RAM
3172	* range. So, what we have to check for is other memory
3173	* types (other than RAM that is) and that we don't span
3174	* more than one RAM range (layz).
3175	*/
3176	bool fRamExists = false;
3177	PPGMRAMRANGE pRamPrev = NULL;
3178	PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
3179	while (pRam && GCPhysLast >= pRam->GCPhys)
3180	{
3181	if ( GCPhys <= pRam->GCPhysLast
3182	&& GCPhysLast >= pRam->GCPhys)
3183	{
3184	/* completely within? */
3185	AssertLogRelMsgReturn( GCPhys >= pRam->GCPhys
3186	&& GCPhysLast <= pRam->GCPhysLast,
3187	("%RGp-%RGp (%s) falls partly outside %RGp-%RGp (%s)\n",
3188	GCPhys, GCPhysLast, pszDesc,
3189	pRam->GCPhys, pRam->GCPhysLast, pRam->pszDesc),
3190	VERR_PGM_RAM_CONFLICT);
3191	fRamExists = true;
3192	break;
3193	}
3194
3195	/* next */
3196	pRamPrev = pRam;
3197	pRam = pRam->pNextR3;
3198	}
3199	if (fRamExists)
3200	{
3201	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3202	uint32_t cPagesLeft = cPages;
3203	while (cPagesLeft-- > 0)
3204	{
3205	AssertLogRelMsgReturn(PGM_PAGE_GET_TYPE(pPage) == PGMPAGETYPE_RAM,
3206	("%RGp (%R[pgmpage]) isn't a RAM page - registering %RGp-%RGp (%s).\n",
3207	pRam->GCPhys + ((RTGCPHYS)(uintptr_t)(pPage - &pRam->aPages[0]) << PAGE_SHIFT),
3208	pPage, GCPhys, GCPhysLast, pszDesc), VERR_PGM_RAM_CONFLICT);
3209	Assert(PGM_PAGE_IS_ZERO(pPage));
3210	pPage++;
3211	}
3212	}
3213
3214	/*
3215	* Update the base memory reservation if necessary.
3216	*/
3217	uint32_t cExtraBaseCost = fRamExists ? 0 : cPages;
3218	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
3219	cExtraBaseCost += cPages;
3220	if (cExtraBaseCost)
3221	{
3222	int rc = MMR3IncreaseBaseReservation(pVM, cExtraBaseCost);
3223	if (RT_FAILURE(rc))
3224	return rc;
3225	}
3226
3227	/*
3228	* Allocate memory for the virgin copy of the RAM.
3229	*/
3230	PGMMALLOCATEPAGESREQ pReq;
3231	int rc = GMMR3AllocatePagesPrepare(pVM, &pReq, cPages, GMMACCOUNT_BASE);
3232	AssertRCReturn(rc, rc);
3233
3234	for (uint32_t iPage = 0; iPage < cPages; iPage++)
3235	{
3236	pReq->aPages[iPage].HCPhysGCPhys = GCPhys + (iPage << PAGE_SHIFT);
3237	pReq->aPages[iPage].idPage = NIL_GMM_PAGEID;
3238	pReq->aPages[iPage].idSharedPage = NIL_GMM_PAGEID;
3239	}
3240
3241	rc = GMMR3AllocatePagesPerform(pVM, pReq);
3242	if (RT_FAILURE(rc))
3243	{
3244	GMMR3AllocatePagesCleanup(pReq);
3245	return rc;
3246	}
3247
3248	/*
3249	* Allocate the new ROM range and RAM range (if necessary).
3250	*/
3251	PPGMROMRANGE pRomNew;
3252	rc = MMHyperAlloc(pVM, RT_OFFSETOF(PGMROMRANGE, aPages[cPages]), 0, MM_TAG_PGM_PHYS, (void **)&pRomNew);
3253	if (RT_SUCCESS(rc))
3254	{
3255	PPGMRAMRANGE pRamNew = NULL;
3256	if (!fRamExists)
3257	rc = MMHyperAlloc(pVM, RT_OFFSETOF(PGMRAMRANGE, aPages[cPages]), sizeof(PGMPAGE), MM_TAG_PGM_PHYS, (void **)&pRamNew);
3258	if (RT_SUCCESS(rc))
3259	{
3260	/*
3261	* Initialize and insert the RAM range (if required).
3262	*/
3263	PPGMROMPAGE pRomPage = &pRomNew->aPages[0];
3264	if (!fRamExists)
3265	{
3266	pRamNew->pSelfR0 = MMHyperCCToR0(pVM, pRamNew);
3267	pRamNew->pSelfRC = MMHyperCCToRC(pVM, pRamNew);
3268	pRamNew->GCPhys = GCPhys;
3269	pRamNew->GCPhysLast = GCPhysLast;
3270	pRamNew->cb = cb;
3271	pRamNew->pszDesc = pszDesc;
3272	pRamNew->fFlags = PGM_RAM_RANGE_FLAGS_AD_HOC_ROM;
3273	pRamNew->pvR3 = NULL;
3274	pRamNew->paLSPages = NULL;
3275
3276	PPGMPAGE pPage = &pRamNew->aPages[0];
3277	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
3278	{
3279	PGM_PAGE_INIT(pPage,
3280	pReq->aPages[iPage].HCPhysGCPhys,
3281	pReq->aPages[iPage].idPage,
3282	PGMPAGETYPE_ROM,
3283	PGM_PAGE_STATE_ALLOCATED);
3284
3285	pRomPage->Virgin = *pPage;
3286	}
3287
3288	pVM->pgm.s.cAllPages += cPages;
3289	pgmR3PhysLinkRamRange(pVM, pRamNew, pRamPrev);
3290	}
3291	else
3292	{
3293	PPGMPAGE pPage = &pRam->aPages[(GCPhys - pRam->GCPhys) >> PAGE_SHIFT];
3294	for (uint32_t iPage = 0; iPage < cPages; iPage++, pPage++, pRomPage++)
3295	{
3296	PGM_PAGE_SET_TYPE(pVM, pPage, PGMPAGETYPE_ROM);
3297	PGM_PAGE_SET_HCPHYS(pVM, pPage, pReq->aPages[iPage].HCPhysGCPhys);
3298	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
3299	PGM_PAGE_SET_PAGEID(pVM, pPage, pReq->aPages[iPage].idPage);
3300	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
3301	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
3302	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
3303
3304	pRomPage->Virgin = *pPage;
3305	}
3306
3307	pRamNew = pRam;
3308
3309	pVM->pgm.s.cZeroPages -= cPages;
3310	}
3311	pVM->pgm.s.cPrivatePages += cPages;
3312
3313	/* Flush physical page map TLB. */
3314	pgmPhysInvalidatePageMapTLB(pVM);
3315
3316
3317	/*
3318	* !HACK ALERT! REM + (Shadowed) ROM ==> mess.
3319	*
3320	* If it's shadowed we'll register the handler after the ROM notification
3321	* so we get the access handler callbacks that we should. If it isn't
3322	* shadowed we'll do it the other way around to make REM use the built-in
3323	* ROM behavior and not the handler behavior (which is to route all access
3324	* to PGM atm).
3325	*/
3326	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
3327	{
3328	#ifdef VBOX_WITH_REM
3329	REMR3NotifyPhysRomRegister(pVM, GCPhys, cb, NULL, true /* fShadowed */);
3330	#endif
3331	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, pVM->pgm.s.hRomPhysHandlerType,
3332	pRomNew, MMHyperCCToR0(pVM, pRomNew), MMHyperCCToRC(pVM, pRomNew),
3333	pszDesc);
3334	}
3335	else
3336	{
3337	rc = PGMHandlerPhysicalRegister(pVM, GCPhys, GCPhysLast, pVM->pgm.s.hRomPhysHandlerType,
3338	pRomNew, MMHyperCCToR0(pVM, pRomNew), MMHyperCCToRC(pVM, pRomNew),
3339	pszDesc);
3340	#ifdef VBOX_WITH_REM
3341	REMR3NotifyPhysRomRegister(pVM, GCPhys, cb, NULL, false /* fShadowed */);
3342	#endif
3343	}
3344	if (RT_SUCCESS(rc))
3345	{
3346	/*
3347	* Copy the image over to the virgin pages.
3348	* This must be done after linking in the RAM range.
3349	*/
3350	size_t cbBinaryLeft = cbBinary;
3351	PPGMPAGE pRamPage = &pRamNew->aPages[(GCPhys - pRamNew->GCPhys) >> PAGE_SHIFT];
3352	for (uint32_t iPage = 0; iPage < cPages; iPage++, pRamPage++)
3353	{
3354	void *pvDstPage;
3355	rc = pgmPhysPageMap(pVM, pRamPage, GCPhys + (iPage << PAGE_SHIFT), &pvDstPage);
3356	if (RT_FAILURE(rc))
3357	{
3358	VMSetError(pVM, rc, RT_SRC_POS, "Failed to map virgin ROM page at %RGp", GCPhys);
3359	break;
3360	}
3361	if (cbBinaryLeft >= PAGE_SIZE)
3362	{
3363	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), PAGE_SIZE);
3364	cbBinaryLeft -= PAGE_SIZE;
3365	}
3366	else
3367	{
3368	ASMMemZeroPage(pvDstPage); /* (shouldn't be necessary, but can't hurt either) */
3369	if (cbBinaryLeft > 0)
3370	{
3371	memcpy(pvDstPage, (uint8_t const *)pvBinary + ((size_t)iPage << PAGE_SHIFT), cbBinaryLeft);
3372	cbBinaryLeft = 0;
3373	}
3374	}
3375	}
3376	if (RT_SUCCESS(rc))
3377	{
3378	/*
3379	* Initialize the ROM range.
3380	* Note that the Virgin member of the pages has already been initialized above.
3381	*/
3382	pRomNew->GCPhys = GCPhys;
3383	pRomNew->GCPhysLast = GCPhysLast;
3384	pRomNew->cb = cb;
3385	pRomNew->fFlags = fFlags;
3386	pRomNew->idSavedState = UINT8_MAX;
3387	pRomNew->cbOriginal = cbBinary;
3388	pRomNew->pszDesc = pszDesc;
3389	pRomNew->pvOriginal = fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY
3390	? pvBinary : RTMemDup(pvBinary, cbBinary);
3391	if (pRomNew->pvOriginal)
3392	{
3393	for (unsigned iPage = 0; iPage < cPages; iPage++)
3394	{
3395	PPGMROMPAGE pPage = &pRomNew->aPages[iPage];
3396	pPage->enmProt = PGMROMPROT_READ_ROM_WRITE_IGNORE;
3397	PGM_PAGE_INIT_ZERO(&pPage->Shadow, pVM, PGMPAGETYPE_ROM_SHADOW);
3398	}
3399
3400	/* update the page count stats for the shadow pages. */
3401	if (fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
3402	{
3403	pVM->pgm.s.cZeroPages += cPages;
3404	pVM->pgm.s.cAllPages += cPages;
3405	}
3406
3407	/*
3408	* Insert the ROM range, tell REM and return successfully.
3409	*/
3410	pRomNew->pNextR3 = pRom;
3411	pRomNew->pNextR0 = pRom ? MMHyperCCToR0(pVM, pRom) : NIL_RTR0PTR;
3412	pRomNew->pNextRC = pRom ? MMHyperCCToRC(pVM, pRom) : NIL_RTRCPTR;
3413
3414	if (pRomPrev)
3415	{
3416	pRomPrev->pNextR3 = pRomNew;
3417	pRomPrev->pNextR0 = MMHyperCCToR0(pVM, pRomNew);
3418	pRomPrev->pNextRC = MMHyperCCToRC(pVM, pRomNew);
3419	}
3420	else
3421	{
3422	pVM->pgm.s.pRomRangesR3 = pRomNew;
3423	pVM->pgm.s.pRomRangesR0 = MMHyperCCToR0(pVM, pRomNew);
3424	pVM->pgm.s.pRomRangesRC = MMHyperCCToRC(pVM, pRomNew);
3425	}
3426
3427	pgmPhysInvalidatePageMapTLB(pVM);
3428	GMMR3AllocatePagesCleanup(pReq);
3429	return VINF_SUCCESS;
3430	}
3431
3432	/* bail out */
3433	rc = VERR_NO_MEMORY;
3434	}
3435
3436	int rc2 = PGMHandlerPhysicalDeregister(pVM, GCPhys);
3437	AssertRC(rc2);
3438	}
3439
3440	if (!fRamExists)
3441	{
3442	pgmR3PhysUnlinkRamRange2(pVM, pRamNew, pRamPrev);
3443	MMHyperFree(pVM, pRamNew);
3444	}
3445	}
3446	MMHyperFree(pVM, pRomNew);
3447	}
3448
3449	/** @todo Purge the mapping cache or something... */
3450	GMMR3FreeAllocatedPages(pVM, pReq);
3451	GMMR3AllocatePagesCleanup(pReq);
3452	return rc;
3453	}
3454
3455
3456	/**
3457	* Registers a ROM image.
3458	*
3459	* Shadowed ROM images requires double the amount of backing memory, so,
3460	* don't use that unless you have to. Shadowing of ROM images is process
3461	* where we can select where the reads go and where the writes go. On real
3462	* hardware the chipset provides means to configure this. We provide
3463	* PGMR3PhysProtectROM() for this purpose.
3464	*
3465	* A read-only copy of the ROM image will always be kept around while we
3466	* will allocate RAM pages for the changes on demand (unless all memory
3467	* is configured to be preallocated).
3468	*
3469	* @returns VBox status code.
3470	* @param pVM The cross context VM structure.
3471	* @param pDevIns The device instance owning the ROM.
3472	* @param GCPhys First physical address in the range.
3473	* Must be page aligned!
3474	* @param cb The size of the range (in bytes).
3475	* Must be page aligned!
3476	* @param pvBinary Pointer to the binary data backing the ROM image.
3477	* @param cbBinary The size of the binary data pvBinary points to.
3478	* This must be less or equal to @a cb.
3479	* @param fFlags Mask of flags. PGMPHYS_ROM_FLAGS_SHADOWED
3480	* and/or PGMPHYS_ROM_FLAGS_PERMANENT_BINARY.
3481	* @param pszDesc Pointer to description string. This must not be freed.
3482	*
3483	* @remark There is no way to remove the rom, automatically on device cleanup or
3484	* manually from the device yet. This isn't difficult in any way, it's
3485	* just not something we expect to be necessary for a while.
3486	*/
3487	VMMR3DECL(int) PGMR3PhysRomRegister(PVM pVM, PPDMDEVINS pDevIns, RTGCPHYS GCPhys, RTGCPHYS cb,
3488	const void pvBinary, uint32_t cbBinary, uint32_t fFlags, const char pszDesc)
3489	{
3490	Log(("PGMR3PhysRomRegister: pDevIns=%p GCPhys=%RGp(-%RGp) cb=%RGp pvBinary=%p cbBinary=%#x fFlags=%#x pszDesc=%s\n",
3491	pDevIns, GCPhys, GCPhys + cb, cb, pvBinary, cbBinary, fFlags, pszDesc));
3492	pgmLock(pVM);
3493	int rc = pgmR3PhysRomRegister(pVM, pDevIns, GCPhys, cb, pvBinary, cbBinary, fFlags, pszDesc);
3494	pgmUnlock(pVM);
3495	return rc;
3496	}
3497
3498
3499	/**
3500	* Called by PGMR3MemSetup to reset the shadow, switch to the virgin, and verify
3501	* that the virgin part is untouched.
3502	*
3503	* This is done after the normal memory has been cleared.
3504	*
3505	* ASSUMES that the caller owns the PGM lock.
3506	*
3507	* @param pVM The cross context VM structure.
3508	*/
3509	int pgmR3PhysRomReset(PVM pVM)
3510	{
3511	PGM_LOCK_ASSERT_OWNER(pVM);
3512	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
3513	{
3514	const uint32_t cPages = pRom->cb >> PAGE_SHIFT;
3515
3516	if (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED)
3517	{
3518	/*
3519	* Reset the physical handler.
3520	*/
3521	int rc = PGMR3PhysRomProtect(pVM, pRom->GCPhys, pRom->cb, PGMROMPROT_READ_ROM_WRITE_IGNORE);
3522	AssertRCReturn(rc, rc);
3523
3524	/*
3525	* What we do with the shadow pages depends on the memory
3526	* preallocation option. If not enabled, we'll just throw
3527	* out all the dirty pages and replace them by the zero page.
3528	*/
3529	if (!pVM->pgm.s.fRamPreAlloc)
3530	{
3531	/* Free the dirty pages. */
3532	uint32_t cPendingPages = 0;
3533	PGMMFREEPAGESREQ pReq;
3534	rc = GMMR3FreePagesPrepare(pVM, &pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
3535	AssertRCReturn(rc, rc);
3536
3537	for (uint32_t iPage = 0; iPage < cPages; iPage++)
3538	if ( !PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow)
3539	&& !PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow))
3540	{
3541	Assert(PGM_PAGE_GET_STATE(&pRom->aPages[iPage].Shadow) == PGM_PAGE_STATE_ALLOCATED);
3542	rc = pgmPhysFreePage(pVM, pReq, &cPendingPages, &pRom->aPages[iPage].Shadow,
3543	pRom->GCPhys + (iPage << PAGE_SHIFT));
3544	AssertLogRelRCReturn(rc, rc);
3545	}
3546
3547	if (cPendingPages)
3548	{
3549	rc = GMMR3FreePagesPerform(pVM, pReq, cPendingPages);
3550	AssertLogRelRCReturn(rc, rc);
3551	}
3552	GMMR3FreePagesCleanup(pReq);
3553	}
3554	else
3555	{
3556	/* clear all the shadow pages. */
3557	for (uint32_t iPage = 0; iPage < cPages; iPage++)
3558	{
3559	if (PGM_PAGE_IS_ZERO(&pRom->aPages[iPage].Shadow))
3560	continue;
3561	Assert(!PGM_PAGE_IS_BALLOONED(&pRom->aPages[iPage].Shadow));
3562	void *pvDstPage;
3563	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
3564	rc = pgmPhysPageMakeWritableAndMap(pVM, &pRom->aPages[iPage].Shadow, GCPhys, &pvDstPage);
3565	if (RT_FAILURE(rc))
3566	break;
3567	ASMMemZeroPage(pvDstPage);
3568	}
3569	AssertRCReturn(rc, rc);
3570	}
3571	}
3572
3573	/*
3574	* Restore the original ROM pages after a saved state load.
3575	* Also, in strict builds check that ROM pages remain unmodified.
3576	*/
3577	#ifndef VBOX_STRICT
3578	if (pVM->pgm.s.fRestoreRomPagesOnReset)
3579	#endif
3580	{
3581	size_t cbSrcLeft = pRom->cbOriginal;
3582	uint8_t const pbSrcPage = (uint8_t const )pRom->pvOriginal;
3583	uint32_t cRestored = 0;
3584	for (uint32_t iPage = 0; iPage < cPages && cbSrcLeft > 0; iPage++, pbSrcPage += PAGE_SIZE)
3585	{
3586	const RTGCPHYS GCPhys = pRom->GCPhys + (iPage << PAGE_SHIFT);
3587	void const *pvDstPage;
3588	int rc = pgmPhysPageMapReadOnly(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPage);
3589	if (RT_FAILURE(rc))
3590	break;
3591
3592	if (memcmp(pvDstPage, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE)))
3593	{
3594	if (pVM->pgm.s.fRestoreRomPagesOnReset)
3595	{
3596	void *pvDstPageW;
3597	rc = pgmPhysPageMap(pVM, &pRom->aPages[iPage].Virgin, GCPhys, &pvDstPageW);
3598	AssertLogRelRCReturn(rc, rc);
3599	memcpy(pvDstPageW, pbSrcPage, RT_MIN(cbSrcLeft, PAGE_SIZE));
3600	cRestored++;
3601	}
3602	else
3603	LogRel(("pgmR3PhysRomReset: %RGp: ROM page changed (%s)\n", GCPhys, pRom->pszDesc));
3604	}
3605	cbSrcLeft -= RT_MIN(cbSrcLeft, PAGE_SIZE);
3606	}
3607	if (cRestored > 0)
3608	LogRel(("PGM: ROM \"%s\": Reloaded %u of %u pages.\n", pRom->pszDesc, cRestored, cPages));
3609	}
3610	}
3611
3612	/* Clear the ROM restore flag now as we only need to do this once after
3613	loading saved state. */
3614	pVM->pgm.s.fRestoreRomPagesOnReset = false;
3615
3616	return VINF_SUCCESS;
3617	}
3618
3619
3620	/**
3621	* Called by PGMR3Term to free resources.
3622	*
3623	* ASSUMES that the caller owns the PGM lock.
3624	*
3625	* @param pVM The cross context VM structure.
3626	*/
3627	void pgmR3PhysRomTerm(PVM pVM)
3628	{
3629	/*
3630	* Free the heap copy of the original bits.
3631	*/
3632	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
3633	{
3634	if ( pRom->pvOriginal
3635	&& !(pRom->fFlags & PGMPHYS_ROM_FLAGS_PERMANENT_BINARY))
3636	{
3637	RTMemFree((void *)pRom->pvOriginal);
3638	pRom->pvOriginal = NULL;
3639	}
3640	}
3641	}
3642
3643
3644	/**
3645	* Change the shadowing of a range of ROM pages.
3646	*
3647	* This is intended for implementing chipset specific memory registers
3648	* and will not be very strict about the input. It will silently ignore
3649	* any pages that are not the part of a shadowed ROM.
3650	*
3651	* @returns VBox status code.
3652	* @retval VINF_PGM_SYNC_CR3
3653	*
3654	* @param pVM The cross context VM structure.
3655	* @param GCPhys Where to start. Page aligned.
3656	* @param cb How much to change. Page aligned.
3657	* @param enmProt The new ROM protection.
3658	*/
3659	VMMR3DECL(int) PGMR3PhysRomProtect(PVM pVM, RTGCPHYS GCPhys, RTGCPHYS cb, PGMROMPROT enmProt)
3660	{
3661	/*
3662	* Check input
3663	*/
3664	if (!cb)
3665	return VINF_SUCCESS;
3666	AssertReturn(!(GCPhys & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3667	AssertReturn(!(cb & PAGE_OFFSET_MASK), VERR_INVALID_PARAMETER);
3668	RTGCPHYS GCPhysLast = GCPhys + (cb - 1);
3669	AssertReturn(GCPhysLast > GCPhys, VERR_INVALID_PARAMETER);
3670	AssertReturn(enmProt >= PGMROMPROT_INVALID && enmProt <= PGMROMPROT_END, VERR_INVALID_PARAMETER);
3671
3672	/*
3673	* Process the request.
3674	*/
3675	pgmLock(pVM);
3676	int rc = VINF_SUCCESS;
3677	bool fFlushTLB = false;
3678	for (PPGMROMRANGE pRom = pVM->pgm.s.pRomRangesR3; pRom; pRom = pRom->pNextR3)
3679	{
3680	if ( GCPhys <= pRom->GCPhysLast
3681	&& GCPhysLast >= pRom->GCPhys
3682	&& (pRom->fFlags & PGMPHYS_ROM_FLAGS_SHADOWED))
3683	{
3684	/*
3685	* Iterate the relevant pages and make necessary the changes.
3686	*/
3687	bool fChanges = false;
3688	uint32_t const cPages = pRom->GCPhysLast <= GCPhysLast
3689	? pRom->cb >> PAGE_SHIFT
3690	: (GCPhysLast - pRom->GCPhys + 1) >> PAGE_SHIFT;
3691	for (uint32_t iPage = (GCPhys - pRom->GCPhys) >> PAGE_SHIFT;
3692	iPage < cPages;
3693	iPage++)
3694	{
3695	PPGMROMPAGE pRomPage = &pRom->aPages[iPage];
3696	if (PGMROMPROT_IS_ROM(pRomPage->enmProt) != PGMROMPROT_IS_ROM(enmProt))
3697	{
3698	fChanges = true;
3699
3700	/* flush references to the page. */
3701	PPGMPAGE pRamPage = pgmPhysGetPage(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT));
3702	int rc2 = pgmPoolTrackUpdateGCPhys(pVM, pRom->GCPhys + (iPage << PAGE_SHIFT), pRamPage,
3703	true /fFlushPTEs/, &fFlushTLB);
3704	if (rc2 != VINF_SUCCESS && (rc == VINF_SUCCESS \|\| RT_FAILURE(rc2)))
3705	rc = rc2;
3706
3707	PPGMPAGE pOld = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Virgin : &pRomPage->Shadow;
3708	PPGMPAGE pNew = PGMROMPROT_IS_ROM(pRomPage->enmProt) ? &pRomPage->Shadow : &pRomPage->Virgin;
3709
3710	pOld = pRamPage;
3711	pRamPage = pNew;
3712	/** @todo preserve the volatile flags (handlers) when these have been moved out of HCPhys! */
3713	}
3714	pRomPage->enmProt = enmProt;
3715	}
3716
3717	/*
3718	* Reset the access handler if we made changes, no need
3719	* to optimize this.
3720	*/
3721	if (fChanges)
3722	{
3723	int rc2 = PGMHandlerPhysicalReset(pVM, pRom->GCPhys);
3724	if (RT_FAILURE(rc2))
3725	{
3726	pgmUnlock(pVM);
3727	AssertRC(rc);
3728	return rc2;
3729	}
3730	}
3731
3732	/* Advance - cb isn't updated. */
3733	GCPhys = pRom->GCPhys + (cPages << PAGE_SHIFT);
3734	}
3735	}
3736	pgmUnlock(pVM);
3737	if (fFlushTLB)
3738	PGM_INVL_ALL_VCPU_TLBS(pVM);
3739
3740	return rc;
3741	}
3742
3743
3744	/**
3745	* Sets the Address Gate 20 state.
3746	*
3747	* @param pVCpu The cross context virtual CPU structure.
3748	* @param fEnable True if the gate should be enabled.
3749	* False if the gate should be disabled.
3750	*/
3751	VMMDECL(void) PGMR3PhysSetA20(PVMCPU pVCpu, bool fEnable)
3752	{
3753	LogFlow(("PGMR3PhysSetA20 %d (was %d)\n", fEnable, pVCpu->pgm.s.fA20Enabled));
3754	if (pVCpu->pgm.s.fA20Enabled != fEnable)
3755	{
3756	pVCpu->pgm.s.fA20Enabled = fEnable;
3757	pVCpu->pgm.s.GCPhysA20Mask = ~((RTGCPHYS)!fEnable << 20);
3758	#ifdef VBOX_WITH_REM
3759	REMR3A20Set(pVCpu->pVMR3, pVCpu, fEnable);
3760	#endif
3761	#ifdef PGM_WITH_A20
3762	pVCpu->pgm.s.fSyncFlags \|= PGM_SYNC_UPDATE_PAGE_BIT_VIRTUAL;
3763	VMCPU_FF_SET(pVCpu, VMCPU_FF_PGM_SYNC_CR3);
3764	pgmR3RefreshShadowModeAfterA20Change(pVCpu);
3765	HMFlushTLB(pVCpu);
3766	#endif
3767	IEMTlbInvalidateAllPhysical(pVCpu);
3768	STAM_REL_COUNTER_INC(&pVCpu->pgm.s.cA20Changes);
3769	}
3770	}
3771
3772
3773	/**
3774	* Tree enumeration callback for dealing with age rollover.
3775	* It will perform a simple compression of the current age.
3776	*/
3777	static DECLCALLBACK(int) pgmR3PhysChunkAgeingRolloverCallback(PAVLU32NODECORE pNode, void *pvUser)
3778	{
3779	/* Age compression - ASSUMES iNow == 4. */
3780	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
3781	if (pChunk->iLastUsed >= UINT32_C(0xffffff00))
3782	pChunk->iLastUsed = 3;
3783	else if (pChunk->iLastUsed >= UINT32_C(0xfffff000))
3784	pChunk->iLastUsed = 2;
3785	else if (pChunk->iLastUsed)
3786	pChunk->iLastUsed = 1;
3787	else /* iLastUsed = 0 */
3788	pChunk->iLastUsed = 4;
3789
3790	NOREF(pvUser);
3791	return 0;
3792	}
3793
3794
3795	/**
3796	* The structure passed in the pvUser argument of pgmR3PhysChunkUnmapCandidateCallback().
3797	*/
3798	typedef struct PGMR3PHYSCHUNKUNMAPCB
3799	{
3800	PVM pVM; /*< Pointer to the VM. /
3801	PPGMCHUNKR3MAP pChunk; /*< The chunk to unmap. /
3802	} PGMR3PHYSCHUNKUNMAPCB, *PPGMR3PHYSCHUNKUNMAPCB;
3803
3804
3805	/**
3806	* Callback used to find the mapping that's been unused for
3807	* the longest time.
3808	*/
3809	static DECLCALLBACK(int) pgmR3PhysChunkUnmapCandidateCallback(PAVLU32NODECORE pNode, void *pvUser)
3810	{
3811	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)pNode;
3812	PPGMR3PHYSCHUNKUNMAPCB pArg = (PPGMR3PHYSCHUNKUNMAPCB)pvUser;
3813
3814	/*
3815	* Check for locks and compare when last used.
3816	*/
3817	if (pChunk->cRefs)
3818	return 0;
3819	if (pChunk->cPermRefs)
3820	return 0;
3821	if ( pArg->pChunk
3822	&& pChunk->iLastUsed >= pArg->pChunk->iLastUsed)
3823	return 0;
3824
3825	/*
3826	* Check that it's not in any of the TLBs.
3827	*/
3828	PVM pVM = pArg->pVM;
3829	if ( pVM->pgm.s.ChunkR3Map.Tlb.aEntries[PGM_CHUNKR3MAPTLB_IDX(pChunk->Core.Key)].idChunk
3830	== pChunk->Core.Key)
3831	{
3832	pChunk = NULL;
3833	return 0;
3834	}
3835	#ifdef VBOX_STRICT
3836	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
3837	{
3838	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk != pChunk);
3839	Assert(pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk != pChunk->Core.Key);
3840	}
3841	#endif
3842
3843	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.PhysTlbHC.aEntries); i++)
3844	if (pVM->pgm.s.PhysTlbHC.aEntries[i].pMap == pChunk)
3845	return 0;
3846
3847	pArg->pChunk = pChunk;
3848	return 0;
3849	}
3850
3851
3852	/**
3853	* Finds a good candidate for unmapping when the ring-3 mapping cache is full.
3854	*
3855	* The candidate will not be part of any TLBs, so no need to flush
3856	* anything afterwards.
3857	*
3858	* @returns Chunk id.
3859	* @param pVM The cross context VM structure.
3860	*/
3861	static int32_t pgmR3PhysChunkFindUnmapCandidate(PVM pVM)
3862	{
3863	PGM_LOCK_ASSERT_OWNER(pVM);
3864
3865	/*
3866	* Enumerate the age tree starting with the left most node.
3867	*/
3868	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
3869	PGMR3PHYSCHUNKUNMAPCB Args;
3870	Args.pVM = pVM;
3871	Args.pChunk = NULL;
3872	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkUnmapCandidateCallback, &Args);
3873	Assert(Args.pChunk);
3874	if (Args.pChunk)
3875	{
3876	Assert(Args.pChunk->cRefs == 0);
3877	Assert(Args.pChunk->cPermRefs == 0);
3878	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
3879	return Args.pChunk->Core.Key;
3880	}
3881
3882	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkFindCandidate, a);
3883	return INT32_MAX;
3884	}
3885
3886
3887	/**
3888	* Rendezvous callback used by pgmR3PhysUnmapChunk that unmaps a chunk
3889	*
3890	* This is only called on one of the EMTs while the other ones are waiting for
3891	* it to complete this function.
3892	*
3893	* @returns VINF_SUCCESS (VBox strict status code).
3894	* @param pVM The cross context VM structure.
3895	* @param pVCpu The cross context virtual CPU structure of the calling EMT. Unused.
3896	* @param pvUser User pointer. Unused
3897	*
3898	*/
3899	static DECLCALLBACK(VBOXSTRICTRC) pgmR3PhysUnmapChunkRendezvous(PVM pVM, PVMCPU pVCpu, void *pvUser)
3900	{
3901	int rc = VINF_SUCCESS;
3902	pgmLock(pVM);
3903	NOREF(pVCpu); NOREF(pvUser);
3904
3905	if (pVM->pgm.s.ChunkR3Map.c >= pVM->pgm.s.ChunkR3Map.cMax)
3906	{
3907	/* Flush the pgm pool cache; call the internal rendezvous handler as we're already in a rendezvous handler here. */
3908	/** @todo also not really efficient to unmap a chunk that contains PD
3909	* or PT pages. */
3910	pgmR3PoolClearAllRendezvous(pVM, &pVM->aCpus[0], NULL /* no need to flush the REM TLB as we already did that above */);
3911
3912	/*
3913	* Request the ring-0 part to unmap a chunk to make space in the mapping cache.
3914	*/
3915	GMMMAPUNMAPCHUNKREQ Req;
3916	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
3917	Req.Hdr.cbReq = sizeof(Req);
3918	Req.pvR3 = NULL;
3919	Req.idChunkMap = NIL_GMM_CHUNKID;
3920	Req.idChunkUnmap = pgmR3PhysChunkFindUnmapCandidate(pVM);
3921	if (Req.idChunkUnmap != INT32_MAX)
3922	{
3923	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkUnmap, a);
3924	rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
3925	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkUnmap, a);
3926	if (RT_SUCCESS(rc))
3927	{
3928	/*
3929	* Remove the unmapped one.
3930	*/
3931	PPGMCHUNKR3MAP pUnmappedChunk = (PPGMCHUNKR3MAP)RTAvlU32Remove(&pVM->pgm.s.ChunkR3Map.pTree, Req.idChunkUnmap);
3932	AssertRelease(pUnmappedChunk);
3933	AssertRelease(!pUnmappedChunk->cRefs);
3934	AssertRelease(!pUnmappedChunk->cPermRefs);
3935	pUnmappedChunk->pv = NULL;
3936	pUnmappedChunk->Core.Key = UINT32_MAX;
3937	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
3938	MMR3HeapFree(pUnmappedChunk);
3939	#else
3940	MMR3UkHeapFree(pVM, pUnmappedChunk, MM_TAG_PGM_CHUNK_MAPPING);
3941	#endif
3942	pVM->pgm.s.ChunkR3Map.c--;
3943	pVM->pgm.s.cUnmappedChunks++;
3944
3945	/*
3946	* Flush dangling PGM pointers (R3 & R0 ptrs to GC physical addresses).
3947	*/
3948	/** @todo We should not flush chunks which include cr3 mappings. */
3949	for (VMCPUID idCpu = 0; idCpu < pVM->cCpus; idCpu++)
3950	{
3951	PPGMCPU pPGM = &pVM->aCpus[idCpu].pgm.s;
3952
3953	pPGM->pGst32BitPdR3 = NULL;
3954	pPGM->pGstPaePdptR3 = NULL;
3955	pPGM->pGstAmd64Pml4R3 = NULL;
3956	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
3957	pPGM->pGst32BitPdR0 = NIL_RTR0PTR;
3958	pPGM->pGstPaePdptR0 = NIL_RTR0PTR;
3959	pPGM->pGstAmd64Pml4R0 = NIL_RTR0PTR;
3960	#endif
3961	for (unsigned i = 0; i < RT_ELEMENTS(pPGM->apGstPaePDsR3); i++)
3962	{
3963	pPGM->apGstPaePDsR3[i] = NULL;
3964	#ifndef VBOX_WITH_2X_4GB_ADDR_SPACE
3965	pPGM->apGstPaePDsR0[i] = NIL_RTR0PTR;
3966	#endif
3967	}
3968
3969	/* Flush REM TLBs. */
3970	CPUMSetChangedFlags(&pVM->aCpus[idCpu], CPUM_CHANGED_GLOBAL_TLB_FLUSH);
3971	}
3972	#ifdef VBOX_WITH_REM
3973	/* Flush REM translation blocks. */
3974	REMFlushTBs(pVM);
3975	#endif
3976	}
3977	}
3978	}
3979	pgmUnlock(pVM);
3980	return rc;
3981	}
3982
3983	/**
3984	* Unmap a chunk to free up virtual address space (request packet handler for pgmR3PhysChunkMap)
3985	*
3986	* @returns VBox status code.
3987	* @param pVM The cross context VM structure.
3988	*/
3989	void pgmR3PhysUnmapChunk(PVM pVM)
3990	{
3991	int rc = VMMR3EmtRendezvous(pVM, VMMEMTRENDEZVOUS_FLAGS_TYPE_ONCE, pgmR3PhysUnmapChunkRendezvous, NULL);
3992	AssertRC(rc);
3993	}
3994
3995
3996	/**
3997	* Maps the given chunk into the ring-3 mapping cache.
3998	*
3999	* This will call ring-0.
4000	*
4001	* @returns VBox status code.
4002	* @param pVM The cross context VM structure.
4003	* @param idChunk The chunk in question.
4004	* @param ppChunk Where to store the chunk tracking structure.
4005	*
4006	* @remarks Called from within the PGM critical section.
4007	* @remarks Can be called from any thread!
4008	*/
4009	int pgmR3PhysChunkMap(PVM pVM, uint32_t idChunk, PPPGMCHUNKR3MAP ppChunk)
4010	{
4011	int rc;
4012
4013	PGM_LOCK_ASSERT_OWNER(pVM);
4014
4015	/*
4016	* Move the chunk time forward.
4017	*/
4018	pVM->pgm.s.ChunkR3Map.iNow++;
4019	if (pVM->pgm.s.ChunkR3Map.iNow == 0)
4020	{
4021	pVM->pgm.s.ChunkR3Map.iNow = 4;
4022	RTAvlU32DoWithAll(&pVM->pgm.s.ChunkR3Map.pTree, true /fFromLeft/, pgmR3PhysChunkAgeingRolloverCallback, NULL);
4023	}
4024
4025	/*
4026	* Allocate a new tracking structure first.
4027	*/
4028	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
4029	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3HeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk));
4030	#else
4031	PPGMCHUNKR3MAP pChunk = (PPGMCHUNKR3MAP)MMR3UkHeapAllocZ(pVM, MM_TAG_PGM_CHUNK_MAPPING, sizeof(*pChunk), NULL);
4032	#endif
4033	AssertReturn(pChunk, VERR_NO_MEMORY);
4034	pChunk->Core.Key = idChunk;
4035	pChunk->iLastUsed = pVM->pgm.s.ChunkR3Map.iNow;
4036
4037	/*
4038	* Request the ring-0 part to map the chunk in question.
4039	*/
4040	GMMMAPUNMAPCHUNKREQ Req;
4041	Req.Hdr.u32Magic = SUPVMMR0REQHDR_MAGIC;
4042	Req.Hdr.cbReq = sizeof(Req);
4043	Req.pvR3 = NULL;
4044	Req.idChunkMap = idChunk;
4045	Req.idChunkUnmap = NIL_GMM_CHUNKID;
4046
4047	/* Must be callable from any thread, so can't use VMMR3CallR0. */
4048	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkMap, a);
4049	rc = SUPR3CallVMMR0Ex(pVM->pVMR0, NIL_VMCPUID, VMMR0_DO_GMM_MAP_UNMAP_CHUNK, 0, &Req.Hdr);
4050	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatChunkMap, a);
4051	if (RT_SUCCESS(rc))
4052	{
4053	pChunk->pv = Req.pvR3;
4054
4055	/*
4056	* If we're running out of virtual address space, then we should
4057	* unmap another chunk.
4058	*
4059	* Currently, an unmap operation requires that all other virtual CPUs
4060	* are idling and not by chance making use of the memory we're
4061	* unmapping. So, we create an async unmap operation here.
4062	*
4063	* Now, when creating or restoring a saved state this wont work very
4064	* well since we may want to restore all guest RAM + a little something.
4065	* So, we have to do the unmap synchronously. Fortunately for us
4066	* though, during these operations the other virtual CPUs are inactive
4067	* and it should be safe to do this.
4068	*/
4069	/** @todo Eventually we should lock all memory when used and do
4070	* map+unmap as one kernel call without any rendezvous or
4071	* other precautions. */
4072	if (pVM->pgm.s.ChunkR3Map.c + 1 >= pVM->pgm.s.ChunkR3Map.cMax)
4073	{
4074	switch (VMR3GetState(pVM))
4075	{
4076	case VMSTATE_LOADING:
4077	case VMSTATE_SAVING:
4078	{
4079	PVMCPU pVCpu = VMMGetCpu(pVM);
4080	if ( pVCpu
4081	&& pVM->pgm.s.cDeprecatedPageLocks == 0)
4082	{
4083	pgmR3PhysUnmapChunkRendezvous(pVM, pVCpu, NULL);
4084	break;
4085	}
4086	/* fall thru */
4087	}
4088	default:
4089	rc = VMR3ReqCallNoWait(pVM, VMCPUID_ANY_QUEUE, (PFNRT)pgmR3PhysUnmapChunk, 1, pVM);
4090	AssertRC(rc);
4091	break;
4092	}
4093	}
4094
4095	/*
4096	* Update the tree. We must do this after any unmapping to make sure
4097	* the chunk we're going to return isn't unmapped by accident.
4098	*/
4099	AssertPtr(Req.pvR3);
4100	bool fRc = RTAvlU32Insert(&pVM->pgm.s.ChunkR3Map.pTree, &pChunk->Core);
4101	AssertRelease(fRc);
4102	pVM->pgm.s.ChunkR3Map.c++;
4103	pVM->pgm.s.cMappedChunks++;
4104	}
4105	else
4106	{
4107	/** @todo this may fail because of /proc/sys/vm/max_map_count, so we
4108	* should probably restrict ourselves on linux. */
4109	AssertRC(rc);
4110	#ifdef VBOX_WITH_2X_4GB_ADDR_SPACE
4111	MMR3HeapFree(pChunk);
4112	#else
4113	MMR3UkHeapFree(pVM, pChunk, MM_TAG_PGM_CHUNK_MAPPING);
4114	#endif
4115	pChunk = NULL;
4116	}
4117
4118	*ppChunk = pChunk;
4119	return rc;
4120	}
4121
4122
4123	/**
4124	* For VMMCALLRING3_PGM_MAP_CHUNK, considered internal.
4125	*
4126	* @returns see pgmR3PhysChunkMap.
4127	* @param pVM The cross context VM structure.
4128	* @param idChunk The chunk to map.
4129	*/
4130	VMMR3DECL(int) PGMR3PhysChunkMap(PVM pVM, uint32_t idChunk)
4131	{
4132	PPGMCHUNKR3MAP pChunk;
4133	int rc;
4134
4135	pgmLock(pVM);
4136	rc = pgmR3PhysChunkMap(pVM, idChunk, &pChunk);
4137	pgmUnlock(pVM);
4138	return rc;
4139	}
4140
4141
4142	/**
4143	* Invalidates the TLB for the ring-3 mapping cache.
4144	*
4145	* @param pVM The cross context VM structure.
4146	*/
4147	VMMR3DECL(void) PGMR3PhysChunkInvalidateTLB(PVM pVM)
4148	{
4149	pgmLock(pVM);
4150	for (unsigned i = 0; i < RT_ELEMENTS(pVM->pgm.s.ChunkR3Map.Tlb.aEntries); i++)
4151	{
4152	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].idChunk = NIL_GMM_CHUNKID;
4153	pVM->pgm.s.ChunkR3Map.Tlb.aEntries[i].pChunk = NULL;
4154	}
4155	/* The page map TLB references chunks, so invalidate that one too. */
4156	pgmPhysInvalidatePageMapTLB(pVM);
4157	pgmUnlock(pVM);
4158	}
4159
4160
4161	/**
4162	* Response to VMMCALLRING3_PGM_ALLOCATE_LARGE_HANDY_PAGE to allocate a large
4163	* (2MB) page for use with a nested paging PDE.
4164	*
4165	* @returns The following VBox status codes.
4166	* @retval VINF_SUCCESS on success.
4167	* @retval VINF_EM_NO_MEMORY if we're out of memory.
4168	*
4169	* @param pVM The cross context VM structure.
4170	* @param GCPhys GC physical start address of the 2 MB range
4171	*/
4172	VMMR3DECL(int) PGMR3PhysAllocateLargeHandyPage(PVM pVM, RTGCPHYS GCPhys)
4173	{
4174	#ifdef PGM_WITH_LARGE_PAGES
4175	uint64_t u64TimeStamp1, u64TimeStamp2;
4176
4177	pgmLock(pVM);
4178
4179	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatAllocLargePage, a);
4180	u64TimeStamp1 = RTTimeMilliTS();
4181	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_LARGE_HANDY_PAGE, 0, NULL);
4182	u64TimeStamp2 = RTTimeMilliTS();
4183	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatAllocLargePage, a);
4184	if (RT_SUCCESS(rc))
4185	{
4186	Assert(pVM->pgm.s.cLargeHandyPages == 1);
4187
4188	uint32_t idPage = pVM->pgm.s.aLargeHandyPage[0].idPage;
4189	RTHCPHYS HCPhys = pVM->pgm.s.aLargeHandyPage[0].HCPhysGCPhys;
4190
4191	void *pv;
4192
4193	/* Map the large page into our address space.
4194	*
4195	* Note: assuming that within the 2 MB range:
4196	* - GCPhys + PAGE_SIZE = HCPhys + PAGE_SIZE (whole point of this exercise)
4197	* - user space mapping is continuous as well
4198	* - page id (GCPhys) + 1 = page id (GCPhys + PAGE_SIZE)
4199	*/
4200	rc = pgmPhysPageMapByPageID(pVM, idPage, HCPhys, &pv);
4201	AssertLogRelMsg(RT_SUCCESS(rc), ("idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n", idPage, HCPhys, rc));
4202
4203	if (RT_SUCCESS(rc))
4204	{
4205	/*
4206	* Clear the pages.
4207	*/
4208	STAM_PROFILE_START(&pVM->pgm.s.CTX_SUFF(pStats)->StatClearLargePage, b);
4209	for (unsigned i = 0; i < _2M/PAGE_SIZE; i++)
4210	{
4211	ASMMemZeroPage(pv);
4212
4213	PPGMPAGE pPage;
4214	rc = pgmPhysGetPageEx(pVM, GCPhys, &pPage);
4215	AssertRC(rc);
4216
4217	Assert(PGM_PAGE_IS_ZERO(pPage));
4218	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatRZPageReplaceZero);
4219	pVM->pgm.s.cZeroPages--;
4220
4221	/*
4222	* Do the PGMPAGE modifications.
4223	*/
4224	pVM->pgm.s.cPrivatePages++;
4225	PGM_PAGE_SET_HCPHYS(pVM, pPage, HCPhys);
4226	PGM_PAGE_SET_PAGEID(pVM, pPage, idPage);
4227	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ALLOCATED);
4228	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_PDE);
4229	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
4230	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
4231
4232	/* Somewhat dirty assumption that page ids are increasing. */
4233	idPage++;
4234
4235	HCPhys += PAGE_SIZE;
4236	GCPhys += PAGE_SIZE;
4237
4238	pv = (void *)((uintptr_t)pv + PAGE_SIZE);
4239
4240	Log3(("PGMR3PhysAllocateLargePage: idPage=%#x HCPhys=%RGp\n", idPage, HCPhys));
4241	}
4242	STAM_PROFILE_STOP(&pVM->pgm.s.CTX_SUFF(pStats)->StatClearLargePage, b);
4243
4244	/* Flush all TLBs. */
4245	PGM_INVL_ALL_VCPU_TLBS(pVM);
4246	pgmPhysInvalidatePageMapTLB(pVM);
4247	}
4248	pVM->pgm.s.cLargeHandyPages = 0;
4249	}
4250
4251	if (RT_SUCCESS(rc))
4252	{
4253	static uint32_t cTimeOut = 0;
4254	uint64_t u64TimeStampDelta = u64TimeStamp2 - u64TimeStamp1;
4255
4256	if (u64TimeStampDelta > 100)
4257	{
4258	STAM_COUNTER_INC(&pVM->pgm.s.CTX_SUFF(pStats)->StatLargePageOverflow);
4259	if ( ++cTimeOut > 10
4260	\|\| u64TimeStampDelta > 1000 /* more than one second forces an early retirement from allocating large pages. */)
4261	{
4262	/* If repeated attempts to allocate a large page takes more than 100 ms, then we fall back to normal 4k pages.
4263	* E.g. Vista 64 tries to move memory around, which takes a huge amount of time.
4264	*/
4265	LogRel(("PGMR3PhysAllocateLargePage: allocating large pages takes too long (last attempt %d ms; nr of timeouts %d); DISABLE\n", u64TimeStampDelta, cTimeOut));
4266	PGMSetLargePageUsage(pVM, false);
4267	}
4268	}
4269	else
4270	if (cTimeOut > 0)
4271	cTimeOut--;
4272	}
4273
4274	pgmUnlock(pVM);
4275	return rc;
4276	#else
4277	RT_NOREF(pVM, GCPhys);
4278	return VERR_NOT_IMPLEMENTED;
4279	#endif /* PGM_WITH_LARGE_PAGES */
4280	}
4281
4282
4283	/**
4284	* Response to VM_FF_PGM_NEED_HANDY_PAGES and VMMCALLRING3_PGM_ALLOCATE_HANDY_PAGES.
4285	*
4286	* This function will also work the VM_FF_PGM_NO_MEMORY force action flag, to
4287	* signal and clear the out of memory condition. When contracted, this API is
4288	* used to try clear the condition when the user wants to resume.
4289	*
4290	* @returns The following VBox status codes.
4291	* @retval VINF_SUCCESS on success. FFs cleared.
4292	* @retval VINF_EM_NO_MEMORY if we're out of memory. The FF is not cleared in
4293	* this case and it gets accompanied by VM_FF_PGM_NO_MEMORY.
4294	*
4295	* @param pVM The cross context VM structure.
4296	*
4297	* @remarks The VINF_EM_NO_MEMORY status is for the benefit of the FF processing
4298	* in EM.cpp and shouldn't be propagated outside TRPM, HM, EM and
4299	* pgmPhysEnsureHandyPage. There is one exception to this in the \#PF
4300	* handler.
4301	*/
4302	VMMR3DECL(int) PGMR3PhysAllocateHandyPages(PVM pVM)
4303	{
4304	pgmLock(pVM);
4305
4306	/*
4307	* Allocate more pages, noting down the index of the first new page.
4308	*/
4309	uint32_t iClear = pVM->pgm.s.cHandyPages;
4310	AssertMsgReturn(iClear <= RT_ELEMENTS(pVM->pgm.s.aHandyPages), ("%d", iClear), VERR_PGM_HANDY_PAGE_IPE);
4311	Log(("PGMR3PhysAllocateHandyPages: %d -> %d\n", iClear, RT_ELEMENTS(pVM->pgm.s.aHandyPages)));
4312	int rcAlloc = VINF_SUCCESS;
4313	int rcSeed = VINF_SUCCESS;
4314	int rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
4315	while (rc == VERR_GMM_SEED_ME)
4316	{
4317	void *pvChunk;
4318	rcAlloc = rc = SUPR3PageAlloc(GMM_CHUNK_SIZE >> PAGE_SHIFT, &pvChunk);
4319	if (RT_SUCCESS(rc))
4320	{
4321	rcSeed = rc = VMMR3CallR0(pVM, VMMR0_DO_GMM_SEED_CHUNK, (uintptr_t)pvChunk, NULL);
4322	if (RT_FAILURE(rc))
4323	SUPR3PageFree(pvChunk, GMM_CHUNK_SIZE >> PAGE_SHIFT);
4324	}
4325	if (RT_SUCCESS(rc))
4326	rc = VMMR3CallR0(pVM, VMMR0_DO_PGM_ALLOCATE_HANDY_PAGES, 0, NULL);
4327	}
4328
4329	/** @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) */
4330	if ( rc == VERR_GMM_HIT_VM_ACCOUNT_LIMIT
4331	&& pVM->pgm.s.cHandyPages > 0)
4332	{
4333	/* Still handy pages left, so don't panic. */
4334	rc = VINF_SUCCESS;
4335	}
4336
4337	if (RT_SUCCESS(rc))
4338	{
4339	AssertMsg(rc == VINF_SUCCESS, ("%Rrc\n", rc));
4340	Assert(pVM->pgm.s.cHandyPages > 0);
4341	VM_FF_CLEAR(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
4342	VM_FF_CLEAR(pVM, VM_FF_PGM_NO_MEMORY);
4343
4344	#ifdef VBOX_STRICT
4345	uint32_t i;
4346	for (i = iClear; i < pVM->pgm.s.cHandyPages; i++)
4347	if ( pVM->pgm.s.aHandyPages[i].idPage == NIL_GMM_PAGEID
4348	\|\| pVM->pgm.s.aHandyPages[i].idSharedPage != NIL_GMM_PAGEID
4349	\|\| (pVM->pgm.s.aHandyPages[i].HCPhysGCPhys & PAGE_OFFSET_MASK))
4350	break;
4351	if (i != pVM->pgm.s.cHandyPages)
4352	{
4353	RTAssertMsg1Weak(NULL, __LINE__, __FILE__, __FUNCTION__);
4354	RTAssertMsg2Weak("i=%d iClear=%d cHandyPages=%d\n", i, iClear, pVM->pgm.s.cHandyPages);
4355	for (uint32_t j = iClear; j < pVM->pgm.s.cHandyPages; j++)
4356	RTAssertMsg2Add("%03d: idPage=%d HCPhysGCPhys=%RHp idSharedPage=%d%\n", j,
4357	pVM->pgm.s.aHandyPages[j].idPage,
4358	pVM->pgm.s.aHandyPages[j].HCPhysGCPhys,
4359	pVM->pgm.s.aHandyPages[j].idSharedPage,
4360	j == i ? " <---" : "");
4361	RTAssertPanic();
4362	}
4363	#endif
4364	/*
4365	* Clear the pages.
4366	*/
4367	while (iClear < pVM->pgm.s.cHandyPages)
4368	{
4369	PGMMPAGEDESC pPage = &pVM->pgm.s.aHandyPages[iClear];
4370	void *pv;
4371	rc = pgmPhysPageMapByPageID(pVM, pPage->idPage, pPage->HCPhysGCPhys, &pv);
4372	AssertLogRelMsgBreak(RT_SUCCESS(rc),
4373	("%u/%u: idPage=%#x HCPhysGCPhys=%RHp rc=%Rrc\n",
4374	iClear, pVM->pgm.s.cHandyPages, pPage->idPage, pPage->HCPhysGCPhys, rc));
4375	ASMMemZeroPage(pv);
4376	iClear++;
4377	Log3(("PGMR3PhysAllocateHandyPages: idPage=%#x HCPhys=%RGp\n", pPage->idPage, pPage->HCPhysGCPhys));
4378	}
4379	}
4380	else
4381	{
4382	uint64_t cAllocPages, cMaxPages, cBalloonPages;
4383
4384	/*
4385	* We should never get here unless there is a genuine shortage of
4386	* memory (or some internal error). Flag the error so the VM can be
4387	* suspended ASAP and the user informed. If we're totally out of
4388	* handy pages we will return failure.
4389	*/
4390	/* Report the failure. */
4391	LogRel(("PGM: Failed to procure handy pages; rc=%Rrc rcAlloc=%Rrc rcSeed=%Rrc cHandyPages=%#x\n"
4392	" cAllPages=%#x cPrivatePages=%#x cSharedPages=%#x cZeroPages=%#x\n",
4393	rc, rcAlloc, rcSeed,
4394	pVM->pgm.s.cHandyPages,
4395	pVM->pgm.s.cAllPages,
4396	pVM->pgm.s.cPrivatePages,
4397	pVM->pgm.s.cSharedPages,
4398	pVM->pgm.s.cZeroPages));
4399
4400	if (GMMR3QueryMemoryStats(pVM, &cAllocPages, &cMaxPages, &cBalloonPages) == VINF_SUCCESS)
4401	{
4402	LogRel(("GMM: Statistics:\n"
4403	" Allocated pages: %RX64\n"
4404	" Maximum pages: %RX64\n"
4405	" Ballooned pages: %RX64\n", cAllocPages, cMaxPages, cBalloonPages));
4406	}
4407
4408	if ( rc != VERR_NO_MEMORY
4409	&& rc != VERR_NO_PHYS_MEMORY
4410	&& rc != VERR_LOCK_FAILED)
4411	{
4412	for (uint32_t i = 0; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
4413	{
4414	LogRel(("PGM: aHandyPages[#%#04x] = {.HCPhysGCPhys=%RHp, .idPage=%#08x, .idSharedPage=%#08x}\n",
4415	i, pVM->pgm.s.aHandyPages[i].HCPhysGCPhys, pVM->pgm.s.aHandyPages[i].idPage,
4416	pVM->pgm.s.aHandyPages[i].idSharedPage));
4417	uint32_t const idPage = pVM->pgm.s.aHandyPages[i].idPage;
4418	if (idPage != NIL_GMM_PAGEID)
4419	{
4420	for (PPGMRAMRANGE pRam = pVM->pgm.s.pRamRangesXR3;
4421	pRam;
4422	pRam = pRam->pNextR3)
4423	{
4424	uint32_t const cPages = pRam->cb >> PAGE_SHIFT;
4425	for (uint32_t iPage = 0; iPage < cPages; iPage++)
4426	if (PGM_PAGE_GET_PAGEID(&pRam->aPages[iPage]) == idPage)
4427	LogRel(("PGM: Used by %RGp %R[pgmpage] (%s)\n",
4428	pRam->GCPhys + ((RTGCPHYS)iPage << PAGE_SHIFT), &pRam->aPages[iPage], pRam->pszDesc));
4429	}
4430	}
4431	}
4432	}
4433
4434	if (rc == VERR_NO_MEMORY)
4435	{
4436	uint64_t cbHostRamAvail = 0;
4437	int rc2 = RTSystemQueryAvailableRam(&cbHostRamAvail);
4438	if (RT_SUCCESS(rc2))
4439	LogRel(("Host RAM: %RU64MB available\n", cbHostRamAvail / _1M));
4440	else
4441	LogRel(("Cannot determine the amount of available host memory\n"));
4442	}
4443
4444	/* Set the FFs and adjust rc. */
4445	VM_FF_SET(pVM, VM_FF_PGM_NEED_HANDY_PAGES);
4446	VM_FF_SET(pVM, VM_FF_PGM_NO_MEMORY);
4447	if ( rc == VERR_NO_MEMORY
4448	\|\| rc == VERR_NO_PHYS_MEMORY
4449	\|\| rc == VERR_LOCK_FAILED)
4450	rc = VINF_EM_NO_MEMORY;
4451	}
4452
4453	pgmUnlock(pVM);
4454	return rc;
4455	}
4456
4457
4458	/**
4459	* Frees the specified RAM page and replaces it with the ZERO page.
4460	*
4461	* This is used by ballooning, remapping MMIO2, RAM reset and state loading.
4462	*
4463	* @param pVM The cross context VM structure.
4464	* @param pReq Pointer to the request.
4465	* @param pcPendingPages Where the number of pages waiting to be freed are
4466	* kept. This will normally be incremented.
4467	* @param pPage Pointer to the page structure.
4468	* @param GCPhys The guest physical address of the page, if applicable.
4469	*
4470	* @remarks The caller must own the PGM lock.
4471	*/
4472	int pgmPhysFreePage(PVM pVM, PGMMFREEPAGESREQ pReq, uint32_t *pcPendingPages, PPGMPAGE pPage, RTGCPHYS GCPhys)
4473	{
4474	/*
4475	* Assert sanity.
4476	*/
4477	PGM_LOCK_ASSERT_OWNER(pVM);
4478	if (RT_UNLIKELY( PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_RAM
4479	&& PGM_PAGE_GET_TYPE(pPage) != PGMPAGETYPE_ROM_SHADOW))
4480	{
4481	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
4482	return VMSetError(pVM, VERR_PGM_PHYS_NOT_RAM, RT_SRC_POS, "GCPhys=%RGp type=%d", GCPhys, PGM_PAGE_GET_TYPE(pPage));
4483	}
4484
4485	/** @todo What about ballooning of large pages??! */
4486	Assert( PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE
4487	&& PGM_PAGE_GET_PDE_TYPE(pPage) != PGM_PAGE_PDE_TYPE_PDE_DISABLED);
4488
4489	if ( PGM_PAGE_IS_ZERO(pPage)
4490	\|\| PGM_PAGE_IS_BALLOONED(pPage))
4491	return VINF_SUCCESS;
4492
4493	const uint32_t idPage = PGM_PAGE_GET_PAGEID(pPage);
4494	Log3(("pgmPhysFreePage: idPage=%#x GCPhys=%RGp pPage=%R[pgmpage]\n", idPage, GCPhys, pPage));
4495	if (RT_UNLIKELY( idPage == NIL_GMM_PAGEID
4496	\|\| idPage > GMM_PAGEID_LAST
4497	\|\| PGM_PAGE_GET_CHUNKID(pPage) == NIL_GMM_CHUNKID))
4498	{
4499	AssertMsgFailed(("GCPhys=%RGp pPage=%R[pgmpage]\n", GCPhys, pPage));
4500	return VMSetError(pVM, VERR_PGM_PHYS_INVALID_PAGE_ID, RT_SRC_POS, "GCPhys=%RGp idPage=%#x", GCPhys, pPage);
4501	}
4502
4503	/* update page count stats. */
4504	if (PGM_PAGE_IS_SHARED(pPage))
4505	pVM->pgm.s.cSharedPages--;
4506	else
4507	pVM->pgm.s.cPrivatePages--;
4508	pVM->pgm.s.cZeroPages++;
4509
4510	/* Deal with write monitored pages. */
4511	if (PGM_PAGE_GET_STATE(pPage) == PGM_PAGE_STATE_WRITE_MONITORED)
4512	{
4513	PGM_PAGE_SET_WRITTEN_TO(pVM, pPage);
4514	pVM->pgm.s.cWrittenToPages++;
4515	}
4516
4517	/*
4518	* pPage = ZERO page.
4519	*/
4520	PGM_PAGE_SET_HCPHYS(pVM, pPage, pVM->pgm.s.HCPhysZeroPg);
4521	PGM_PAGE_SET_STATE(pVM, pPage, PGM_PAGE_STATE_ZERO);
4522	PGM_PAGE_SET_PAGEID(pVM, pPage, NIL_GMM_PAGEID);
4523	PGM_PAGE_SET_PDE_TYPE(pVM, pPage, PGM_PAGE_PDE_TYPE_DONTCARE);
4524	PGM_PAGE_SET_PTE_INDEX(pVM, pPage, 0);
4525	PGM_PAGE_SET_TRACKING(pVM, pPage, 0);
4526
4527	/* Flush physical page map TLB entry. */
4528	pgmPhysInvalidatePageMapTLBEntry(pVM, GCPhys);
4529
4530	/*
4531	* Make sure it's not in the handy page array.
4532	*/
4533	for (uint32_t i = pVM->pgm.s.cHandyPages; i < RT_ELEMENTS(pVM->pgm.s.aHandyPages); i++)
4534	{
4535	if (pVM->pgm.s.aHandyPages[i].idPage == idPage)
4536	{
4537	pVM->pgm.s.aHandyPages[i].idPage = NIL_GMM_PAGEID;
4538	break;
4539	}
4540	if (pVM->pgm.s.aHandyPages[i].idSharedPage == idPage)
4541	{
4542	pVM->pgm.s.aHandyPages[i].idSharedPage = NIL_GMM_PAGEID;
4543	break;
4544	}
4545	}
4546
4547	/*
4548	* Push it onto the page array.
4549	*/
4550	uint32_t iPage = *pcPendingPages;
4551	Assert(iPage < PGMPHYS_FREE_PAGE_BATCH_SIZE);
4552	*pcPendingPages += 1;
4553
4554	pReq->aPages[iPage].idPage = idPage;
4555
4556	if (iPage + 1 < PGMPHYS_FREE_PAGE_BATCH_SIZE)
4557	return VINF_SUCCESS;
4558
4559	/*
4560	* Flush the pages.
4561	*/
4562	int rc = GMMR3FreePagesPerform(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE);
4563	if (RT_SUCCESS(rc))
4564	{
4565	GMMR3FreePagesRePrep(pVM, pReq, PGMPHYS_FREE_PAGE_BATCH_SIZE, GMMACCOUNT_BASE);
4566	*pcPendingPages = 0;
4567	}
4568	return rc;
4569	}
4570
4571
4572	/**
4573	* Converts a GC physical address to a HC ring-3 pointer, with some
4574	* additional checks.
4575	*
4576	* @returns VBox status code.
4577	* @retval VINF_SUCCESS on success.
4578	* @retval VINF_PGM_PHYS_TLB_CATCH_WRITE and *ppv set if the page has a write
4579	* access handler of some kind.
4580	* @retval VERR_PGM_PHYS_TLB_CATCH_ALL if the page has a handler catching all
4581	* accesses or is odd in any way.
4582	* @retval VERR_PGM_PHYS_TLB_UNASSIGNED if the page doesn't exist.
4583	*
4584	* @param pVM The cross context VM structure.
4585	* @param GCPhys The GC physical address to convert. Since this is only
4586	* used for filling the REM TLB, the A20 mask must be
4587	* applied before calling this API.
4588	* @param fWritable Whether write access is required.
4589	* @param ppv Where to store the pointer corresponding to GCPhys on
4590	* success.
4591	*/
4592	VMMR3DECL(int) PGMR3PhysTlbGCPhys2Ptr(PVM pVM, RTGCPHYS GCPhys, bool fWritable, void **ppv)
4593	{
4594	pgmLock(pVM);
4595	PGM_A20_ASSERT_MASKED(VMMGetCpu(pVM), GCPhys);
4596
4597	PPGMRAMRANGE pRam;
4598	PPGMPAGE pPage;
4599	int rc = pgmPhysGetPageAndRangeEx(pVM, GCPhys, &pPage, &pRam);
4600	if (RT_SUCCESS(rc))
4601	{
4602	if (PGM_PAGE_IS_BALLOONED(pPage))
4603	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
4604	else if (!PGM_PAGE_HAS_ANY_HANDLERS(pPage))
4605	rc = VINF_SUCCESS;
4606	else
4607	{
4608	if (PGM_PAGE_HAS_ACTIVE_ALL_HANDLERS(pPage)) /* catches MMIO */
4609	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
4610	else if (PGM_PAGE_HAS_ACTIVE_HANDLERS(pPage))
4611	{
4612	/** @todo Handle TLB loads of virtual handlers so ./test.sh can be made to work
4613	* in -norawr0 mode. */
4614	if (fWritable)
4615	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
4616	}
4617	else
4618	{
4619	/* Temporarily disabled physical handler(s), since the recompiler
4620	doesn't get notified when it's reset we'll have to pretend it's
4621	operating normally. */
4622	if (pgmHandlerPhysicalIsAll(pVM, GCPhys))
4623	rc = VERR_PGM_PHYS_TLB_CATCH_ALL;
4624	else
4625	rc = VINF_PGM_PHYS_TLB_CATCH_WRITE;
4626	}
4627	}
4628	if (RT_SUCCESS(rc))
4629	{
4630	int rc2;
4631
4632	/* Make sure what we return is writable. */
4633	if (fWritable)
4634	switch (PGM_PAGE_GET_STATE(pPage))
4635	{
4636	case PGM_PAGE_STATE_ALLOCATED:
4637	break;
4638	case PGM_PAGE_STATE_BALLOONED:
4639	AssertFailed();
4640	break;
4641	case PGM_PAGE_STATE_ZERO:
4642	case PGM_PAGE_STATE_SHARED:
4643	if (rc == VINF_PGM_PHYS_TLB_CATCH_WRITE)
4644	break;
4645	case PGM_PAGE_STATE_WRITE_MONITORED:
4646	rc2 = pgmPhysPageMakeWritable(pVM, pPage, GCPhys & ~(RTGCPHYS)PAGE_OFFSET_MASK);
4647	AssertLogRelRCReturn(rc2, rc2);
4648	break;
4649	}
4650
4651	/* Get a ring-3 mapping of the address. */
4652	PPGMPAGER3MAPTLBE pTlbe;
4653	rc2 = pgmPhysPageQueryTlbe(pVM, GCPhys, &pTlbe);
4654	AssertLogRelRCReturn(rc2, rc2);
4655	ppv = (void )((uintptr_t)pTlbe->pv \| (uintptr_t)(GCPhys & PAGE_OFFSET_MASK));
4656	/** @todo mapping/locking hell; this isn't horribly efficient since
4657	* pgmPhysPageLoadIntoTlb will repeat the lookup we've done here. */
4658
4659	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage] ppv=%p\n", GCPhys, rc, pPage, ppv));
4660	}
4661	else
4662	Log6(("PGMR3PhysTlbGCPhys2Ptr: GCPhys=%RGp rc=%Rrc pPage=%R[pgmpage]\n", GCPhys, rc, pPage));
4663
4664	/* else: handler catching all access, no pointer returned. */
4665	}
4666	else
4667	rc = VERR_PGM_PHYS_TLB_UNASSIGNED;
4668
4669	pgmUnlock(pVM);
4670	return rc;
4671	}
4672

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PGMPhys.cpp@ 63648

Download in other formats: