PDMAsyncCompletionFileCache.cpp@ 34646

Last change on this file since 34646 was 34406, checked in by vboxsync, 14 years ago
iprt/list.h: RTListNodeGetFirst/Last -> RTListGetFirst/Last; added RTListGetNext, RTListGetPrev, RTListNodeInsertAfter and RTListNodeInsertBefore.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 83.1 KB

Line
1	/* $Id: PDMAsyncCompletionFileCache.cpp 34406 2010-11-26 16:45:34Z vboxsync $ */
2	/** @file
3	* PDM Async I/O - Transport data asynchronous in R3 using EMT.
4	* File data cache.
5	*/
6
7	/*
8	* Copyright (C) 2006-2008 Oracle Corporation
9	*
10	* This file is part of VirtualBox Open Source Edition (OSE), as
11	* available from http://www.virtualbox.org. This file is free software;
12	* you can redistribute it and/or modify it under the terms of the GNU
13	* General Public License (GPL) as published by the Free Software
14	* Foundation, in version 2 as it comes in the "COPYING" file of the
15	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
16	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
17	*/
18
19	/** @page pg_pdm_async_completion_cache PDM Async Completion Cache - The file I/O cache
20	* This component implements an I/O cache for file endpoints based on the 2Q cache algorithm.
21	*/
22
23	/*******************************************************************************
24	* Header Files *
25	*******************************************************************************/
26	#define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION
27	#include <iprt/asm.h>
28	#include <iprt/mem.h>
29	#include <iprt/path.h>
30	#include <VBox/log.h>
31	#include <VBox/stam.h>
32
33	#include "PDMAsyncCompletionFileInternal.h"
34
35	/**
36	* A I/O memory context.
37	*/
38	typedef struct PDMIOMEMCTX
39	{
40	/** Pointer to the scatter/gather list. */
41	PCRTSGSEG paDataSeg;
42	/** Number of segments. */
43	size_t cSegments;
44	/** Current segment we are in. */
45	unsigned iSegIdx;
46	/** Pointer to the current buffer. */
47	uint8_t *pbBuf;
48	/** Number of bytes left in the current buffer. */
49	size_t cbBufLeft;
50	} PDMIOMEMCTX, *PPDMIOMEMCTX;
51
52	#ifdef VBOX_STRICT
53	# define PDMACFILECACHE_IS_CRITSECT_OWNER(Cache) \
54	do \
55	{ \
56	AssertMsg(RTCritSectIsOwner(&Cache->CritSect), \
57	("Thread does not own critical section\n"));\
58	} while(0)
59
60	# define PDMACFILECACHE_EP_IS_SEMRW_WRITE_OWNER(pEpCache) \
61	do \
62	{ \
63	AssertMsg(RTSemRWIsWriteOwner(pEpCache->SemRWEntries), \
64	("Thread is not exclusive owner of the per endpoint RW semaphore\n")); \
65	} while(0)
66
67	# define PDMACFILECACHE_EP_IS_SEMRW_READ_OWNER(pEpCache) \
68	do \
69	{ \
70	AssertMsg(RTSemRWIsReadOwner(pEpCache->SemRWEntries), \
71	("Thread is not read owner of the per endpoint RW semaphore\n")); \
72	} while(0)
73
74	#else
75	# define PDMACFILECACHE_IS_CRITSECT_OWNER(Cache) do { } while(0)
76	# define PDMACFILECACHE_EP_IS_SEMRW_WRITE_OWNER(pEpCache) do { } while(0)
77	# define PDMACFILECACHE_EP_IS_SEMRW_READ_OWNER(pEpCache) do { } while(0)
78	#endif
79
80	/*******************************************************************************
81	* Internal Functions *
82	*******************************************************************************/
83	static void pdmacFileCacheTaskCompleted(PPDMACTASKFILE pTask, void *pvUser, int rc);
84
85	/**
86	* Decrement the reference counter of the given cache entry.
87	*
88	* @returns nothing.
89	* @param pEntry The entry to release.
90	*/
91	DECLINLINE(void) pdmacFileEpCacheEntryRelease(PPDMACFILECACHEENTRY pEntry)
92	{
93	AssertMsg(pEntry->cRefs > 0, ("Trying to release a not referenced entry\n"));
94	ASMAtomicDecU32(&pEntry->cRefs);
95	}
96
97	/**
98	* Increment the reference counter of the given cache entry.
99	*
100	* @returns nothing.
101	* @param pEntry The entry to reference.
102	*/
103	DECLINLINE(void) pdmacFileEpCacheEntryRef(PPDMACFILECACHEENTRY pEntry)
104	{
105	ASMAtomicIncU32(&pEntry->cRefs);
106	}
107
108	/**
109	* Initialize a I/O memory context.
110	*
111	* @returns nothing
112	* @param pIoMemCtx Pointer to a unitialized I/O memory context.
113	* @param paDataSeg Pointer to the S/G list.
114	* @param cSegments Number of segments in the S/G list.
115	*/
116	DECLINLINE(void) pdmIoMemCtxInit(PPDMIOMEMCTX pIoMemCtx, PCRTSGSEG paDataSeg, size_t cSegments)
117	{
118	AssertMsg((cSegments > 0) && paDataSeg, ("Trying to initialize a I/O memory context without a S/G list\n"));
119
120	pIoMemCtx->paDataSeg = paDataSeg;
121	pIoMemCtx->cSegments = cSegments;
122	pIoMemCtx->iSegIdx = 0;
123	pIoMemCtx->pbBuf = (uint8_t *)paDataSeg[0].pvSeg;
124	pIoMemCtx->cbBufLeft = paDataSeg[0].cbSeg;
125	}
126
127	/**
128	* Return a buffer from the I/O memory context.
129	*
130	* @returns Pointer to the buffer
131	* @param pIoMemCtx Pointer to the I/O memory context.
132	* @param pcbData Pointer to the amount of byte requested.
133	* If the current buffer doesn't have enough bytes left
134	* the amount is returned in the variable.
135	*/
136	DECLINLINE(uint8_t ) pdmIoMemCtxGetBuffer(PPDMIOMEMCTX pIoMemCtx, size_t pcbData)
137	{
138	size_t cbData = RT_MIN(*pcbData, pIoMemCtx->cbBufLeft);
139	uint8_t *pbBuf = pIoMemCtx->pbBuf;
140
141	pIoMemCtx->cbBufLeft -= cbData;
142
143	/* Advance to the next segment if required. */
144	if (!pIoMemCtx->cbBufLeft)
145	{
146	pIoMemCtx->iSegIdx++;
147
148	if (RT_UNLIKELY(pIoMemCtx->iSegIdx == pIoMemCtx->cSegments))
149	{
150	pIoMemCtx->cbBufLeft = 0;
151	pIoMemCtx->pbBuf = NULL;
152	}
153	else
154	{
155	pIoMemCtx->pbBuf = (uint8_t *)pIoMemCtx->paDataSeg[pIoMemCtx->iSegIdx].pvSeg;
156	pIoMemCtx->cbBufLeft = pIoMemCtx->paDataSeg[pIoMemCtx->iSegIdx].cbSeg;
157	}
158
159	*pcbData = cbData;
160	}
161	else
162	pIoMemCtx->pbBuf += cbData;
163
164	return pbBuf;
165	}
166
167	#ifdef DEBUG
168	static void pdmacFileCacheValidate(PPDMACFILECACHEGLOBAL pCache)
169	{
170	/* Amount of cached data should never exceed the maximum amount. */
171	AssertMsg(pCache->cbCached <= pCache->cbMax,
172	("Current amount of cached data exceeds maximum\n"));
173
174	/* The amount of cached data in the LRU and FRU list should match cbCached */
175	AssertMsg(pCache->LruRecentlyUsedIn.cbCached + pCache->LruFrequentlyUsed.cbCached == pCache->cbCached,
176	("Amount of cached data doesn't match\n"));
177
178	AssertMsg(pCache->LruRecentlyUsedOut.cbCached <= pCache->cbRecentlyUsedOutMax,
179	("Paged out list exceeds maximum\n"));
180	}
181	#endif
182
183	DECLINLINE(void) pdmacFileCacheLockEnter(PPDMACFILECACHEGLOBAL pCache)
184	{
185	RTCritSectEnter(&pCache->CritSect);
186	#ifdef DEBUG
187	pdmacFileCacheValidate(pCache);
188	#endif
189	}
190
191	DECLINLINE(void) pdmacFileCacheLockLeave(PPDMACFILECACHEGLOBAL pCache)
192	{
193	#ifdef DEBUG
194	pdmacFileCacheValidate(pCache);
195	#endif
196	RTCritSectLeave(&pCache->CritSect);
197	}
198
199	DECLINLINE(void) pdmacFileCacheSub(PPDMACFILECACHEGLOBAL pCache, uint32_t cbAmount)
200	{
201	PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
202	pCache->cbCached -= cbAmount;
203	}
204
205	DECLINLINE(void) pdmacFileCacheAdd(PPDMACFILECACHEGLOBAL pCache, uint32_t cbAmount)
206	{
207	PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
208	pCache->cbCached += cbAmount;
209	}
210
211	DECLINLINE(void) pdmacFileCacheListAdd(PPDMACFILELRULIST pList, uint32_t cbAmount)
212	{
213	pList->cbCached += cbAmount;
214	}
215
216	DECLINLINE(void) pdmacFileCacheListSub(PPDMACFILELRULIST pList, uint32_t cbAmount)
217	{
218	pList->cbCached -= cbAmount;
219	}
220
221	#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
222	/**
223	* Checks consistency of a LRU list.
224	*
225	* @returns nothing
226	* @param pList The LRU list to check.
227	* @param pNotInList Element which is not allowed to occur in the list.
228	*/
229	static void pdmacFileCacheCheckList(PPDMACFILELRULIST pList, PPDMACFILECACHEENTRY pNotInList)
230	{
231	PPDMACFILECACHEENTRY pCurr = pList->pHead;
232
233	/* Check that there are no double entries and no cycles in the list. */
234	while (pCurr)
235	{
236	PPDMACFILECACHEENTRY pNext = pCurr->pNext;
237
238	while (pNext)
239	{
240	AssertMsg(pCurr != pNext,
241	("Entry %#p is at least two times in list %#p or there is a cycle in the list\n",
242	pCurr, pList));
243	pNext = pNext->pNext;
244	}
245
246	AssertMsg(pCurr != pNotInList, ("Not allowed entry %#p is in list\n", pCurr));
247
248	if (!pCurr->pNext)
249	AssertMsg(pCurr == pList->pTail, ("End of list reached but last element is not list tail\n"));
250
251	pCurr = pCurr->pNext;
252	}
253	}
254	#endif
255
256	/**
257	* Unlinks a cache entry from the LRU list it is assigned to.
258	*
259	* @returns nothing.
260	* @param pEntry The entry to unlink.
261	*/
262	static void pdmacFileCacheEntryRemoveFromList(PPDMACFILECACHEENTRY pEntry)
263	{
264	PPDMACFILELRULIST pList = pEntry->pList;
265	PPDMACFILECACHEENTRY pPrev, pNext;
266
267	LogFlowFunc((": Deleting entry %#p from list %#p\n", pEntry, pList));
268
269	AssertPtr(pList);
270
271	#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
272	pdmacFileCacheCheckList(pList, NULL);
273	#endif
274
275	pPrev = pEntry->pPrev;
276	pNext = pEntry->pNext;
277
278	AssertMsg(pEntry != pPrev, ("Entry links to itself as previous element\n"));
279	AssertMsg(pEntry != pNext, ("Entry links to itself as next element\n"));
280
281	if (pPrev)
282	pPrev->pNext = pNext;
283	else
284	{
285	pList->pHead = pNext;
286
287	if (pNext)
288	pNext->pPrev = NULL;
289	}
290
291	if (pNext)
292	pNext->pPrev = pPrev;
293	else
294	{
295	pList->pTail = pPrev;
296
297	if (pPrev)
298	pPrev->pNext = NULL;
299	}
300
301	pEntry->pList = NULL;
302	pEntry->pPrev = NULL;
303	pEntry->pNext = NULL;
304	pdmacFileCacheListSub(pList, pEntry->cbData);
305	#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
306	pdmacFileCacheCheckList(pList, pEntry);
307	#endif
308	}
309
310	/**
311	* Adds a cache entry to the given LRU list unlinking it from the currently
312	* assigned list if needed.
313	*
314	* @returns nothing.
315	* @param pList List to the add entry to.
316	* @param pEntry Entry to add.
317	*/
318	static void pdmacFileCacheEntryAddToList(PPDMACFILELRULIST pList, PPDMACFILECACHEENTRY pEntry)
319	{
320	LogFlowFunc((": Adding entry %#p to list %#p\n", pEntry, pList));
321	#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
322	pdmacFileCacheCheckList(pList, NULL);
323	#endif
324
325	/* Remove from old list if needed */
326	if (pEntry->pList)
327	pdmacFileCacheEntryRemoveFromList(pEntry);
328
329	pEntry->pNext = pList->pHead;
330	if (pList->pHead)
331	pList->pHead->pPrev = pEntry;
332	else
333	{
334	Assert(!pList->pTail);
335	pList->pTail = pEntry;
336	}
337
338	pEntry->pPrev = NULL;
339	pList->pHead = pEntry;
340	pdmacFileCacheListAdd(pList, pEntry->cbData);
341	pEntry->pList = pList;
342	#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
343	pdmacFileCacheCheckList(pList, NULL);
344	#endif
345	}
346
347	/**
348	* Destroys a LRU list freeing all entries.
349	*
350	* @returns nothing
351	* @param pList Pointer to the LRU list to destroy.
352	*
353	* @note The caller must own the critical section of the cache.
354	*/
355	static void pdmacFileCacheDestroyList(PPDMACFILELRULIST pList)
356	{
357	while (pList->pHead)
358	{
359	PPDMACFILECACHEENTRY pEntry = pList->pHead;
360
361	pList->pHead = pEntry->pNext;
362
363	AssertMsg(!(pEntry->fFlags & (PDMACFILECACHE_ENTRY_IO_IN_PROGRESS \| PDMACFILECACHE_ENTRY_IS_DIRTY)),
364	("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
365
366	RTMemPageFree(pEntry->pbData, pEntry->cbData);
367	RTMemFree(pEntry);
368	}
369	}
370
371	/**
372	* Tries to remove the given amount of bytes from a given list in the cache
373	* moving the entries to one of the given ghosts lists
374	*
375	* @returns Amount of data which could be freed.
376	* @param pCache Pointer to the global cache data.
377	* @param cbData The amount of the data to free.
378	* @param pListSrc The source list to evict data from.
379	* @param pGhostListSrc The ghost list removed entries should be moved to
380	* NULL if the entry should be freed.
381	* @param fReuseBuffer Flag whether a buffer should be reused if it has the same size
382	* @param ppbBuf Where to store the address of the buffer if an entry with the
383	* same size was found and fReuseBuffer is true.
384	*
385	* @note This function may return fewer bytes than requested because entries
386	* may be marked as non evictable if they are used for I/O at the
387	* moment.
388	*/
389	static size_t pdmacFileCacheEvictPagesFrom(PPDMACFILECACHEGLOBAL pCache, size_t cbData,
390	PPDMACFILELRULIST pListSrc, PPDMACFILELRULIST pGhostListDst,
391	bool fReuseBuffer, uint8_t **ppbBuffer)
392	{
393	size_t cbEvicted = 0;
394
395	PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
396
397	AssertMsg(cbData > 0, ("Evicting 0 bytes not possible\n"));
398	AssertMsg( !pGhostListDst
399	\|\| (pGhostListDst == &pCache->LruRecentlyUsedOut),
400	("Destination list must be NULL or the recently used but paged out list\n"));
401
402	if (fReuseBuffer)
403	{
404	AssertPtr(ppbBuffer);
405	*ppbBuffer = NULL;
406	}
407
408	/* Start deleting from the tail. */
409	PPDMACFILECACHEENTRY pEntry = pListSrc->pTail;
410
411	while ((cbEvicted < cbData) && pEntry)
412	{
413	PPDMACFILECACHEENTRY pCurr = pEntry;
414
415	pEntry = pEntry->pPrev;
416
417	/* We can't evict pages which are currently in progress or dirty but not in progress */
418	if ( !(pCurr->fFlags & PDMACFILECACHE_NOT_EVICTABLE)
419	&& (ASMAtomicReadU32(&pCurr->cRefs) == 0))
420	{
421	/* Ok eviction candidate. Grab the endpoint semaphore and check again
422	* because somebody else might have raced us. */
423	PPDMACFILEENDPOINTCACHE pEndpointCache = &pCurr->pEndpoint->DataCache;
424	RTSemRWRequestWrite(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
425
426	if (!(pCurr->fFlags & PDMACFILECACHE_NOT_EVICTABLE)
427	&& (ASMAtomicReadU32(&pCurr->cRefs) == 0))
428	{
429	LogFlow(("Evicting entry %#p (%u bytes)\n", pCurr, pCurr->cbData));
430
431	if (fReuseBuffer && (pCurr->cbData == cbData))
432	{
433	STAM_COUNTER_INC(&pCache->StatBuffersReused);
434	*ppbBuffer = pCurr->pbData;
435	}
436	else if (pCurr->pbData)
437	RTMemPageFree(pCurr->pbData, pCurr->cbData);
438
439	pCurr->pbData = NULL;
440	cbEvicted += pCurr->cbData;
441
442	pdmacFileCacheEntryRemoveFromList(pCurr);
443	pdmacFileCacheSub(pCache, pCurr->cbData);
444
445	if (pGhostListDst)
446	{
447	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
448
449	PPDMACFILECACHEENTRY pGhostEntFree = pGhostListDst->pTail;
450
451	/* We have to remove the last entries from the paged out list. */
452	while ( ((pGhostListDst->cbCached + pCurr->cbData) > pCache->cbRecentlyUsedOutMax)
453	&& pGhostEntFree)
454	{
455	PPDMACFILECACHEENTRY pFree = pGhostEntFree;
456	PPDMACFILEENDPOINTCACHE pEndpointCacheFree = &pFree->pEndpoint->DataCache;
457
458	pGhostEntFree = pGhostEntFree->pPrev;
459
460	RTSemRWRequestWrite(pEndpointCacheFree->SemRWEntries, RT_INDEFINITE_WAIT);
461
462	if (ASMAtomicReadU32(&pFree->cRefs) == 0)
463	{
464	pdmacFileCacheEntryRemoveFromList(pFree);
465
466	STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
467	RTAvlrFileOffsetRemove(pEndpointCacheFree->pTree, pFree->Core.Key);
468	STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
469
470	RTMemFree(pFree);
471	}
472
473	RTSemRWReleaseWrite(pEndpointCacheFree->SemRWEntries);
474	}
475
476	if (pGhostListDst->cbCached + pCurr->cbData > pCache->cbRecentlyUsedOutMax)
477	{
478	/* Couldn't remove enough entries. Delete */
479	STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
480	RTAvlrFileOffsetRemove(pCurr->pEndpoint->DataCache.pTree, pCurr->Core.Key);
481	STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
482
483	RTMemFree(pCurr);
484	}
485	else
486	pdmacFileCacheEntryAddToList(pGhostListDst, pCurr);
487	}
488	else
489	{
490	/* Delete the entry from the AVL tree it is assigned to. */
491	STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
492	RTAvlrFileOffsetRemove(pCurr->pEndpoint->DataCache.pTree, pCurr->Core.Key);
493	STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
494
495	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
496	RTMemFree(pCurr);
497	}
498	}
499
500	}
501	else
502	LogFlow(("Entry %#p (%u bytes) is still in progress and can't be evicted\n", pCurr, pCurr->cbData));
503	}
504
505	return cbEvicted;
506	}
507
508	static bool pdmacFileCacheReclaim(PPDMACFILECACHEGLOBAL pCache, size_t cbData, bool fReuseBuffer, uint8_t **ppbBuffer)
509	{
510	size_t cbRemoved = 0;
511
512	if ((pCache->cbCached + cbData) < pCache->cbMax)
513	return true;
514	else if ((pCache->LruRecentlyUsedIn.cbCached + cbData) > pCache->cbRecentlyUsedInMax)
515	{
516	/* Try to evict as many bytes as possible from A1in */
517	cbRemoved = pdmacFileCacheEvictPagesFrom(pCache, cbData, &pCache->LruRecentlyUsedIn,
518	&pCache->LruRecentlyUsedOut, fReuseBuffer, ppbBuffer);
519
520	/*
521	* If it was not possible to remove enough entries
522	* try the frequently accessed cache.
523	*/
524	if (cbRemoved < cbData)
525	{
526	Assert(!fReuseBuffer \|\| !ppbBuffer); / It is not possible that we got a buffer with the correct size but we didn't freed enough data. */
527
528	/*
529	* If we removed something we can't pass the reuse buffer flag anymore because
530	* we don't need to evict that much data
531	*/
532	if (!cbRemoved)
533	cbRemoved += pdmacFileCacheEvictPagesFrom(pCache, cbData, &pCache->LruFrequentlyUsed,
534	NULL, fReuseBuffer, ppbBuffer);
535	else
536	cbRemoved += pdmacFileCacheEvictPagesFrom(pCache, cbData - cbRemoved, &pCache->LruFrequentlyUsed,
537	NULL, false, NULL);
538	}
539	}
540	else
541	{
542	/* We have to remove entries from frequently access list. */
543	cbRemoved = pdmacFileCacheEvictPagesFrom(pCache, cbData, &pCache->LruFrequentlyUsed,
544	NULL, fReuseBuffer, ppbBuffer);
545	}
546
547	LogFlowFunc((": removed %u bytes, requested %u\n", cbRemoved, cbData));
548	return (cbRemoved >= cbData);
549	}
550
551	/**
552	* Initiates a read I/O task for the given entry.
553	*
554	* @returns nothing.
555	* @param pEntry The entry to fetch the data to.
556	*/
557	static void pdmacFileCacheReadFromEndpoint(PPDMACFILECACHEENTRY pEntry)
558	{
559	LogFlowFunc((": Reading data into cache entry %#p\n", pEntry));
560
561	/* Make sure no one evicts the entry while it is accessed. */
562	pEntry->fFlags \|= PDMACFILECACHE_ENTRY_IO_IN_PROGRESS;
563
564	PPDMACTASKFILE pIoTask = pdmacFileTaskAlloc(pEntry->pEndpoint);
565	AssertPtr(pIoTask);
566
567	AssertMsg(pEntry->pbData, ("Entry is in ghost state\n"));
568
569	pIoTask->pEndpoint = pEntry->pEndpoint;
570	pIoTask->enmTransferType = PDMACTASKFILETRANSFER_READ;
571	pIoTask->Off = pEntry->Core.Key;
572	pIoTask->DataSeg.cbSeg = pEntry->cbData;
573	pIoTask->DataSeg.pvSeg = pEntry->pbData;
574	pIoTask->pvUser = pEntry;
575	pIoTask->pfnCompleted = pdmacFileCacheTaskCompleted;
576
577	/* Send it off to the I/O manager. */
578	pdmacFileEpAddTask(pEntry->pEndpoint, pIoTask);
579	}
580
581	/**
582	* Initiates a write I/O task for the given entry.
583	*
584	* @returns nothing.
585	* @param pEntry The entry to read the data from.
586	*/
587	static void pdmacFileCacheWriteToEndpoint(PPDMACFILECACHEENTRY pEntry)
588	{
589	LogFlowFunc((": Writing data from cache entry %#p\n", pEntry));
590
591	/* Make sure no one evicts the entry while it is accessed. */
592	pEntry->fFlags \|= PDMACFILECACHE_ENTRY_IO_IN_PROGRESS;
593
594	PPDMACTASKFILE pIoTask = pdmacFileTaskAlloc(pEntry->pEndpoint);
595	AssertPtr(pIoTask);
596
597	AssertMsg(pEntry->pbData, ("Entry is in ghost state\n"));
598
599	pIoTask->pEndpoint = pEntry->pEndpoint;
600	pIoTask->enmTransferType = PDMACTASKFILETRANSFER_WRITE;
601	pIoTask->Off = pEntry->Core.Key;
602	pIoTask->DataSeg.cbSeg = pEntry->cbData;
603	pIoTask->DataSeg.pvSeg = pEntry->pbData;
604	pIoTask->pvUser = pEntry;
605	pIoTask->pfnCompleted = pdmacFileCacheTaskCompleted;
606
607	/* Send it off to the I/O manager. */
608	pdmacFileEpAddTask(pEntry->pEndpoint, pIoTask);
609	}
610
611	/**
612	* Commit a single dirty entry to the endpoint
613	*
614	* @returns nothing
615	* @param pEntry The entry to commit.
616	*/
617	static void pdmacFileCacheEntryCommit(PPDMACFILEENDPOINTCACHE pEndpointCache, PPDMACFILECACHEENTRY pEntry)
618	{
619	NOREF(pEndpointCache);
620	AssertMsg( (pEntry->fFlags & PDMACFILECACHE_ENTRY_IS_DIRTY)
621	&& !(pEntry->fFlags & PDMACFILECACHE_ENTRY_IO_IN_PROGRESS),
622	("Invalid flags set for entry %#p\n", pEntry));
623
624	pdmacFileCacheWriteToEndpoint(pEntry);
625	}
626
627	/**
628	* Commit all dirty entries for a single endpoint.
629	*
630	* @returns nothing.
631	* @param pEndpointCache The endpoint cache to commit.
632	*/
633	static void pdmacFileCacheEndpointCommit(PPDMACFILEENDPOINTCACHE pEndpointCache)
634	{
635	uint32_t cbCommitted = 0;
636	RTSemRWRequestWrite(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
637
638	/* The list is moved to a new header to reduce locking overhead. */
639	RTLISTNODE ListDirtyNotCommitted;
640	RTSPINLOCKTMP Tmp;
641
642	RTListInit(&ListDirtyNotCommitted);
643	RTSpinlockAcquire(pEndpointCache->LockList, &Tmp);
644	RTListMove(&ListDirtyNotCommitted, &pEndpointCache->ListDirtyNotCommitted);
645	RTSpinlockRelease(pEndpointCache->LockList, &Tmp);
646
647	if (!RTListIsEmpty(&ListDirtyNotCommitted))
648	{
649	PPDMACFILECACHEENTRY pEntry = RTListGetFirst(&ListDirtyNotCommitted,
650	PDMACFILECACHEENTRY,
651	NodeNotCommitted);
652
653	while (!RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted))
654	{
655	PPDMACFILECACHEENTRY pNext = RTListNodeGetNext(&pEntry->NodeNotCommitted, PDMACFILECACHEENTRY,
656	NodeNotCommitted);
657	pdmacFileCacheEntryCommit(pEndpointCache, pEntry);
658	cbCommitted += pEntry->cbData;
659	RTListNodeRemove(&pEntry->NodeNotCommitted);
660	pEntry = pNext;
661	}
662
663	/* Commit the last endpoint */
664	Assert(RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted));
665	pdmacFileCacheEntryCommit(pEndpointCache, pEntry);
666	RTListNodeRemove(&pEntry->NodeNotCommitted);
667	AssertMsg(RTListIsEmpty(&ListDirtyNotCommitted),
668	("Committed all entries but list is not empty\n"));
669	}
670
671	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
672	AssertMsg(pEndpointCache->pCache->cbDirty >= cbCommitted,
673	("Number of committed bytes exceeds number of dirty bytes\n"));
674	ASMAtomicSubU32(&pEndpointCache->pCache->cbDirty, cbCommitted);
675	}
676
677	/**
678	* Commit all dirty entries in the cache.
679	*
680	* @returns nothing.
681	* @param pCache The global cache instance.
682	*/
683	static void pdmacFileCacheCommitDirtyEntries(PPDMACFILECACHEGLOBAL pCache)
684	{
685	bool fCommitInProgress = ASMAtomicXchgBool(&pCache->fCommitInProgress, true);
686
687	if (!fCommitInProgress)
688	{
689	pdmacFileCacheLockEnter(pCache);
690	Assert(!RTListIsEmpty(&pCache->ListEndpoints));
691
692	PPDMACFILEENDPOINTCACHE pEndpointCache = RTListGetFirst(&pCache->ListEndpoints,
693	PDMACFILEENDPOINTCACHE,
694	NodeCacheEndpoint);
695	AssertPtr(pEndpointCache);
696
697	while (!RTListNodeIsLast(&pCache->ListEndpoints, &pEndpointCache->NodeCacheEndpoint))
698	{
699	pdmacFileCacheEndpointCommit(pEndpointCache);
700
701	pEndpointCache = RTListNodeGetNext(&pEndpointCache->NodeCacheEndpoint, PDMACFILEENDPOINTCACHE,
702	NodeCacheEndpoint);
703	}
704
705	/* Commit the last endpoint */
706	Assert(RTListNodeIsLast(&pCache->ListEndpoints, &pEndpointCache->NodeCacheEndpoint));
707	pdmacFileCacheEndpointCommit(pEndpointCache);
708
709	pdmacFileCacheLockLeave(pCache);
710	ASMAtomicWriteBool(&pCache->fCommitInProgress, false);
711	}
712	}
713
714	/**
715	* Adds the given entry as a dirty to the cache.
716	*
717	* @returns Flag whether the amount of dirty bytes in the cache exceeds the threshold
718	* @param pEndpointCache The endpoint cache the entry belongs to.
719	* @param pEntry The entry to add.
720	*/
721	static bool pdmacFileCacheAddDirtyEntry(PPDMACFILEENDPOINTCACHE pEndpointCache, PPDMACFILECACHEENTRY pEntry)
722	{
723	bool fDirtyBytesExceeded = false;
724	PPDMACFILECACHEGLOBAL pCache = pEndpointCache->pCache;
725
726	/* If the commit timer is disabled we commit right away. */
727	if (pCache->u32CommitTimeoutMs == 0)
728	{
729	pEntry->fFlags \|= PDMACFILECACHE_ENTRY_IS_DIRTY;
730	pdmacFileCacheEntryCommit(pEndpointCache, pEntry);
731	}
732	else if (!(pEntry->fFlags & PDMACFILECACHE_ENTRY_IS_DIRTY))
733	{
734	pEntry->fFlags \|= PDMACFILECACHE_ENTRY_IS_DIRTY;
735
736	RTSPINLOCKTMP Tmp;
737	RTSpinlockAcquire(pEndpointCache->LockList, &Tmp);
738	RTListAppend(&pEndpointCache->ListDirtyNotCommitted, &pEntry->NodeNotCommitted);
739	RTSpinlockRelease(pEndpointCache->LockList, &Tmp);
740
741	uint32_t cbDirty = ASMAtomicAddU32(&pCache->cbDirty, pEntry->cbData);
742
743	fDirtyBytesExceeded = (cbDirty >= pCache->cbCommitDirtyThreshold);
744	}
745
746	return fDirtyBytesExceeded;
747	}
748
749
750	/**
751	* Completes a task segment freeing all resources and completes the task handle
752	* if everything was transferred.
753	*
754	* @returns Next task segment handle.
755	* @param pTaskSeg Task segment to complete.
756	* @param rc Status code to set.
757	*/
758	static PPDMACFILETASKSEG pdmacFileCacheTaskComplete(PPDMACFILETASKSEG pTaskSeg, int rc)
759	{
760	PPDMACFILETASKSEG pNext = pTaskSeg->pNext;
761	PPDMASYNCCOMPLETIONTASKFILE pTaskFile = pTaskSeg->pTask;
762
763	if (RT_FAILURE(rc))
764	ASMAtomicCmpXchgS32(&pTaskFile->rc, rc, VINF_SUCCESS);
765
766	uint32_t uOld = ASMAtomicSubS32(&pTaskFile->cbTransferLeft, pTaskSeg->cbTransfer);
767	AssertMsg(uOld >= pTaskSeg->cbTransfer, ("New value would overflow\n"));
768	if (!(uOld - pTaskSeg->cbTransfer)
769	&& !ASMAtomicXchgBool(&pTaskFile->fCompleted, true))
770	pdmR3AsyncCompletionCompleteTask(&pTaskFile->Core, pTaskFile->rc, true);
771
772	RTMemFree(pTaskSeg);
773
774	return pNext;
775	}
776
777	/**
778	* Completion callback for I/O tasks.
779	*
780	* @returns nothing.
781	* @param pTask The completed task.
782	* @param pvUser Opaque user data.
783	* @param rc Status code of the completed request.
784	*/
785	static void pdmacFileCacheTaskCompleted(PPDMACTASKFILE pTask, void *pvUser, int rc)
786	{
787	PPDMACFILECACHEENTRY pEntry = (PPDMACFILECACHEENTRY)pvUser;
788	PPDMACFILECACHEGLOBAL pCache = pEntry->pCache;
789	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint = pEntry->pEndpoint;
790	PPDMACFILEENDPOINTCACHE pEndpointCache = &pEndpoint->DataCache;
791
792	/* Reference the entry now as we are clearing the I/O in progress flag
793	* which protects the entry till now. */
794	pdmacFileEpCacheEntryRef(pEntry);
795
796	RTSemRWRequestWrite(pEndpoint->DataCache.SemRWEntries, RT_INDEFINITE_WAIT);
797	pEntry->fFlags &= ~PDMACFILECACHE_ENTRY_IO_IN_PROGRESS;
798
799	/* Process waiting segment list. The data in entry might have changed in-between. */
800	bool fDirty = false;
801	PPDMACFILETASKSEG pComplete = pEntry->pWaitingHead;
802	PPDMACFILETASKSEG pCurr = pComplete;
803
804	AssertMsg((pCurr && pEntry->pWaitingTail) \|\| (!pCurr && !pEntry->pWaitingTail),
805	("The list tail was not updated correctly\n"));
806	pEntry->pWaitingTail = NULL;
807	pEntry->pWaitingHead = NULL;
808
809	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
810	{
811	/*
812	* An error here is difficult to handle as the original request completed already.
813	* The error is logged for now and the VM is paused.
814	* If the user continues the entry is written again in the hope
815	* the user fixed the problem and the next write succeeds.
816	*/
817	/** @todo r=aeichner: This solution doesn't work
818	* The user will get the message but the VM will hang afterwards
819	* VMR3Suspend() returns when the VM is suspended but suspending
820	* the VM will reopen the images readonly in DrvVD. They are closed first
821	* which will close the endpoints. This will block EMT while the
822	* I/O manager processes the close request but the IO manager is stuck
823	* in the VMR3Suspend call and can't process the request.
824	* Another problem is that closing the VM means flushing the cache
825	* but the entry failed and will probably fail again.
826	* No idea so far how to solve this problem... but the user gets informed
827	* at least.
828	*/
829	if (RT_FAILURE(rc))
830	{
831	LogRel(("I/O cache: Error while writing entry at offset %RTfoff (%u bytes) to file \"%s\"\n",
832	pEntry->Core.Key, pEntry->cbData, pEndpoint->Core.pszUri));
833
834	rc = VMSetRuntimeError(pEndpoint->Core.pEpClass->pVM, 0, "CACHE_IOERR",
835	N_("The I/O cache encountered an error while updating data in file \"%s\" (rc=%Rrc). Make sure there is enough free space on the disk and that the disk is working properly. Operation can be resumed afterwards."), pEndpoint->Core.pszUri, rc);
836	AssertRC(rc);
837	rc = VMR3Suspend(pEndpoint->Core.pEpClass->pVM);
838	}
839	else
840	{
841	pEntry->fFlags &= ~PDMACFILECACHE_ENTRY_IS_DIRTY;
842
843	while (pCurr)
844	{
845	AssertMsg(pCurr->fWrite, ("Completed write entries should never have read tasks attached\n"));
846
847	memcpy(pEntry->pbData + pCurr->uBufOffset, pCurr->pvBuf, pCurr->cbTransfer);
848	fDirty = true;
849
850	pCurr = pCurr->pNext;
851	}
852	}
853	}
854	else
855	{
856	AssertMsg(pTask->enmTransferType == PDMACTASKFILETRANSFER_READ, ("Invalid transfer type\n"));
857	AssertMsg(!(pEntry->fFlags & PDMACFILECACHE_ENTRY_IS_DIRTY),
858	("Invalid flags set\n"));
859
860	while (pCurr)
861	{
862	if (pCurr->fWrite)
863	{
864	memcpy(pEntry->pbData + pCurr->uBufOffset, pCurr->pvBuf, pCurr->cbTransfer);
865	fDirty = true;
866	}
867	else
868	memcpy(pCurr->pvBuf, pEntry->pbData + pCurr->uBufOffset, pCurr->cbTransfer);
869
870	pCurr = pCurr->pNext;
871	}
872	}
873
874	bool fCommit = false;
875	if (fDirty)
876	fCommit = pdmacFileCacheAddDirtyEntry(pEndpointCache, pEntry);
877
878	RTSemRWReleaseWrite(pEndpoint->DataCache.SemRWEntries);
879
880	/* Dereference so that it isn't protected anymore except we issued anyother write for it. */
881	pdmacFileEpCacheEntryRelease(pEntry);
882
883	if (fCommit)
884	pdmacFileCacheCommitDirtyEntries(pCache);
885
886	/* Complete waiters now. */
887	while (pComplete)
888	pComplete = pdmacFileCacheTaskComplete(pComplete, rc);
889	}
890
891	/**
892	* Commit timer callback.
893	*/
894	static void pdmacFileCacheCommitTimerCallback(PVM pVM, PTMTIMER pTimer, void *pvUser)
895	{
896	PPDMASYNCCOMPLETIONEPCLASSFILE pClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pvUser;
897	PPDMACFILECACHEGLOBAL pCache = &pClassFile->Cache;
898
899	LogFlowFunc(("Commit interval expired, commiting dirty entries\n"));
900
901	if (ASMAtomicReadU32(&pCache->cbDirty) > 0)
902	pdmacFileCacheCommitDirtyEntries(pCache);
903
904	TMTimerSetMillies(pTimer, pCache->u32CommitTimeoutMs);
905	LogFlowFunc(("Entries committed, going to sleep\n"));
906	}
907
908	/**
909	* Initializes the I/O cache.
910	*
911	* returns VBox status code.
912	* @param pClassFile The global class data for file endpoints.
913	* @param pCfgNode CFGM node to query configuration data from.
914	*/
915	int pdmacFileCacheInit(PPDMASYNCCOMPLETIONEPCLASSFILE pClassFile, PCFGMNODE pCfgNode)
916	{
917	int rc = VINF_SUCCESS;
918	PPDMACFILECACHEGLOBAL pCache = &pClassFile->Cache;
919
920	rc = CFGMR3QueryU32Def(pCfgNode, "CacheSize", &pCache->cbMax, 5 * _1M);
921	AssertLogRelRCReturn(rc, rc);
922
923	RTListInit(&pCache->ListEndpoints);
924	pCache->cRefs = 0;
925	pCache->cbCached = 0;
926	pCache->fCommitInProgress = 0;
927	LogFlowFunc((": Maximum number of bytes cached %u\n", pCache->cbMax));
928
929	/* Initialize members */
930	pCache->LruRecentlyUsedIn.pHead = NULL;
931	pCache->LruRecentlyUsedIn.pTail = NULL;
932	pCache->LruRecentlyUsedIn.cbCached = 0;
933
934	pCache->LruRecentlyUsedOut.pHead = NULL;
935	pCache->LruRecentlyUsedOut.pTail = NULL;
936	pCache->LruRecentlyUsedOut.cbCached = 0;
937
938	pCache->LruFrequentlyUsed.pHead = NULL;
939	pCache->LruFrequentlyUsed.pTail = NULL;
940	pCache->LruFrequentlyUsed.cbCached = 0;
941
942	pCache->cbRecentlyUsedInMax = (pCache->cbMax / 100) * 25; /* 25% of the buffer size */
943	pCache->cbRecentlyUsedOutMax = (pCache->cbMax / 100) * 50; /* 50% of the buffer size */
944	LogFlowFunc((": cbRecentlyUsedInMax=%u cbRecentlyUsedOutMax=%u\n", pCache->cbRecentlyUsedInMax, pCache->cbRecentlyUsedOutMax));
945
946	/** @todo r=aeichner: Experiment to find optimal default values */
947	rc = CFGMR3QueryU32Def(pCfgNode, "CacheCommitIntervalMs", &pCache->u32CommitTimeoutMs, 10000 /* 10sec */);
948	AssertLogRelRCReturn(rc, rc);
949	rc = CFGMR3QueryU32(pCfgNode, "CacheCommitThreshold", &pCache->cbCommitDirtyThreshold);
950	if ( rc == VERR_CFGM_VALUE_NOT_FOUND
951	\|\| rc == VERR_CFGM_NO_PARENT)
952	{
953	/* Start committing after 50% of the cache are dirty */
954	pCache->cbCommitDirtyThreshold = pCache->cbMax / 2;
955	}
956	else
957	return rc;
958
959	STAMR3Register(pClassFile->Core.pVM, &pCache->cbMax,
960	STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
961	"/PDM/AsyncCompletion/File/cbMax",
962	STAMUNIT_BYTES,
963	"Maximum cache size");
964	STAMR3Register(pClassFile->Core.pVM, &pCache->cbCached,
965	STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
966	"/PDM/AsyncCompletion/File/cbCached",
967	STAMUNIT_BYTES,
968	"Currently used cache");
969	STAMR3Register(pClassFile->Core.pVM, &pCache->LruRecentlyUsedIn.cbCached,
970	STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
971	"/PDM/AsyncCompletion/File/cbCachedMruIn",
972	STAMUNIT_BYTES,
973	"Number of bytes cached in MRU list");
974	STAMR3Register(pClassFile->Core.pVM, &pCache->LruRecentlyUsedOut.cbCached,
975	STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
976	"/PDM/AsyncCompletion/File/cbCachedMruOut",
977	STAMUNIT_BYTES,
978	"Number of bytes cached in FRU list");
979	STAMR3Register(pClassFile->Core.pVM, &pCache->LruFrequentlyUsed.cbCached,
980	STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
981	"/PDM/AsyncCompletion/File/cbCachedFru",
982	STAMUNIT_BYTES,
983	"Number of bytes cached in FRU ghost list");
984
985	#ifdef VBOX_WITH_STATISTICS
986	STAMR3Register(pClassFile->Core.pVM, &pCache->cHits,
987	STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
988	"/PDM/AsyncCompletion/File/CacheHits",
989	STAMUNIT_COUNT, "Number of hits in the cache");
990	STAMR3Register(pClassFile->Core.pVM, &pCache->cPartialHits,
991	STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
992	"/PDM/AsyncCompletion/File/CachePartialHits",
993	STAMUNIT_COUNT, "Number of partial hits in the cache");
994	STAMR3Register(pClassFile->Core.pVM, &pCache->cMisses,
995	STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
996	"/PDM/AsyncCompletion/File/CacheMisses",
997	STAMUNIT_COUNT, "Number of misses when accessing the cache");
998	STAMR3Register(pClassFile->Core.pVM, &pCache->StatRead,
999	STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1000	"/PDM/AsyncCompletion/File/CacheRead",
1001	STAMUNIT_BYTES, "Number of bytes read from the cache");
1002	STAMR3Register(pClassFile->Core.pVM, &pCache->StatWritten,
1003	STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1004	"/PDM/AsyncCompletion/File/CacheWritten",
1005	STAMUNIT_BYTES, "Number of bytes written to the cache");
1006	STAMR3Register(pClassFile->Core.pVM, &pCache->StatTreeGet,
1007	STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1008	"/PDM/AsyncCompletion/File/CacheTreeGet",
1009	STAMUNIT_TICKS_PER_CALL, "Time taken to access an entry in the tree");
1010	STAMR3Register(pClassFile->Core.pVM, &pCache->StatTreeInsert,
1011	STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1012	"/PDM/AsyncCompletion/File/CacheTreeInsert",
1013	STAMUNIT_TICKS_PER_CALL, "Time taken to insert an entry in the tree");
1014	STAMR3Register(pClassFile->Core.pVM, &pCache->StatTreeRemove,
1015	STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1016	"/PDM/AsyncCompletion/File/CacheTreeRemove",
1017	STAMUNIT_TICKS_PER_CALL, "Time taken to remove an entry an the tree");
1018	STAMR3Register(pClassFile->Core.pVM, &pCache->StatBuffersReused,
1019	STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1020	"/PDM/AsyncCompletion/File/CacheBuffersReused",
1021	STAMUNIT_COUNT, "Number of times a buffer could be reused");
1022	#endif
1023
1024	/* Initialize the critical section */
1025	rc = RTCritSectInit(&pCache->CritSect);
1026
1027	if (RT_SUCCESS(rc))
1028	{
1029	/* Create the commit timer */
1030	if (pCache->u32CommitTimeoutMs > 0)
1031	rc = TMR3TimerCreateInternal(pClassFile->Core.pVM, TMCLOCK_REAL,
1032	pdmacFileCacheCommitTimerCallback,
1033	pClassFile,
1034	"Cache-Commit",
1035	&pClassFile->Cache.pTimerCommit);
1036
1037	if (RT_SUCCESS(rc))
1038	{
1039	LogRel(("AIOMgr: Cache successfully initialised. Cache size is %u bytes\n", pCache->cbMax));
1040	LogRel(("AIOMgr: Cache commit interval is %u ms\n", pCache->u32CommitTimeoutMs));
1041	LogRel(("AIOMgr: Cache commit threshold is %u bytes\n", pCache->cbCommitDirtyThreshold));
1042	return VINF_SUCCESS;
1043	}
1044
1045	RTCritSectDelete(&pCache->CritSect);
1046	}
1047
1048	return rc;
1049	}
1050
1051	/**
1052	* Destroysthe cache freeing all data.
1053	*
1054	* returns nothing.
1055	* @param pClassFile The global class data for file endpoints.
1056	*/
1057	void pdmacFileCacheDestroy(PPDMASYNCCOMPLETIONEPCLASSFILE pClassFile)
1058	{
1059	PPDMACFILECACHEGLOBAL pCache = &pClassFile->Cache;
1060
1061	/* Make sure no one else uses the cache now */
1062	pdmacFileCacheLockEnter(pCache);
1063
1064	/* Cleanup deleting all cache entries waiting for in progress entries to finish. */
1065	pdmacFileCacheDestroyList(&pCache->LruRecentlyUsedIn);
1066	pdmacFileCacheDestroyList(&pCache->LruRecentlyUsedOut);
1067	pdmacFileCacheDestroyList(&pCache->LruFrequentlyUsed);
1068
1069	pdmacFileCacheLockLeave(pCache);
1070
1071	RTCritSectDelete(&pCache->CritSect);
1072	}
1073
1074	/**
1075	* Initializes per endpoint cache data
1076	* like the AVL tree used to access cached entries.
1077	*
1078	* @returns VBox status code.
1079	* @param pEndpoint The endpoint to init the cache for,
1080	* @param pClassFile The global class data for file endpoints.
1081	*/
1082	int pdmacFileEpCacheInit(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMASYNCCOMPLETIONEPCLASSFILE pClassFile)
1083	{
1084	PPDMACFILEENDPOINTCACHE pEndpointCache = &pEndpoint->DataCache;
1085
1086	pEndpointCache->pCache = &pClassFile->Cache;
1087	RTListInit(&pEndpointCache->ListDirtyNotCommitted);
1088	int rc = RTSpinlockCreate(&pEndpointCache->LockList);
1089
1090	if (RT_SUCCESS(rc))
1091	{
1092	rc = RTSemRWCreate(&pEndpointCache->SemRWEntries);
1093	if (RT_SUCCESS(rc))
1094	{
1095	pEndpointCache->pTree = (PAVLRFOFFTREE)RTMemAllocZ(sizeof(AVLRFOFFTREE));
1096	if (pEndpointCache->pTree)
1097	{
1098	pClassFile->Cache.cRefs++;
1099	RTListAppend(&pClassFile->Cache.ListEndpoints, &pEndpointCache->NodeCacheEndpoint);
1100
1101	/* Arm the timer if this is the first endpoint. */
1102	if ( pClassFile->Cache.cRefs == 1
1103	&& pClassFile->Cache.u32CommitTimeoutMs > 0)
1104	rc = TMTimerSetMillies(pClassFile->Cache.pTimerCommit, pClassFile->Cache.u32CommitTimeoutMs);
1105	}
1106	else
1107	rc = VERR_NO_MEMORY;
1108
1109	if (RT_FAILURE(rc))
1110	RTSemRWDestroy(pEndpointCache->SemRWEntries);
1111	}
1112
1113	if (RT_FAILURE(rc))
1114	RTSpinlockDestroy(pEndpointCache->LockList);
1115	}
1116
1117	#ifdef VBOX_WITH_STATISTICS
1118	if (RT_SUCCESS(rc))
1119	{
1120	STAMR3RegisterF(pClassFile->Core.pVM, &pEndpointCache->StatWriteDeferred,
1121	STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1122	STAMUNIT_COUNT, "Number of deferred writes",
1123	"/PDM/AsyncCompletion/File/%s/Cache/DeferredWrites", RTPathFilename(pEndpoint->Core.pszUri));
1124	}
1125	#endif
1126
1127	LogFlowFunc(("Leave rc=%Rrc\n", rc));
1128	return rc;
1129	}
1130
1131	/**
1132	* Callback for the AVL destroy routine. Frees a cache entry for this endpoint.
1133	*
1134	* @returns IPRT status code.
1135	* @param pNode The node to destroy.
1136	* @param pvUser Opaque user data.
1137	*/
1138	static int pdmacFileEpCacheEntryDestroy(PAVLRFOFFNODECORE pNode, void *pvUser)
1139	{
1140	PPDMACFILECACHEENTRY pEntry = (PPDMACFILECACHEENTRY)pNode;
1141	PPDMACFILECACHEGLOBAL pCache = (PPDMACFILECACHEGLOBAL)pvUser;
1142	PPDMACFILEENDPOINTCACHE pEndpointCache = &pEntry->pEndpoint->DataCache;
1143
1144	while (ASMAtomicReadU32(&pEntry->fFlags) & (PDMACFILECACHE_ENTRY_IO_IN_PROGRESS \| PDMACFILECACHE_ENTRY_IS_DIRTY))
1145	{
1146	/* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
1147	pdmacFileEpCacheEntryRef(pEntry);
1148	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
1149	pdmacFileCacheLockLeave(pCache);
1150
1151	RTThreadSleep(250);
1152
1153	/* Re-enter all locks */
1154	pdmacFileCacheLockEnter(pCache);
1155	RTSemRWRequestWrite(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
1156	pdmacFileEpCacheEntryRelease(pEntry);
1157	}
1158
1159	AssertMsg(!(pEntry->fFlags & (PDMACFILECACHE_ENTRY_IO_IN_PROGRESS \| PDMACFILECACHE_ENTRY_IS_DIRTY)),
1160	("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
1161
1162	bool fUpdateCache = pEntry->pList == &pCache->LruFrequentlyUsed
1163	\|\| pEntry->pList == &pCache->LruRecentlyUsedIn;
1164
1165	pdmacFileCacheEntryRemoveFromList(pEntry);
1166
1167	if (fUpdateCache)
1168	pdmacFileCacheSub(pCache, pEntry->cbData);
1169
1170	RTMemPageFree(pEntry->pbData, pEntry->cbData);
1171	RTMemFree(pEntry);
1172
1173	return VINF_SUCCESS;
1174	}
1175
1176	/**
1177	* Destroys all cache resources used by the given endpoint.
1178	*
1179	* @returns nothing.
1180	* @param pEndpoint The endpoint to the destroy.
1181	*/
1182	void pdmacFileEpCacheDestroy(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
1183	{
1184	PPDMACFILEENDPOINTCACHE pEndpointCache = &pEndpoint->DataCache;
1185	PPDMACFILECACHEGLOBAL pCache = pEndpointCache->pCache;
1186
1187	/* Make sure nobody is accessing the cache while we delete the tree. */
1188	pdmacFileCacheLockEnter(pCache);
1189	RTSemRWRequestWrite(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
1190	RTAvlrFileOffsetDestroy(pEndpointCache->pTree, pdmacFileEpCacheEntryDestroy, pCache);
1191	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
1192
1193	RTSpinlockDestroy(pEndpointCache->LockList);
1194
1195	pCache->cRefs--;
1196	RTListNodeRemove(&pEndpointCache->NodeCacheEndpoint);
1197
1198	if ( !pCache->cRefs
1199	&& pCache->u32CommitTimeoutMs > 0)
1200	TMTimerStop(pCache->pTimerCommit);
1201
1202	pdmacFileCacheLockLeave(pCache);
1203
1204	RTSemRWDestroy(pEndpointCache->SemRWEntries);
1205
1206	#ifdef VBOX_WITH_STATISTICS
1207	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
1208
1209	STAMR3Deregister(pEpClassFile->Core.pVM, &pEndpointCache->StatWriteDeferred);
1210	#endif
1211	}
1212
1213	static PPDMACFILECACHEENTRY pdmacFileEpCacheGetCacheEntryByOffset(PPDMACFILEENDPOINTCACHE pEndpointCache, RTFOFF off)
1214	{
1215	PPDMACFILECACHEGLOBAL pCache = pEndpointCache->pCache;
1216	PPDMACFILECACHEENTRY pEntry = NULL;
1217
1218	STAM_PROFILE_ADV_START(&pCache->StatTreeGet, Cache);
1219
1220	RTSemRWRequestRead(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
1221	pEntry = (PPDMACFILECACHEENTRY)RTAvlrFileOffsetRangeGet(pEndpointCache->pTree, off);
1222	if (pEntry)
1223	pdmacFileEpCacheEntryRef(pEntry);
1224	RTSemRWReleaseRead(pEndpointCache->SemRWEntries);
1225
1226	STAM_PROFILE_ADV_STOP(&pCache->StatTreeGet, Cache);
1227
1228	return pEntry;
1229	}
1230
1231	/**
1232	* Return the best fit cache entries for the given offset.
1233	*
1234	* @returns nothing.
1235	* @param pEndpointCache The endpoint cache.
1236	* @param off The offset.
1237	* @param pEntryAbove Where to store the pointer to the best fit entry above the
1238	* the given offset. NULL if not required.
1239	* @param pEntryBelow Where to store the pointer to the best fit entry below the
1240	* the given offset. NULL if not required.
1241	*/
1242	static void pdmacFileEpCacheGetCacheBestFitEntryByOffset(PPDMACFILEENDPOINTCACHE pEndpointCache, RTFOFF off,
1243	PPDMACFILECACHEENTRY *ppEntryAbove,
1244	PPDMACFILECACHEENTRY *ppEntryBelow)
1245	{
1246	PPDMACFILECACHEGLOBAL pCache = pEndpointCache->pCache;
1247
1248	STAM_PROFILE_ADV_START(&pCache->StatTreeGet, Cache);
1249
1250	RTSemRWRequestRead(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
1251	if (ppEntryAbove)
1252	{
1253	ppEntryAbove = (PPDMACFILECACHEENTRY)RTAvlrFileOffsetGetBestFit(pEndpointCache->pTree, off, true /fAbove*/);
1254	if (*ppEntryAbove)
1255	pdmacFileEpCacheEntryRef(*ppEntryAbove);
1256	}
1257
1258	if (ppEntryBelow)
1259	{
1260	ppEntryBelow = (PPDMACFILECACHEENTRY)RTAvlrFileOffsetGetBestFit(pEndpointCache->pTree, off, false /fAbove*/);
1261	if (*ppEntryBelow)
1262	pdmacFileEpCacheEntryRef(*ppEntryBelow);
1263	}
1264	RTSemRWReleaseRead(pEndpointCache->SemRWEntries);
1265
1266	STAM_PROFILE_ADV_STOP(&pCache->StatTreeGet, Cache);
1267	}
1268
1269	static void pdmacFileEpCacheInsertEntry(PPDMACFILEENDPOINTCACHE pEndpointCache, PPDMACFILECACHEENTRY pEntry)
1270	{
1271	PPDMACFILECACHEGLOBAL pCache = pEndpointCache->pCache;
1272
1273	STAM_PROFILE_ADV_START(&pCache->StatTreeInsert, Cache);
1274	RTSemRWRequestWrite(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
1275	bool fInserted = RTAvlrFileOffsetInsert(pEndpointCache->pTree, &pEntry->Core);
1276	AssertMsg(fInserted, ("Node was not inserted into tree\n"));
1277	STAM_PROFILE_ADV_STOP(&pCache->StatTreeInsert, Cache);
1278	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
1279	}
1280
1281	/**
1282	* Allocates and initializes a new entry for the cache.
1283	* The entry has a reference count of 1.
1284	*
1285	* @returns Pointer to the new cache entry or NULL if out of memory.
1286	* @param pCache The cache the entry belongs to.
1287	* @param pEndoint The endpoint the entry holds data for.
1288	* @param off Start offset.
1289	* @param cbData Size of the cache entry.
1290	* @param pbBuffer Pointer to the buffer to use.
1291	* NULL if a new buffer should be allocated.
1292	* The buffer needs to have the same size of the entry.
1293	*/
1294	static PPDMACFILECACHEENTRY pdmacFileCacheEntryAlloc(PPDMACFILECACHEGLOBAL pCache,
1295	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
1296	RTFOFF off, size_t cbData, uint8_t *pbBuffer)
1297	{
1298	PPDMACFILECACHEENTRY pEntryNew = (PPDMACFILECACHEENTRY)RTMemAllocZ(sizeof(PDMACFILECACHEENTRY));
1299
1300	if (RT_UNLIKELY(!pEntryNew))
1301	return NULL;
1302
1303	pEntryNew->Core.Key = off;
1304	pEntryNew->Core.KeyLast = off + cbData - 1;
1305	pEntryNew->pEndpoint = pEndpoint;
1306	pEntryNew->pCache = pCache;
1307	pEntryNew->fFlags = 0;
1308	pEntryNew->cRefs = 1; /* We are using it now. */
1309	pEntryNew->pList = NULL;
1310	pEntryNew->cbData = cbData;
1311	pEntryNew->pWaitingHead = NULL;
1312	pEntryNew->pWaitingTail = NULL;
1313	if (pbBuffer)
1314	pEntryNew->pbData = pbBuffer;
1315	else
1316	pEntryNew->pbData = (uint8_t *)RTMemPageAlloc(cbData);
1317
1318	if (RT_UNLIKELY(!pEntryNew->pbData))
1319	{
1320	RTMemFree(pEntryNew);
1321	return NULL;
1322	}
1323
1324	return pEntryNew;
1325	}
1326
1327	/**
1328	* Adds a segment to the waiting list for a cache entry
1329	* which is currently in progress.
1330	*
1331	* @returns nothing.
1332	* @param pEntry The cache entry to add the segment to.
1333	* @param pSeg The segment to add.
1334	*/
1335	DECLINLINE(void) pdmacFileEpCacheEntryAddWaitingSegment(PPDMACFILECACHEENTRY pEntry, PPDMACFILETASKSEG pSeg)
1336	{
1337	pSeg->pNext = NULL;
1338
1339	if (pEntry->pWaitingHead)
1340	{
1341	AssertPtr(pEntry->pWaitingTail);
1342
1343	pEntry->pWaitingTail->pNext = pSeg;
1344	pEntry->pWaitingTail = pSeg;
1345	}
1346	else
1347	{
1348	Assert(!pEntry->pWaitingTail);
1349
1350	pEntry->pWaitingHead = pSeg;
1351	pEntry->pWaitingTail = pSeg;
1352	}
1353	}
1354
1355	/**
1356	* Checks that a set of flags is set/clear acquiring the R/W semaphore
1357	* in exclusive mode.
1358	*
1359	* @returns true if the flag in fSet is set and the one in fClear is clear.
1360	* false otherwise.
1361	* The R/W semaphore is only held if true is returned.
1362	*
1363	* @param pEndpointCache The endpoint cache instance data.
1364	* @param pEntry The entry to check the flags for.
1365	* @param fSet The flag which is tested to be set.
1366	* @param fClear The flag which is tested to be clear.
1367	*/
1368	DECLINLINE(bool) pdmacFileEpCacheEntryFlagIsSetClearAcquireLock(PPDMACFILEENDPOINTCACHE pEndpointCache,
1369	PPDMACFILECACHEENTRY pEntry,
1370	uint32_t fSet, uint32_t fClear)
1371	{
1372	uint32_t fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1373	bool fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1374
1375	if (fPassed)
1376	{
1377	/* Acquire the lock and check again because the completion callback might have raced us. */
1378	RTSemRWRequestWrite(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
1379
1380	fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1381	fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1382
1383	/* Drop the lock if we didn't passed the test. */
1384	if (!fPassed)
1385	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
1386	}
1387
1388	return fPassed;
1389	}
1390
1391	/**
1392	* Copies data to a buffer described by a I/O memory context.
1393	*
1394	* @returns nothing.
1395	* @param pIoMemCtx The I/O memory context to copy the data into.
1396	* @param pbData Pointer to the data data to copy.
1397	* @param cbData Amount of data to copy.
1398	*/
1399	static void pdmacFileEpCacheCopyToIoMemCtx(PPDMIOMEMCTX pIoMemCtx,
1400	uint8_t *pbData,
1401	size_t cbData)
1402	{
1403	while (cbData)
1404	{
1405	size_t cbCopy = cbData;
1406	uint8_t *pbBuf = pdmIoMemCtxGetBuffer(pIoMemCtx, &cbCopy);
1407
1408	AssertPtr(pbBuf);
1409
1410	memcpy(pbBuf, pbData, cbCopy);
1411
1412	cbData -= cbCopy;
1413	pbData += cbCopy;
1414	}
1415	}
1416
1417	/**
1418	* Copies data from a buffer described by a I/O memory context.
1419	*
1420	* @returns nothing.
1421	* @param pIoMemCtx The I/O memory context to copy the data from.
1422	* @param pbData Pointer to the destination buffer.
1423	* @param cbData Amount of data to copy.
1424	*/
1425	static void pdmacFileEpCacheCopyFromIoMemCtx(PPDMIOMEMCTX pIoMemCtx,
1426	uint8_t *pbData,
1427	size_t cbData)
1428	{
1429	while (cbData)
1430	{
1431	size_t cbCopy = cbData;
1432	uint8_t *pbBuf = pdmIoMemCtxGetBuffer(pIoMemCtx, &cbCopy);
1433
1434	AssertPtr(pbBuf);
1435
1436	memcpy(pbData, pbBuf, cbCopy);
1437
1438	cbData -= cbCopy;
1439	pbData += cbCopy;
1440	}
1441	}
1442
1443	/**
1444	* Add a buffer described by the I/O memory context
1445	* to the entry waiting for completion.
1446	*
1447	* @returns nothing.
1448	* @param pEntry The entry to add the buffer to.
1449	* @param pTask Task associated with the buffer.
1450	* @param pIoMemCtx The memory context to use.
1451	* @param OffDiff Offset from the start of the buffer
1452	* in the entry.
1453	* @param cbData Amount of data to wait for onthis entry.
1454	* @param fWrite Flag whether the task waits because it wants to write
1455	* to the cache entry.
1456	*/
1457	static void pdmacFileEpCacheEntryWaitersAdd(PPDMACFILECACHEENTRY pEntry,
1458	PPDMASYNCCOMPLETIONTASKFILE pTask,
1459	PPDMIOMEMCTX pIoMemCtx,
1460	RTFOFF OffDiff,
1461	size_t cbData,
1462	bool fWrite)
1463	{
1464	while (cbData)
1465	{
1466	PPDMACFILETASKSEG pSeg = (PPDMACFILETASKSEG)RTMemAllocZ(sizeof(PDMACFILETASKSEG));
1467	size_t cbSeg = cbData;
1468	uint8_t *pbBuf = pdmIoMemCtxGetBuffer(pIoMemCtx, &cbSeg);
1469
1470	pSeg->pTask = pTask;
1471	pSeg->uBufOffset = OffDiff;
1472	pSeg->cbTransfer = cbSeg;
1473	pSeg->pvBuf = pbBuf;
1474	pSeg->fWrite = fWrite;
1475
1476	pdmacFileEpCacheEntryAddWaitingSegment(pEntry, pSeg);
1477
1478	cbData -= cbSeg;
1479	OffDiff += cbSeg;
1480	}
1481	}
1482
1483	/**
1484	* Passthrough a part of a request directly to the I/O manager
1485	* handling the endpoint.
1486	*
1487	* @returns nothing.
1488	* @param pEndpoint The endpoint.
1489	* @param pTask The task.
1490	* @param pIoMemCtx The I/O memory context to use.
1491	* @param offStart Offset to start transfer from.
1492	* @param cbData Amount of data to transfer.
1493	* @param enmTransferType The transfer type (read/write)
1494	*/
1495	static void pdmacFileEpCacheRequestPassthrough(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
1496	PPDMASYNCCOMPLETIONTASKFILE pTask,
1497	PPDMIOMEMCTX pIoMemCtx,
1498	RTFOFF offStart, size_t cbData,
1499	PDMACTASKFILETRANSFER enmTransferType)
1500	{
1501	while (cbData)
1502	{
1503	size_t cbSeg = cbData;
1504	uint8_t *pbBuf = pdmIoMemCtxGetBuffer(pIoMemCtx, &cbSeg);
1505	PPDMACTASKFILE pIoTask = pdmacFileTaskAlloc(pEndpoint);
1506	AssertPtr(pIoTask);
1507
1508	pIoTask->pEndpoint = pEndpoint;
1509	pIoTask->enmTransferType = enmTransferType;
1510	pIoTask->Off = offStart;
1511	pIoTask->DataSeg.cbSeg = cbSeg;
1512	pIoTask->DataSeg.pvSeg = pbBuf;
1513	pIoTask->pvUser = pTask;
1514	pIoTask->pfnCompleted = pdmacFileEpTaskCompleted;
1515
1516	offStart += cbSeg;
1517	cbData -= cbSeg;
1518
1519	/* Send it off to the I/O manager. */
1520	pdmacFileEpAddTask(pEndpoint, pIoTask);
1521	}
1522	}
1523
1524	/**
1525	* Calculate aligned offset and size for a new cache entry
1526	* which do not intersect with an already existing entry and the
1527	* file end.
1528	*
1529	* @returns The number of bytes the entry can hold of the requested amount
1530	* of byte.
1531	* @param pEndpoint The endpoint.
1532	* @param pEndpointCache The endpoint cache.
1533	* @param off The start offset.
1534	* @param cb The number of bytes the entry needs to hold at least.
1535	* @param uAlignment Alignment of the boundary sizes.
1536	* @param poffAligned Where to store the aligned offset.
1537	* @param pcbAligned Where to store the aligned size of the entry.
1538	*/
1539	static size_t pdmacFileEpCacheEntryBoundariesCalc(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
1540	PPDMACFILEENDPOINTCACHE pEndpointCache,
1541	RTFOFF off, size_t cb,
1542	unsigned uAlignment,
1543	RTFOFF poffAligned, size_t pcbAligned)
1544	{
1545	size_t cbAligned;
1546	size_t cbInEntry = 0;
1547	RTFOFF offAligned;
1548	PPDMACFILECACHEENTRY pEntryAbove = NULL;
1549	PPDMACFILECACHEENTRY pEntryBelow = NULL;
1550
1551	/* Get the best fit entries around the offset */
1552	pdmacFileEpCacheGetCacheBestFitEntryByOffset(pEndpointCache, off,
1553	&pEntryAbove, &pEntryBelow);
1554
1555	/* Log the info */
1556	LogFlow(("%sest fit entry below off=%RTfoff (BestFit=%RTfoff BestFitEnd=%RTfoff BestFitSize=%u)\n",
1557	pEntryBelow ? "B" : "No b",
1558	off,
1559	pEntryBelow ? pEntryBelow->Core.Key : 0,
1560	pEntryBelow ? pEntryBelow->Core.KeyLast : 0,
1561	pEntryBelow ? pEntryBelow->cbData : 0));
1562
1563	LogFlow(("%sest fit entry above off=%RTfoff (BestFit=%RTfoff BestFitEnd=%RTfoff BestFitSize=%u)\n",
1564	pEntryAbove ? "B" : "No b",
1565	off,
1566	pEntryAbove ? pEntryAbove->Core.Key : 0,
1567	pEntryAbove ? pEntryAbove->Core.KeyLast : 0,
1568	pEntryAbove ? pEntryAbove->cbData : 0));
1569
1570	/* Align the offset first. */
1571	offAligned = off & ~(RTFOFF)(512-1);
1572	if ( pEntryBelow
1573	&& offAligned <= pEntryBelow->Core.KeyLast)
1574	offAligned = pEntryBelow->Core.KeyLast;
1575
1576	if ( pEntryAbove
1577	&& off + (RTFOFF)cb > pEntryAbove->Core.Key)
1578	{
1579	cbInEntry = pEntryAbove->Core.Key - off;
1580	cbAligned = pEntryAbove->Core.Key - offAligned;
1581	}
1582	else
1583	{
1584	/*
1585	* Align the size to a 4KB boundary.
1586	* Memory size is aligned to a page boundary
1587	* and memory is wasted if the size is rather small.
1588	* (For example reads with a size of 512 bytes).
1589	*/
1590	cbInEntry = cb;
1591	cbAligned = RT_ALIGN_Z(cb + (off - offAligned), uAlignment);
1592
1593	/*
1594	* Clip to file size if the original request doesn't
1595	* exceed the file (not an appending write)
1596	*/
1597	uint64_t cbReq = off + (RTFOFF)cb;
1598	if (cbReq >= pEndpoint->cbFile)
1599	cbAligned = cbReq - offAligned;
1600	else
1601	cbAligned = RT_MIN(pEndpoint->cbFile - offAligned, cbAligned);
1602	if (pEntryAbove)
1603	{
1604	Assert(pEntryAbove->Core.Key >= off);
1605	cbAligned = RT_MIN(cbAligned, (uint64_t)pEntryAbove->Core.Key - offAligned);
1606	}
1607	}
1608
1609	/* A few sanity checks */
1610	AssertMsg(!pEntryBelow \|\| pEntryBelow->Core.KeyLast < offAligned,
1611	("Aligned start offset intersects with another cache entry\n"));
1612	AssertMsg(!pEntryAbove \|\| (offAligned + (RTFOFF)cbAligned) <= pEntryAbove->Core.Key,
1613	("Aligned size intersects with another cache entry\n"));
1614	Assert(cbInEntry <= cbAligned);
1615	AssertMsg( ( offAligned + (RTFOFF)cbAligned <= (RTFOFF)pEndpoint->cbFile
1616	&& off + (RTFOFF)cb <= (RTFOFF)pEndpoint->cbFile)
1617	\|\| (offAligned + (RTFOFF)cbAligned <= off + (RTFOFF)cb),
1618	("Unwanted file size increase\n"));
1619
1620	if (pEntryBelow)
1621	pdmacFileEpCacheEntryRelease(pEntryBelow);
1622	if (pEntryAbove)
1623	pdmacFileEpCacheEntryRelease(pEntryAbove);
1624
1625	LogFlow(("offAligned=%RTfoff cbAligned=%u\n", offAligned, cbAligned));
1626
1627	*poffAligned = offAligned;
1628	*pcbAligned = cbAligned;
1629
1630	return cbInEntry;
1631	}
1632
1633	/**
1634	* Create a new cache entry evicting data from the cache if required.
1635	*
1636	* @returns Pointer to the new cache entry or NULL
1637	* if not enough bytes could be evicted from the cache.
1638	* @param pEndpoint The endpoint.
1639	* @param pEndpointCache The endpoint cache.
1640	* @param off The offset.
1641	* @param cb Number of bytes the cache entry should have.
1642	* @param uAlignment Alignment the size of the entry should have.
1643	* @param pcbData Where to store the number of bytes the new
1644	* entry can hold. May be lower than actually requested
1645	* due to another entry intersecting the access range.
1646	*/
1647	static PPDMACFILECACHEENTRY pdmacFileEpCacheEntryCreate(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
1648	PPDMACFILEENDPOINTCACHE pEndpointCache,
1649	RTFOFF off, size_t cb,
1650	unsigned uAlignment,
1651	size_t *pcbData)
1652	{
1653	RTFOFF offStart = 0;
1654	size_t cbEntry = 0;
1655	PPDMACFILECACHEENTRY pEntryNew = NULL;
1656	PPDMACFILECACHEGLOBAL pCache = pEndpointCache->pCache;
1657	uint8_t *pbBuffer = NULL;
1658
1659	*pcbData = pdmacFileEpCacheEntryBoundariesCalc(pEndpoint,
1660	pEndpointCache,
1661	off, cb,
1662	uAlignment,
1663	&offStart, &cbEntry);
1664
1665	pdmacFileCacheLockEnter(pCache);
1666	bool fEnough = pdmacFileCacheReclaim(pCache, cbEntry, true, &pbBuffer);
1667
1668	if (fEnough)
1669	{
1670	LogFlow(("Evicted enough bytes (%u requested). Creating new cache entry\n", cbEntry));
1671
1672	pEntryNew = pdmacFileCacheEntryAlloc(pCache, pEndpoint,
1673	offStart, cbEntry,
1674	pbBuffer);
1675	if (RT_LIKELY(pEntryNew))
1676	{
1677	pdmacFileCacheEntryAddToList(&pCache->LruRecentlyUsedIn, pEntryNew);
1678	pdmacFileCacheAdd(pCache, cbEntry);
1679	pdmacFileCacheLockLeave(pCache);
1680
1681	pdmacFileEpCacheInsertEntry(pEndpointCache, pEntryNew);
1682
1683	AssertMsg( (off >= pEntryNew->Core.Key)
1684	&& (off + (RTFOFF)*pcbData <= pEntryNew->Core.KeyLast + 1),
1685	("Overflow in calculation off=%RTfoff OffsetAligned=%RTfoff\n",
1686	off, pEntryNew->Core.Key));
1687	}
1688	else
1689	pdmacFileCacheLockLeave(pCache);
1690	}
1691	else
1692	pdmacFileCacheLockLeave(pCache);
1693
1694	return pEntryNew;
1695	}
1696
1697	/**
1698	* Reads the specified data from the endpoint using the cache if possible.
1699	*
1700	* @returns VBox status code.
1701	* @param pEndpoint The endpoint to read from.
1702	* @param pTask The task structure used as identifier for this request.
1703	* @param off The offset to start reading from.
1704	* @param paSegments Pointer to the array holding the destination buffers.
1705	* @param cSegments Number of segments in the array.
1706	* @param cbRead Number of bytes to read.
1707	*/
1708	int pdmacFileEpCacheRead(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMASYNCCOMPLETIONTASKFILE pTask,
1709	RTFOFF off, PCRTSGSEG paSegments, size_t cSegments,
1710	size_t cbRead)
1711	{
1712	int rc = VINF_SUCCESS;
1713	PPDMACFILEENDPOINTCACHE pEndpointCache = &pEndpoint->DataCache;
1714	PPDMACFILECACHEGLOBAL pCache = pEndpointCache->pCache;
1715	PPDMACFILECACHEENTRY pEntry;
1716
1717	LogFlowFunc((": pEndpoint=%#p{%s} pTask=%#p off=%RTfoff paSegments=%#p cSegments=%u cbRead=%u\n",
1718	pEndpoint, pEndpoint->Core.pszUri, pTask, off, paSegments, cSegments, cbRead));
1719
1720	/* Set to completed to make sure that the task is valid while we access it. */
1721	ASMAtomicWriteBool(&pTask->fCompleted, true);
1722
1723	/* Init the I/O memory context */
1724	PDMIOMEMCTX IoMemCtx;
1725	pdmIoMemCtxInit(&IoMemCtx, paSegments, cSegments);
1726
1727	while (cbRead)
1728	{
1729	size_t cbToRead;
1730
1731	pEntry = pdmacFileEpCacheGetCacheEntryByOffset(pEndpointCache, off);
1732
1733	/*
1734	* If there is no entry we try to create a new one eviciting unused pages
1735	* if the cache is full. If this is not possible we will pass the request through
1736	* and skip the caching (all entries may be still in progress so they can't
1737	* be evicted)
1738	* If we have an entry it can be in one of the LRU lists where the entry
1739	* contains data (recently used or frequently used LRU) so we can just read
1740	* the data we need and put the entry at the head of the frequently used LRU list.
1741	* In case the entry is in one of the ghost lists it doesn't contain any data.
1742	* We have to fetch it again evicting pages from either T1 or T2 to make room.
1743	*/
1744	if (pEntry)
1745	{
1746	RTFOFF OffDiff = off - pEntry->Core.Key;
1747
1748	AssertMsg(off >= pEntry->Core.Key,
1749	("Overflow in calculation off=%RTfoff OffsetAligned=%RTfoff\n",
1750	off, pEntry->Core.Key));
1751
1752	AssertPtr(pEntry->pList);
1753
1754	cbToRead = RT_MIN(pEntry->cbData - OffDiff, cbRead);
1755
1756	AssertMsg(off + (RTFOFF)cbToRead <= pEntry->Core.Key + pEntry->Core.KeyLast + 1,
1757	("Buffer of cache entry exceeded off=%RTfoff cbToRead=%d\n",
1758	off, cbToRead));
1759
1760	cbRead -= cbToRead;
1761
1762	if (!cbRead)
1763	STAM_COUNTER_INC(&pCache->cHits);
1764	else
1765	STAM_COUNTER_INC(&pCache->cPartialHits);
1766
1767	STAM_COUNTER_ADD(&pCache->StatRead, cbToRead);
1768
1769	/* Ghost lists contain no data. */
1770	if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
1771	\|\| (pEntry->pList == &pCache->LruFrequentlyUsed))
1772	{
1773	if (pdmacFileEpCacheEntryFlagIsSetClearAcquireLock(pEndpointCache, pEntry,
1774	PDMACFILECACHE_ENTRY_IO_IN_PROGRESS,
1775	PDMACFILECACHE_ENTRY_IS_DIRTY))
1776	{
1777	/* Entry didn't completed yet. Append to the list */
1778	pdmacFileEpCacheEntryWaitersAdd(pEntry, pTask,
1779	&IoMemCtx,
1780	OffDiff, cbToRead,
1781	false /* fWrite */);
1782	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
1783	}
1784	else
1785	{
1786	/* Read as much as we can from the entry. */
1787	pdmacFileEpCacheCopyToIoMemCtx(&IoMemCtx, pEntry->pbData + OffDiff, cbToRead);
1788	ASMAtomicSubS32(&pTask->cbTransferLeft, cbToRead);
1789	}
1790
1791	/* Move this entry to the top position */
1792	if (pEntry->pList == &pCache->LruFrequentlyUsed)
1793	{
1794	pdmacFileCacheLockEnter(pCache);
1795	pdmacFileCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
1796	pdmacFileCacheLockLeave(pCache);
1797	}
1798	/* Release the entry */
1799	pdmacFileEpCacheEntryRelease(pEntry);
1800	}
1801	else
1802	{
1803	uint8_t *pbBuffer = NULL;
1804
1805	LogFlow(("Fetching data for ghost entry %#p from file\n", pEntry));
1806
1807	pdmacFileCacheLockEnter(pCache);
1808	pdmacFileCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
1809	bool fEnough = pdmacFileCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
1810
1811	/* Move the entry to Am and fetch it to the cache. */
1812	if (fEnough)
1813	{
1814	pdmacFileCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
1815	pdmacFileCacheAdd(pCache, pEntry->cbData);
1816	pdmacFileCacheLockLeave(pCache);
1817
1818	if (pbBuffer)
1819	pEntry->pbData = pbBuffer;
1820	else
1821	pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
1822	AssertPtr(pEntry->pbData);
1823
1824	pdmacFileEpCacheEntryWaitersAdd(pEntry, pTask,
1825	&IoMemCtx,
1826	OffDiff, cbToRead,
1827	false /* fWrite */);
1828	pdmacFileCacheReadFromEndpoint(pEntry);
1829	/* Release the entry */
1830	pdmacFileEpCacheEntryRelease(pEntry);
1831	}
1832	else
1833	{
1834	RTSemRWRequestWrite(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
1835	STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
1836	RTAvlrFileOffsetRemove(pEndpointCache->pTree, pEntry->Core.Key);
1837	STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
1838	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
1839
1840	pdmacFileCacheLockLeave(pCache);
1841
1842	RTMemFree(pEntry);
1843
1844	pdmacFileEpCacheRequestPassthrough(pEndpoint, pTask,
1845	&IoMemCtx, off, cbToRead,
1846	PDMACTASKFILETRANSFER_READ);
1847	}
1848	}
1849	}
1850	else
1851	{
1852	#ifdef VBOX_WITH_IO_READ_CACHE
1853	/* No entry found for this offset. Create a new entry and fetch the data to the cache. */
1854	PPDMACFILECACHEENTRY pEntryNew = pdmacFileEpCacheEntryCreate(pEndpoint,
1855	pEndpointCache,
1856	off, cbRead,
1857	PAGE_SIZE,
1858	&cbToRead);
1859
1860	cbRead -= cbToRead;
1861
1862	if (pEntryNew)
1863	{
1864	if (!cbRead)
1865	STAM_COUNTER_INC(&pCache->cMisses);
1866	else
1867	STAM_COUNTER_INC(&pCache->cPartialHits);
1868
1869	pdmacFileEpCacheEntryWaitersAdd(pEntryNew, pTask,
1870	&IoMemCtx,
1871	off - pEntryNew->Core.Key,
1872	cbToRead,
1873	false /* fWrite */);
1874	pdmacFileCacheReadFromEndpoint(pEntryNew);
1875	pdmacFileEpCacheEntryRelease(pEntryNew); /* it is protected by the I/O in progress flag now. */
1876	}
1877	else
1878	{
1879	/*
1880	* There is not enough free space in the cache.
1881	* Pass the request directly to the I/O manager.
1882	*/
1883	LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToRead));
1884
1885	pdmacFileEpCacheRequestPassthrough(pEndpoint, pTask,
1886	&IoMemCtx, off, cbToRead,
1887	PDMACTASKFILETRANSFER_READ);
1888	}
1889	#else
1890	/* Clip read size if necessary. */
1891	PPDMACFILECACHEENTRY pEntryAbove;
1892	pdmacFileEpCacheGetCacheBestFitEntryByOffset(pEndpointCache, off,
1893	&pEntryAbove, NULL);
1894
1895	if (pEntryAbove)
1896	{
1897	if (off + (RTFOFF)cbRead > pEntryAbove->Core.Key)
1898	cbToRead = pEntryAbove->Core.Key - off;
1899	else
1900	cbToRead = cbRead;
1901
1902	pdmacFileEpCacheEntryRelease(pEntryAbove);
1903	}
1904	else
1905	cbToRead = cbRead;
1906
1907	cbRead -= cbToRead;
1908	pdmacFileEpCacheRequestPassthrough(pEndpoint, pTask,
1909	&IoMemCtx, off, cbToRead,
1910	PDMACTASKFILETRANSFER_READ);
1911	#endif
1912	}
1913	off += cbToRead;
1914	}
1915
1916	ASMAtomicWriteBool(&pTask->fCompleted, false);
1917
1918	if (ASMAtomicReadS32(&pTask->cbTransferLeft) == 0
1919	&& !ASMAtomicXchgBool(&pTask->fCompleted, true))
1920	pdmR3AsyncCompletionCompleteTask(&pTask->Core, VINF_SUCCESS, false);
1921	else
1922	rc = VINF_AIO_TASK_PENDING;
1923
1924	LogFlowFunc((": Leave rc=%Rrc\n", rc));
1925
1926	return rc;
1927	}
1928
1929	/**
1930	* Writes the given data to the endpoint using the cache if possible.
1931	*
1932	* @returns VBox status code.
1933	* @param pEndpoint The endpoint to write to.
1934	* @param pTask The task structure used as identifier for this request.
1935	* @param off The offset to start writing to
1936	* @param paSegments Pointer to the array holding the source buffers.
1937	* @param cSegments Number of segments in the array.
1938	* @param cbWrite Number of bytes to write.
1939	*/
1940	int pdmacFileEpCacheWrite(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMASYNCCOMPLETIONTASKFILE pTask,
1941	RTFOFF off, PCRTSGSEG paSegments, size_t cSegments,
1942	size_t cbWrite)
1943	{
1944	int rc = VINF_SUCCESS;
1945	PPDMACFILEENDPOINTCACHE pEndpointCache = &pEndpoint->DataCache;
1946	PPDMACFILECACHEGLOBAL pCache = pEndpointCache->pCache;
1947	PPDMACFILECACHEENTRY pEntry;
1948
1949	LogFlowFunc((": pEndpoint=%#p{%s} pTask=%#p off=%RTfoff paSegments=%#p cSegments=%u cbWrite=%u\n",
1950	pEndpoint, pEndpoint->Core.pszUri, pTask, off, paSegments, cSegments, cbWrite));
1951
1952	/* Set to completed to make sure that the task is valid while we access it. */
1953	ASMAtomicWriteBool(&pTask->fCompleted, true);
1954
1955	/* Init the I/O memory context */
1956	PDMIOMEMCTX IoMemCtx;
1957	pdmIoMemCtxInit(&IoMemCtx, paSegments, cSegments);
1958
1959	while (cbWrite)
1960	{
1961	size_t cbToWrite;
1962
1963	pEntry = pdmacFileEpCacheGetCacheEntryByOffset(pEndpointCache, off);
1964
1965	if (pEntry)
1966	{
1967	/* Write the data into the entry and mark it as dirty */
1968	AssertPtr(pEntry->pList);
1969
1970	RTFOFF OffDiff = off - pEntry->Core.Key;
1971
1972	AssertMsg(off >= pEntry->Core.Key,
1973	("Overflow in calculation off=%RTfoff OffsetAligned=%RTfoff\n",
1974	off, pEntry->Core.Key));
1975
1976	cbToWrite = RT_MIN(pEntry->cbData - OffDiff, cbWrite);
1977	cbWrite -= cbToWrite;
1978
1979	if (!cbWrite)
1980	STAM_COUNTER_INC(&pCache->cHits);
1981	else
1982	STAM_COUNTER_INC(&pCache->cPartialHits);
1983
1984	STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
1985
1986	/* Ghost lists contain no data. */
1987	if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
1988	\|\| (pEntry->pList == &pCache->LruFrequentlyUsed))
1989	{
1990	/* Check if the entry is dirty. */
1991	if(pdmacFileEpCacheEntryFlagIsSetClearAcquireLock(pEndpointCache, pEntry,
1992	PDMACFILECACHE_ENTRY_IS_DIRTY,
1993	0))
1994	{
1995	/* If it is dirty but not in progress just update the data. */
1996	if (!(pEntry->fFlags & PDMACFILECACHE_ENTRY_IO_IN_PROGRESS))
1997	{
1998	pdmacFileEpCacheCopyFromIoMemCtx(&IoMemCtx,
1999	pEntry->pbData + OffDiff,
2000	cbToWrite);
2001	ASMAtomicSubS32(&pTask->cbTransferLeft, cbToWrite);
2002	}
2003	else
2004	{
2005	/* The data isn't written to the file yet */
2006	pdmacFileEpCacheEntryWaitersAdd(pEntry, pTask,
2007	&IoMemCtx,
2008	OffDiff, cbToWrite,
2009	true /* fWrite */);
2010	STAM_COUNTER_INC(&pEndpointCache->StatWriteDeferred);
2011	}
2012
2013	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
2014	}
2015	else /* Dirty bit not set */
2016	{
2017	/*
2018	* Check if a read is in progress for this entry.
2019	* We have to defer processing in that case.
2020	*/
2021	if(pdmacFileEpCacheEntryFlagIsSetClearAcquireLock(pEndpointCache, pEntry,
2022	PDMACFILECACHE_ENTRY_IO_IN_PROGRESS,
2023	0))
2024	{
2025	pdmacFileEpCacheEntryWaitersAdd(pEntry, pTask,
2026	&IoMemCtx,
2027	OffDiff, cbToWrite,
2028	true /* fWrite */);
2029	STAM_COUNTER_INC(&pEndpointCache->StatWriteDeferred);
2030	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
2031	}
2032	else /* I/O in progress flag not set */
2033	{
2034	/* Write as much as we can into the entry and update the file. */
2035	pdmacFileEpCacheCopyFromIoMemCtx(&IoMemCtx,
2036	pEntry->pbData + OffDiff,
2037	cbToWrite);
2038	ASMAtomicSubS32(&pTask->cbTransferLeft, cbToWrite);
2039
2040	bool fCommit = pdmacFileCacheAddDirtyEntry(pEndpointCache, pEntry);
2041	if (fCommit)
2042	pdmacFileCacheCommitDirtyEntries(pCache);
2043	}
2044	} /* Dirty bit not set */
2045
2046	/* Move this entry to the top position */
2047	if (pEntry->pList == &pCache->LruFrequentlyUsed)
2048	{
2049	pdmacFileCacheLockEnter(pCache);
2050	pdmacFileCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2051	pdmacFileCacheLockLeave(pCache);
2052	}
2053
2054	pdmacFileEpCacheEntryRelease(pEntry);
2055	}
2056	else /* Entry is on the ghost list */
2057	{
2058	uint8_t *pbBuffer = NULL;
2059
2060	pdmacFileCacheLockEnter(pCache);
2061	pdmacFileCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2062	bool fEnough = pdmacFileCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2063
2064	if (fEnough)
2065	{
2066	/* Move the entry to Am and fetch it to the cache. */
2067	pdmacFileCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2068	pdmacFileCacheAdd(pCache, pEntry->cbData);
2069	pdmacFileCacheLockLeave(pCache);
2070
2071	if (pbBuffer)
2072	pEntry->pbData = pbBuffer;
2073	else
2074	pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2075	AssertPtr(pEntry->pbData);
2076
2077	pdmacFileEpCacheEntryWaitersAdd(pEntry, pTask,
2078	&IoMemCtx,
2079	OffDiff, cbToWrite,
2080	true /* fWrite */);
2081	STAM_COUNTER_INC(&pEndpointCache->StatWriteDeferred);
2082	pdmacFileCacheReadFromEndpoint(pEntry);
2083
2084	/* Release the reference. If it is still needed the I/O in progress flag should protect it now. */
2085	pdmacFileEpCacheEntryRelease(pEntry);
2086	}
2087	else
2088	{
2089	RTSemRWRequestWrite(pEndpointCache->SemRWEntries, RT_INDEFINITE_WAIT);
2090	STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2091	RTAvlrFileOffsetRemove(pEndpointCache->pTree, pEntry->Core.Key);
2092	STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2093	RTSemRWReleaseWrite(pEndpointCache->SemRWEntries);
2094
2095	pdmacFileCacheLockLeave(pCache);
2096
2097	RTMemFree(pEntry);
2098	pdmacFileEpCacheRequestPassthrough(pEndpoint, pTask,
2099	&IoMemCtx, off, cbToWrite,
2100	PDMACTASKFILETRANSFER_WRITE);
2101	}
2102	}
2103	}
2104	else /* No entry found */
2105	{
2106	/*
2107	* No entry found. Try to create a new cache entry to store the data in and if that fails
2108	* write directly to the file.
2109	*/
2110	PPDMACFILECACHEENTRY pEntryNew = pdmacFileEpCacheEntryCreate(pEndpoint,
2111	pEndpointCache,
2112	off, cbWrite,
2113	512,
2114	&cbToWrite);
2115
2116	cbWrite -= cbToWrite;
2117
2118	if (pEntryNew)
2119	{
2120	RTFOFF offDiff = off - pEntryNew->Core.Key;
2121
2122	STAM_COUNTER_INC(&pCache->cHits);
2123
2124	/*
2125	* Check if it is possible to just write the data without waiting
2126	* for it to get fetched first.
2127	*/
2128	if (!offDiff && pEntryNew->cbData == cbToWrite)
2129	{
2130	pdmacFileEpCacheCopyFromIoMemCtx(&IoMemCtx,
2131	pEntryNew->pbData,
2132	cbToWrite);
2133	ASMAtomicSubS32(&pTask->cbTransferLeft, cbToWrite);
2134
2135	bool fCommit = pdmacFileCacheAddDirtyEntry(pEndpointCache, pEntryNew);
2136	if (fCommit)
2137	pdmacFileCacheCommitDirtyEntries(pCache);
2138	STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2139	}
2140	else
2141	{
2142	/* Defer the write and fetch the data from the endpoint. */
2143	pdmacFileEpCacheEntryWaitersAdd(pEntryNew, pTask,
2144	&IoMemCtx,
2145	offDiff, cbToWrite,
2146	true /* fWrite */);
2147	STAM_COUNTER_INC(&pEndpointCache->StatWriteDeferred);
2148	pdmacFileCacheReadFromEndpoint(pEntryNew);
2149	}
2150
2151	pdmacFileEpCacheEntryRelease(pEntryNew);
2152	}
2153	else
2154	{
2155	/*
2156	* There is not enough free space in the cache.
2157	* Pass the request directly to the I/O manager.
2158	*/
2159	LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToWrite));
2160
2161	STAM_COUNTER_INC(&pCache->cMisses);
2162
2163	pdmacFileEpCacheRequestPassthrough(pEndpoint, pTask,
2164	&IoMemCtx, off, cbToWrite,
2165	PDMACTASKFILETRANSFER_WRITE);
2166	}
2167	}
2168
2169	off += cbToWrite;
2170	}
2171
2172	ASMAtomicWriteBool(&pTask->fCompleted, false);
2173
2174	if (ASMAtomicReadS32(&pTask->cbTransferLeft) == 0
2175	&& !ASMAtomicXchgBool(&pTask->fCompleted, true))
2176	pdmR3AsyncCompletionCompleteTask(&pTask->Core, VINF_SUCCESS, false);
2177	else
2178	rc = VINF_AIO_TASK_PENDING;
2179
2180	LogFlowFunc((": Leave rc=%Rrc\n", rc));
2181
2182	return rc;
2183	}
2184
2185	int pdmacFileEpCacheFlush(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
2186	{
2187	int rc = VINF_SUCCESS;
2188
2189	LogFlowFunc((": pEndpoint=%#p{%s}\n", pEndpoint, pEndpoint->Core.pszUri));
2190
2191	/* Commit dirty entries in the cache. */
2192	pdmacFileCacheEndpointCommit(&pEndpoint->DataCache);
2193
2194	LogFlowFunc((": Leave rc=%Rrc\n", rc));
2195	return rc;
2196	}
2197

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

source: vbox/trunk/src/VBox/VMM/PDMAsyncCompletionFileCache.cpp@ 34646

Download in other formats: