VirtualBox

source: vbox/trunk/src/VBox/VMM/PDMBlkCache.cpp@ 34223

Last change on this file since 34223 was 34223, checked in by vboxsync, 14 years ago

BlockCache: Bug fixes

  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 84.1 KB
Line 
1/* $Id: PDMBlkCache.cpp 34223 2010-11-21 23:51:36Z vboxsync $ */
2/** @file
3 * PDM Block Cache.
4 */
5
6/*
7 * Copyright (C) 2006-2008 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/** @page pg_pdm_block_cache PDM Block Cache - The I/O cache
19 * This component implements an I/O cache based on the 2Q cache algorithm.
20 */
21
22/*******************************************************************************
23* Header Files *
24*******************************************************************************/
25#define LOG_GROUP LOG_GROUP_PDM_BLK_CACHE
26#include "PDMInternal.h"
27#include <iprt/asm.h>
28#include <iprt/mem.h>
29#include <iprt/path.h>
30#include <iprt/string.h>
31#include <VBox/log.h>
32#include <VBox/stam.h>
33#include <VBox/uvm.h>
34#include <VBox/vm.h>
35
36#include "PDMBlkCacheInternal.h"
37
38#ifdef VBOX_STRICT
39# define PDMACFILECACHE_IS_CRITSECT_OWNER(Cache) \
40 do \
41 { \
42 AssertMsg(RTCritSectIsOwner(&Cache->CritSect), \
43 ("Thread does not own critical section\n"));\
44 } while(0)
45
46# define PDMACFILECACHE_EP_IS_SEMRW_WRITE_OWNER(pEpCache) \
47 do \
48 { \
49 AssertMsg(RTSemRWIsWriteOwner(pEpCache->SemRWEntries), \
50 ("Thread is not exclusive owner of the per endpoint RW semaphore\n")); \
51 } while(0)
52
53# define PDMACFILECACHE_EP_IS_SEMRW_READ_OWNER(pEpCache) \
54 do \
55 { \
56 AssertMsg(RTSemRWIsReadOwner(pEpCache->SemRWEntries), \
57 ("Thread is not read owner of the per endpoint RW semaphore\n")); \
58 } while(0)
59
60#else
61# define PDMACFILECACHE_IS_CRITSECT_OWNER(Cache) do { } while(0)
62# define PDMACFILECACHE_EP_IS_SEMRW_WRITE_OWNER(pEpCache) do { } while(0)
63# define PDMACFILECACHE_EP_IS_SEMRW_READ_OWNER(pEpCache) do { } while(0)
64#endif
65
66/*******************************************************************************
67* Internal Functions *
68*******************************************************************************/
69
70/**
71 * Decrement the reference counter of the given cache entry.
72 *
73 * @returns nothing.
74 * @param pEntry The entry to release.
75 */
76DECLINLINE(void) pdmBlkCacheEntryRelease(PPDMBLKCACHEENTRY pEntry)
77{
78 AssertMsg(pEntry->cRefs > 0, ("Trying to release a not referenced entry\n"));
79 ASMAtomicDecU32(&pEntry->cRefs);
80}
81
82/**
83 * Increment the reference counter of the given cache entry.
84 *
85 * @returns nothing.
86 * @param pEntry The entry to reference.
87 */
88DECLINLINE(void) pdmBlkCacheEntryRef(PPDMBLKCACHEENTRY pEntry)
89{
90 ASMAtomicIncU32(&pEntry->cRefs);
91}
92
93#ifdef DEBUG
94static void pdmBlkCacheValidate(PPDMBLKCACHEGLOBAL pCache)
95{
96 /* Amount of cached data should never exceed the maximum amount. */
97 AssertMsg(pCache->cbCached <= pCache->cbMax,
98 ("Current amount of cached data exceeds maximum\n"));
99
100 /* The amount of cached data in the LRU and FRU list should match cbCached */
101 AssertMsg(pCache->LruRecentlyUsedIn.cbCached + pCache->LruFrequentlyUsed.cbCached == pCache->cbCached,
102 ("Amount of cached data doesn't match\n"));
103
104 AssertMsg(pCache->LruRecentlyUsedOut.cbCached <= pCache->cbRecentlyUsedOutMax,
105 ("Paged out list exceeds maximum\n"));
106}
107#endif
108
109DECLINLINE(void) pdmBlkCacheLockEnter(PPDMBLKCACHEGLOBAL pCache)
110{
111 RTCritSectEnter(&pCache->CritSect);
112#ifdef DEBUG
113 pdmBlkCacheValidate(pCache);
114#endif
115}
116
117DECLINLINE(void) pdmBlkCacheLockLeave(PPDMBLKCACHEGLOBAL pCache)
118{
119#ifdef DEBUG
120 pdmBlkCacheValidate(pCache);
121#endif
122 RTCritSectLeave(&pCache->CritSect);
123}
124
125DECLINLINE(void) pdmBlkCacheSub(PPDMBLKCACHEGLOBAL pCache, uint32_t cbAmount)
126{
127 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
128 pCache->cbCached -= cbAmount;
129}
130
131DECLINLINE(void) pdmBlkCacheAdd(PPDMBLKCACHEGLOBAL pCache, uint32_t cbAmount)
132{
133 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
134 pCache->cbCached += cbAmount;
135}
136
137DECLINLINE(void) pdmBlkCacheListAdd(PPDMBLKLRULIST pList, uint32_t cbAmount)
138{
139 pList->cbCached += cbAmount;
140}
141
142DECLINLINE(void) pdmBlkCacheListSub(PPDMBLKLRULIST pList, uint32_t cbAmount)
143{
144 pList->cbCached -= cbAmount;
145}
146
147#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
148/**
149 * Checks consistency of a LRU list.
150 *
151 * @returns nothing
152 * @param pList The LRU list to check.
153 * @param pNotInList Element which is not allowed to occur in the list.
154 */
155static void pdmBlkCacheCheckList(PPDMBLKLRULIST pList, PPDMBLKCACHEENTRY pNotInList)
156{
157 PPDMBLKCACHEENTRY pCurr = pList->pHead;
158
159 /* Check that there are no double entries and no cycles in the list. */
160 while (pCurr)
161 {
162 PPDMBLKCACHEENTRY pNext = pCurr->pNext;
163
164 while (pNext)
165 {
166 AssertMsg(pCurr != pNext,
167 ("Entry %#p is at least two times in list %#p or there is a cycle in the list\n",
168 pCurr, pList));
169 pNext = pNext->pNext;
170 }
171
172 AssertMsg(pCurr != pNotInList, ("Not allowed entry %#p is in list\n", pCurr));
173
174 if (!pCurr->pNext)
175 AssertMsg(pCurr == pList->pTail, ("End of list reached but last element is not list tail\n"));
176
177 pCurr = pCurr->pNext;
178 }
179}
180#endif
181
182/**
183 * Unlinks a cache entry from the LRU list it is assigned to.
184 *
185 * @returns nothing.
186 * @param pEntry The entry to unlink.
187 */
188static void pdmBlkCacheEntryRemoveFromList(PPDMBLKCACHEENTRY pEntry)
189{
190 PPDMBLKLRULIST pList = pEntry->pList;
191 PPDMBLKCACHEENTRY pPrev, pNext;
192
193 LogFlowFunc((": Deleting entry %#p from list %#p\n", pEntry, pList));
194
195 AssertPtr(pList);
196
197#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
198 pdmBlkCacheCheckList(pList, NULL);
199#endif
200
201 pPrev = pEntry->pPrev;
202 pNext = pEntry->pNext;
203
204 AssertMsg(pEntry != pPrev, ("Entry links to itself as previous element\n"));
205 AssertMsg(pEntry != pNext, ("Entry links to itself as next element\n"));
206
207 if (pPrev)
208 pPrev->pNext = pNext;
209 else
210 {
211 pList->pHead = pNext;
212
213 if (pNext)
214 pNext->pPrev = NULL;
215 }
216
217 if (pNext)
218 pNext->pPrev = pPrev;
219 else
220 {
221 pList->pTail = pPrev;
222
223 if (pPrev)
224 pPrev->pNext = NULL;
225 }
226
227 pEntry->pList = NULL;
228 pEntry->pPrev = NULL;
229 pEntry->pNext = NULL;
230 pdmBlkCacheListSub(pList, pEntry->cbData);
231#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
232 pdmBlkCacheCheckList(pList, pEntry);
233#endif
234}
235
236/**
237 * Adds a cache entry to the given LRU list unlinking it from the currently
238 * assigned list if needed.
239 *
240 * @returns nothing.
241 * @param pList List to the add entry to.
242 * @param pEntry Entry to add.
243 */
244static void pdmBlkCacheEntryAddToList(PPDMBLKLRULIST pList, PPDMBLKCACHEENTRY pEntry)
245{
246 LogFlowFunc((": Adding entry %#p to list %#p\n", pEntry, pList));
247#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
248 pdmBlkCacheCheckList(pList, NULL);
249#endif
250
251 /* Remove from old list if needed */
252 if (pEntry->pList)
253 pdmBlkCacheEntryRemoveFromList(pEntry);
254
255 pEntry->pNext = pList->pHead;
256 if (pList->pHead)
257 pList->pHead->pPrev = pEntry;
258 else
259 {
260 Assert(!pList->pTail);
261 pList->pTail = pEntry;
262 }
263
264 pEntry->pPrev = NULL;
265 pList->pHead = pEntry;
266 pdmBlkCacheListAdd(pList, pEntry->cbData);
267 pEntry->pList = pList;
268#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
269 pdmBlkCacheCheckList(pList, NULL);
270#endif
271}
272
273/**
274 * Destroys a LRU list freeing all entries.
275 *
276 * @returns nothing
277 * @param pList Pointer to the LRU list to destroy.
278 *
279 * @note The caller must own the critical section of the cache.
280 */
281static void pdmBlkCacheDestroyList(PPDMBLKLRULIST pList)
282{
283 while (pList->pHead)
284 {
285 PPDMBLKCACHEENTRY pEntry = pList->pHead;
286
287 pList->pHead = pEntry->pNext;
288
289 AssertMsg(!(pEntry->fFlags & (PDMBLKCACHE_ENTRY_IO_IN_PROGRESS | PDMBLKCACHE_ENTRY_IS_DIRTY)),
290 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
291
292 RTMemPageFree(pEntry->pbData, pEntry->cbData);
293 RTMemFree(pEntry);
294 }
295}
296
297/**
298 * Tries to remove the given amount of bytes from a given list in the cache
299 * moving the entries to one of the given ghosts lists
300 *
301 * @returns Amount of data which could be freed.
302 * @param pCache Pointer to the global cache data.
303 * @param cbData The amount of the data to free.
304 * @param pListSrc The source list to evict data from.
305 * @param pGhostListSrc The ghost list removed entries should be moved to
306 * NULL if the entry should be freed.
307 * @param fReuseBuffer Flag whether a buffer should be reused if it has the same size
308 * @param ppbBuf Where to store the address of the buffer if an entry with the
309 * same size was found and fReuseBuffer is true.
310 *
311 * @note This function may return fewer bytes than requested because entries
312 * may be marked as non evictable if they are used for I/O at the
313 * moment.
314 */
315static size_t pdmBlkCacheEvictPagesFrom(PPDMBLKCACHEGLOBAL pCache, size_t cbData,
316 PPDMBLKLRULIST pListSrc, PPDMBLKLRULIST pGhostListDst,
317 bool fReuseBuffer, uint8_t **ppbBuffer)
318{
319 size_t cbEvicted = 0;
320
321 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
322
323 AssertMsg(cbData > 0, ("Evicting 0 bytes not possible\n"));
324 AssertMsg( !pGhostListDst
325 || (pGhostListDst == &pCache->LruRecentlyUsedOut),
326 ("Destination list must be NULL or the recently used but paged out list\n"));
327
328 if (fReuseBuffer)
329 {
330 AssertPtr(ppbBuffer);
331 *ppbBuffer = NULL;
332 }
333
334 /* Start deleting from the tail. */
335 PPDMBLKCACHEENTRY pEntry = pListSrc->pTail;
336
337 while ((cbEvicted < cbData) && pEntry)
338 {
339 PPDMBLKCACHEENTRY pCurr = pEntry;
340
341 pEntry = pEntry->pPrev;
342
343 /* We can't evict pages which are currently in progress or dirty but not in progress */
344 if ( !(pCurr->fFlags & PDMBLKCACHE_NOT_EVICTABLE)
345 && (ASMAtomicReadU32(&pCurr->cRefs) == 0))
346 {
347 /* Ok eviction candidate. Grab the endpoint semaphore and check again
348 * because somebody else might have raced us. */
349 PPDMBLKCACHE pBlkCache = pCurr->pBlkCache;
350 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
351
352 if (!(pCurr->fFlags & PDMBLKCACHE_NOT_EVICTABLE)
353 && (ASMAtomicReadU32(&pCurr->cRefs) == 0))
354 {
355 LogFlow(("Evicting entry %#p (%u bytes)\n", pCurr, pCurr->cbData));
356
357 if (fReuseBuffer && (pCurr->cbData == cbData))
358 {
359 STAM_COUNTER_INC(&pCache->StatBuffersReused);
360 *ppbBuffer = pCurr->pbData;
361 }
362 else if (pCurr->pbData)
363 RTMemPageFree(pCurr->pbData, pCurr->cbData);
364
365 pCurr->pbData = NULL;
366 cbEvicted += pCurr->cbData;
367
368 pdmBlkCacheEntryRemoveFromList(pCurr);
369 pdmBlkCacheSub(pCache, pCurr->cbData);
370
371 if (pGhostListDst)
372 {
373 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
374
375 PPDMBLKCACHEENTRY pGhostEntFree = pGhostListDst->pTail;
376
377 /* We have to remove the last entries from the paged out list. */
378 while ( ((pGhostListDst->cbCached + pCurr->cbData) > pCache->cbRecentlyUsedOutMax)
379 && pGhostEntFree)
380 {
381 PPDMBLKCACHEENTRY pFree = pGhostEntFree;
382 PPDMBLKCACHE pBlkCacheFree = pFree->pBlkCache;
383
384 pGhostEntFree = pGhostEntFree->pPrev;
385
386 RTSemRWRequestWrite(pBlkCacheFree->SemRWEntries, RT_INDEFINITE_WAIT);
387
388 if (ASMAtomicReadU32(&pFree->cRefs) == 0)
389 {
390 pdmBlkCacheEntryRemoveFromList(pFree);
391
392 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
393 RTAvlrU64Remove(pBlkCacheFree->pTree, pFree->Core.Key);
394 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
395
396 RTMemFree(pFree);
397 }
398
399 RTSemRWReleaseWrite(pBlkCacheFree->SemRWEntries);
400 }
401
402 if (pGhostListDst->cbCached + pCurr->cbData > pCache->cbRecentlyUsedOutMax)
403 {
404 /* Couldn't remove enough entries. Delete */
405 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
406 RTAvlrU64Remove(pCurr->pBlkCache->pTree, pCurr->Core.Key);
407 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
408
409 RTMemFree(pCurr);
410 }
411 else
412 pdmBlkCacheEntryAddToList(pGhostListDst, pCurr);
413 }
414 else
415 {
416 /* Delete the entry from the AVL tree it is assigned to. */
417 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
418 RTAvlrU64Remove(pCurr->pBlkCache->pTree, pCurr->Core.Key);
419 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
420
421 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
422 RTMemFree(pCurr);
423 }
424 }
425
426 }
427 else
428 LogFlow(("Entry %#p (%u bytes) is still in progress and can't be evicted\n", pCurr, pCurr->cbData));
429 }
430
431 return cbEvicted;
432}
433
434static bool pdmBlkCacheReclaim(PPDMBLKCACHEGLOBAL pCache, size_t cbData, bool fReuseBuffer, uint8_t **ppbBuffer)
435{
436 size_t cbRemoved = 0;
437
438 if ((pCache->cbCached + cbData) < pCache->cbMax)
439 return true;
440 else if ((pCache->LruRecentlyUsedIn.cbCached + cbData) > pCache->cbRecentlyUsedInMax)
441 {
442 /* Try to evict as many bytes as possible from A1in */
443 cbRemoved = pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruRecentlyUsedIn,
444 &pCache->LruRecentlyUsedOut, fReuseBuffer, ppbBuffer);
445
446 /*
447 * If it was not possible to remove enough entries
448 * try the frequently accessed cache.
449 */
450 if (cbRemoved < cbData)
451 {
452 Assert(!fReuseBuffer || !*ppbBuffer); /* It is not possible that we got a buffer with the correct size but we didn't freed enough data. */
453
454 /*
455 * If we removed something we can't pass the reuse buffer flag anymore because
456 * we don't need to evict that much data
457 */
458 if (!cbRemoved)
459 cbRemoved += pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruFrequentlyUsed,
460 NULL, fReuseBuffer, ppbBuffer);
461 else
462 cbRemoved += pdmBlkCacheEvictPagesFrom(pCache, cbData - cbRemoved, &pCache->LruFrequentlyUsed,
463 NULL, false, NULL);
464 }
465 }
466 else
467 {
468 /* We have to remove entries from frequently access list. */
469 cbRemoved = pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruFrequentlyUsed,
470 NULL, fReuseBuffer, ppbBuffer);
471 }
472
473 LogFlowFunc((": removed %u bytes, requested %u\n", cbRemoved, cbData));
474 return (cbRemoved >= cbData);
475}
476
477DECLINLINE(int) pdmBlkCacheEnqueue(PPDMBLKCACHE pBlkCache, uint64_t off, PPDMBLKCACHEIOXFER pIoXfer)
478{
479 int rc = VINF_SUCCESS;
480
481 switch (pBlkCache->enmType)
482 {
483 case PDMBLKCACHETYPE_DEV:
484 {
485 rc = pBlkCache->u.Dev.pfnXferEnqueue(pBlkCache->u.Dev.pDevIns,
486 pIoXfer->enmXferDir,
487 off, pIoXfer->cbXfer,
488 &pIoXfer->SgBuf, pIoXfer);
489 break;
490 }
491 case PDMBLKCACHETYPE_DRV:
492 {
493 rc = pBlkCache->u.Drv.pfnXferEnqueue(pBlkCache->u.Drv.pDrvIns,
494 pIoXfer->enmXferDir,
495 off, pIoXfer->cbXfer,
496 &pIoXfer->SgBuf, pIoXfer);
497 break;
498 }
499 case PDMBLKCACHETYPE_USB:
500 {
501 rc = pBlkCache->u.Usb.pfnXferEnqueue(pBlkCache->u.Usb.pUsbIns,
502 pIoXfer->enmXferDir,
503 off, pIoXfer->cbXfer,
504 &pIoXfer->SgBuf, pIoXfer);
505 break;
506 }
507 case PDMBLKCACHETYPE_INTERNAL:
508 {
509 rc = pBlkCache->u.Int.pfnXferEnqueue(pBlkCache->u.Int.pvUser,
510 pIoXfer->enmXferDir,
511 off, pIoXfer->cbXfer,
512 &pIoXfer->SgBuf, pIoXfer);
513 break;
514 }
515 default:
516 AssertMsgFailed(("Unknown block cache type!\n"));
517 }
518
519 return rc;
520}
521
522/**
523 * Initiates a read I/O task for the given entry.
524 *
525 * @returns VBox status code.
526 * @param pEntry The entry to fetch the data to.
527 */
528static int pdmBlkCacheEntryReadFromMedium(PPDMBLKCACHEENTRY pEntry)
529{
530 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
531 LogFlowFunc((": Reading data into cache entry %#p\n", pEntry));
532
533 /* Make sure no one evicts the entry while it is accessed. */
534 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
535
536 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
537 if (RT_UNLIKELY(!pIoXfer))
538 return VERR_NO_MEMORY;
539
540 AssertMsg(pEntry->pbData, ("Entry is in ghost state\n"));
541
542 pIoXfer->fIoCache = true;
543 pIoXfer->pEntry = pEntry;
544 pIoXfer->SgSeg.pvSeg = pEntry->pbData;
545 pIoXfer->SgSeg.cbSeg = pEntry->cbData;
546 pIoXfer->cbXfer = pEntry->cbData;
547 pIoXfer->enmXferDir = PDMBLKCACHEXFERDIR_READ;
548 RTSgBufInit(&pIoXfer->SgBuf, &pIoXfer->SgSeg, 1);
549
550 return pdmBlkCacheEnqueue(pBlkCache, pEntry->Core.Key, pIoXfer);
551}
552
553/**
554 * Initiates a write I/O task for the given entry.
555 *
556 * @returns nothing.
557 * @param pEntry The entry to read the data from.
558 */
559static int pdmBlkCacheEntryWriteToMedium(PPDMBLKCACHEENTRY pEntry)
560{
561 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
562 LogFlowFunc((": Writing data from cache entry %#p\n", pEntry));
563
564 /* Make sure no one evicts the entry while it is accessed. */
565 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
566
567 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
568 if (RT_UNLIKELY(!pIoXfer))
569 return VERR_NO_MEMORY;
570
571 AssertMsg(pEntry->pbData, ("Entry is in ghost state\n"));
572
573 pIoXfer->fIoCache = true;
574 pIoXfer->pEntry = pEntry;
575 pIoXfer->SgSeg.pvSeg = pEntry->pbData;
576 pIoXfer->SgSeg.cbSeg = pEntry->cbData;
577 pIoXfer->cbXfer = pEntry->cbData;
578 pIoXfer->enmXferDir = PDMBLKCACHEXFERDIR_WRITE;
579 RTSgBufInit(&pIoXfer->SgBuf, &pIoXfer->SgSeg, 1);
580
581 return pdmBlkCacheEnqueue(pBlkCache, pEntry->Core.Key, pIoXfer);
582}
583
584/**
585 * Passthrough a part of a request directly to the I/O manager
586 * handling the endpoint.
587 *
588 * @returns VBox status code.
589 * @param pEndpoint The endpoint.
590 * @param pTask The task.
591 * @param pIoMemCtx The I/O memory context to use.
592 * @param offStart Offset to start transfer from.
593 * @param cbData Amount of data to transfer.
594 * @param enmTransferType The transfer type (read/write)
595 */
596static int pdmBlkCacheRequestPassthrough(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq,
597 PRTSGBUF pSgBuf, uint64_t offStart, size_t cbData,
598 PDMBLKCACHEXFERDIR enmXferDir)
599{
600
601 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
602 if (RT_UNLIKELY(!pIoXfer))
603 return VERR_NO_MEMORY;
604
605 ASMAtomicIncU32(&pReq->cXfersPending);
606 pIoXfer->fIoCache = false;
607 pIoXfer->pReq = pReq;
608 pIoXfer->cbXfer = cbData;
609 pIoXfer->enmXferDir = enmXferDir;
610 RTSgBufClone(&pIoXfer->SgBuf, pSgBuf);
611
612 return pdmBlkCacheEnqueue(pBlkCache, offStart, pIoXfer);
613}
614
615/**
616 * Commit a single dirty entry to the endpoint
617 *
618 * @returns nothing
619 * @param pEntry The entry to commit.
620 */
621static void pdmBlkCacheEntryCommit(PPDMBLKCACHEENTRY pEntry)
622{
623 AssertMsg( (pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY)
624 && !(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
625 ("Invalid flags set for entry %#p\n", pEntry));
626
627 pdmBlkCacheEntryWriteToMedium(pEntry);
628}
629
630/**
631 * Commit all dirty entries for a single endpoint.
632 *
633 * @returns nothing.
634 * @param pBlkCache The endpoint cache to commit.
635 */
636static void pdmBlkCacheCommit(PPDMBLKCACHE pBlkCache)
637{
638 uint32_t cbCommitted = 0;
639 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
640
641 /* The list is moved to a new header to reduce locking overhead. */
642 RTLISTNODE ListDirtyNotCommitted;
643 RTSPINLOCKTMP Tmp;
644
645 RTListInit(&ListDirtyNotCommitted);
646 RTSpinlockAcquire(pBlkCache->LockList, &Tmp);
647 RTListMove(&ListDirtyNotCommitted, &pBlkCache->ListDirtyNotCommitted);
648 RTSpinlockRelease(pBlkCache->LockList, &Tmp);
649
650 if (!RTListIsEmpty(&ListDirtyNotCommitted))
651 {
652 PPDMBLKCACHEENTRY pEntry = RTListNodeGetFirst(&ListDirtyNotCommitted,
653 PDMBLKCACHEENTRY,
654 NodeNotCommitted);
655
656 while (!RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted))
657 {
658 PPDMBLKCACHEENTRY pNext = RTListNodeGetNext(&pEntry->NodeNotCommitted, PDMBLKCACHEENTRY,
659 NodeNotCommitted);
660 pdmBlkCacheEntryCommit(pEntry);
661 cbCommitted += pEntry->cbData;
662 RTListNodeRemove(&pEntry->NodeNotCommitted);
663 pEntry = pNext;
664 }
665
666 /* Commit the last endpoint */
667 Assert(RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted));
668 pdmBlkCacheEntryCommit(pEntry);
669 RTListNodeRemove(&pEntry->NodeNotCommitted);
670 AssertMsg(RTListIsEmpty(&ListDirtyNotCommitted),
671 ("Committed all entries but list is not empty\n"));
672 }
673
674 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
675 AssertMsg(pBlkCache->pCache->cbDirty >= cbCommitted,
676 ("Number of committed bytes exceeds number of dirty bytes\n"));
677 ASMAtomicSubU32(&pBlkCache->pCache->cbDirty, cbCommitted);
678}
679
680/**
681 * Commit all dirty entries in the cache.
682 *
683 * @returns nothing.
684 * @param pCache The global cache instance.
685 */
686static void pdmBlkCacheCommitDirtyEntries(PPDMBLKCACHEGLOBAL pCache)
687{
688 bool fCommitInProgress = ASMAtomicXchgBool(&pCache->fCommitInProgress, true);
689
690 if (!fCommitInProgress)
691 {
692 pdmBlkCacheLockEnter(pCache);
693 Assert(!RTListIsEmpty(&pCache->ListUsers));
694
695 PPDMBLKCACHE pBlkCache = RTListNodeGetFirst(&pCache->ListUsers,
696 PDMBLKCACHE,
697 NodeCacheUser);
698 AssertPtr(pBlkCache);
699
700 while (!RTListNodeIsLast(&pCache->ListUsers, &pBlkCache->NodeCacheUser))
701 {
702 pdmBlkCacheCommit(pBlkCache);
703
704 pBlkCache = RTListNodeGetNext(&pBlkCache->NodeCacheUser, PDMBLKCACHE,
705 NodeCacheUser);
706 }
707
708 /* Commit the last endpoint */
709 Assert(RTListNodeIsLast(&pCache->ListUsers, &pBlkCache->NodeCacheUser));
710 pdmBlkCacheCommit(pBlkCache);
711
712 pdmBlkCacheLockLeave(pCache);
713 ASMAtomicWriteBool(&pCache->fCommitInProgress, false);
714 }
715}
716
717/**
718 * Adds the given entry as a dirty to the cache.
719 *
720 * @returns Flag whether the amount of dirty bytes in the cache exceeds the threshold
721 * @param pBlkCache The endpoint cache the entry belongs to.
722 * @param pEntry The entry to add.
723 */
724static bool pdmBlkCacheAddDirtyEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
725{
726 bool fDirtyBytesExceeded = false;
727 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
728
729 /* If the commit timer is disabled we commit right away. */
730 if (pCache->u32CommitTimeoutMs == 0)
731 {
732 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IS_DIRTY;
733 pdmBlkCacheEntryCommit(pEntry);
734 }
735 else if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY))
736 {
737 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IS_DIRTY;
738
739 RTSPINLOCKTMP Tmp;
740 RTSpinlockAcquire(pBlkCache->LockList, &Tmp);
741 RTListAppend(&pBlkCache->ListDirtyNotCommitted, &pEntry->NodeNotCommitted);
742 RTSpinlockRelease(pBlkCache->LockList, &Tmp);
743
744 uint32_t cbDirty = ASMAtomicAddU32(&pCache->cbDirty, pEntry->cbData);
745
746 fDirtyBytesExceeded = (cbDirty >= pCache->cbCommitDirtyThreshold);
747 }
748
749 return fDirtyBytesExceeded;
750}
751
752/**
753 * Commit timer callback.
754 */
755static void pdmBlkCacheCommitTimerCallback(PVM pVM, PTMTIMER pTimer, void *pvUser)
756{
757 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
758
759 LogFlowFunc(("Commit interval expired, commiting dirty entries\n"));
760
761 if (ASMAtomicReadU32(&pCache->cbDirty) > 0)
762 pdmBlkCacheCommitDirtyEntries(pCache);
763
764 TMTimerSetMillies(pTimer, pCache->u32CommitTimeoutMs);
765 LogFlowFunc(("Entries committed, going to sleep\n"));
766}
767
768int pdmR3BlkCacheInit(PVM pVM)
769{
770 int rc = VINF_SUCCESS;
771 PUVM pUVM = pVM->pUVM;
772 PPDMBLKCACHEGLOBAL pBlkCacheGlobal;
773
774 LogFlowFunc((": pVM=%p\n", pVM));
775
776 VM_ASSERT_EMT(pVM);
777
778 PCFGMNODE pCfgRoot = CFGMR3GetRoot(pVM);
779 PCFGMNODE pCfgBlkCache = CFGMR3GetChild(CFGMR3GetChild(pCfgRoot, "PDM"), "BlkCache");
780
781 pBlkCacheGlobal = (PPDMBLKCACHEGLOBAL)RTMemAllocZ(sizeof(PDMBLKCACHEGLOBAL));
782 if (!pBlkCacheGlobal)
783 return VERR_NO_MEMORY;
784
785 RTListInit(&pBlkCacheGlobal->ListUsers);
786 pBlkCacheGlobal->pVM = pVM;
787 pBlkCacheGlobal->cRefs = 0;
788 pBlkCacheGlobal->cbCached = 0;
789 pBlkCacheGlobal->fCommitInProgress = false;
790
791 /* Initialize members */
792 pBlkCacheGlobal->LruRecentlyUsedIn.pHead = NULL;
793 pBlkCacheGlobal->LruRecentlyUsedIn.pTail = NULL;
794 pBlkCacheGlobal->LruRecentlyUsedIn.cbCached = 0;
795
796 pBlkCacheGlobal->LruRecentlyUsedOut.pHead = NULL;
797 pBlkCacheGlobal->LruRecentlyUsedOut.pTail = NULL;
798 pBlkCacheGlobal->LruRecentlyUsedOut.cbCached = 0;
799
800 pBlkCacheGlobal->LruFrequentlyUsed.pHead = NULL;
801 pBlkCacheGlobal->LruFrequentlyUsed.pTail = NULL;
802 pBlkCacheGlobal->LruFrequentlyUsed.cbCached = 0;
803
804 do
805 {
806 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheSize", &pBlkCacheGlobal->cbMax, 5 * _1M);
807 AssertLogRelRCBreak(rc);
808 LogFlowFunc(("Maximum number of bytes cached %u\n", pBlkCacheGlobal->cbMax));
809
810 pBlkCacheGlobal->cbRecentlyUsedInMax = (pBlkCacheGlobal->cbMax / 100) * 25; /* 25% of the buffer size */
811 pBlkCacheGlobal->cbRecentlyUsedOutMax = (pBlkCacheGlobal->cbMax / 100) * 50; /* 50% of the buffer size */
812 LogFlowFunc(("cbRecentlyUsedInMax=%u cbRecentlyUsedOutMax=%u\n",
813 pBlkCacheGlobal->cbRecentlyUsedInMax, pBlkCacheGlobal->cbRecentlyUsedOutMax));
814
815 /** @todo r=aeichner: Experiment to find optimal default values */
816 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheCommitIntervalMs", &pBlkCacheGlobal->u32CommitTimeoutMs, 10000 /* 10sec */);
817 AssertLogRelRCBreak(rc);
818 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheCommitThreshold", &pBlkCacheGlobal->cbCommitDirtyThreshold, pBlkCacheGlobal->cbMax / 2);
819 AssertLogRelRCBreak(rc);
820 } while (0);
821
822 if (RT_SUCCESS(rc))
823 {
824 STAMR3Register(pVM, &pBlkCacheGlobal->cbMax,
825 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
826 "/PDM/BlkCache/cbMax",
827 STAMUNIT_BYTES,
828 "Maximum cache size");
829 STAMR3Register(pVM, &pBlkCacheGlobal->cbCached,
830 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
831 "/PDM/BlkCache/cbCached",
832 STAMUNIT_BYTES,
833 "Currently used cache");
834 STAMR3Register(pVM, &pBlkCacheGlobal->LruRecentlyUsedIn.cbCached,
835 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
836 "/PDM/BlkCache/cbCachedMruIn",
837 STAMUNIT_BYTES,
838 "Number of bytes cached in MRU list");
839 STAMR3Register(pVM, &pBlkCacheGlobal->LruRecentlyUsedOut.cbCached,
840 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
841 "/PDM/BlkCache/cbCachedMruOut",
842 STAMUNIT_BYTES,
843 "Number of bytes cached in FRU list");
844 STAMR3Register(pVM, &pBlkCacheGlobal->LruFrequentlyUsed.cbCached,
845 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
846 "/PDM/BlkCache/cbCachedFru",
847 STAMUNIT_BYTES,
848 "Number of bytes cached in FRU ghost list");
849
850#ifdef VBOX_WITH_STATISTICS
851 STAMR3Register(pVM, &pBlkCacheGlobal->cHits,
852 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
853 "/PDM/BlkCache/CacheHits",
854 STAMUNIT_COUNT, "Number of hits in the cache");
855 STAMR3Register(pVM, &pBlkCacheGlobal->cPartialHits,
856 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
857 "/PDM/BlkCache/CachePartialHits",
858 STAMUNIT_COUNT, "Number of partial hits in the cache");
859 STAMR3Register(pVM, &pBlkCacheGlobal->cMisses,
860 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
861 "/PDM/BlkCache/CacheMisses",
862 STAMUNIT_COUNT, "Number of misses when accessing the cache");
863 STAMR3Register(pVM, &pBlkCacheGlobal->StatRead,
864 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
865 "/PDM/BlkCache/CacheRead",
866 STAMUNIT_BYTES, "Number of bytes read from the cache");
867 STAMR3Register(pVM, &pBlkCacheGlobal->StatWritten,
868 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
869 "/PDM/BlkCache/CacheWritten",
870 STAMUNIT_BYTES, "Number of bytes written to the cache");
871 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeGet,
872 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
873 "/PDM/BlkCache/CacheTreeGet",
874 STAMUNIT_TICKS_PER_CALL, "Time taken to access an entry in the tree");
875 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeInsert,
876 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
877 "/PDM/BlkCache/CacheTreeInsert",
878 STAMUNIT_TICKS_PER_CALL, "Time taken to insert an entry in the tree");
879 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeRemove,
880 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
881 "/PDM/BlkCache/CacheTreeRemove",
882 STAMUNIT_TICKS_PER_CALL, "Time taken to remove an entry an the tree");
883 STAMR3Register(pVM, &pBlkCacheGlobal->StatBuffersReused,
884 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
885 "/PDM/BlkCache/CacheBuffersReused",
886 STAMUNIT_COUNT, "Number of times a buffer could be reused");
887#endif
888
889 /* Initialize the critical section */
890 rc = RTCritSectInit(&pBlkCacheGlobal->CritSect);
891 }
892
893 if (RT_SUCCESS(rc))
894 {
895 /* Create the commit timer */
896 if (pBlkCacheGlobal->u32CommitTimeoutMs > 0)
897 rc = TMR3TimerCreateInternal(pVM, TMCLOCK_REAL,
898 pdmBlkCacheCommitTimerCallback,
899 pBlkCacheGlobal,
900 "BlkCache-Commit",
901 &pBlkCacheGlobal->pTimerCommit);
902
903 if (RT_SUCCESS(rc))
904 {
905 LogRel(("BlkCache: Cache successfully initialised. Cache size is %u bytes\n", pBlkCacheGlobal->cbMax));
906 LogRel(("BlkCache: Cache commit interval is %u ms\n", pBlkCacheGlobal->u32CommitTimeoutMs));
907 LogRel(("BlkCache: Cache commit threshold is %u bytes\n", pBlkCacheGlobal->cbCommitDirtyThreshold));
908 pUVM->pdm.s.pBlkCacheGlobal = pBlkCacheGlobal;
909 return VINF_SUCCESS;
910 }
911
912 RTCritSectDelete(&pBlkCacheGlobal->CritSect);
913 }
914
915 if (pBlkCacheGlobal)
916 RTMemFree(pBlkCacheGlobal);
917
918 LogFlowFunc((": returns rc=%Rrc\n", pVM, rc));
919 return rc;
920}
921
922void pdmR3BlkCacheTerm(PVM pVM)
923{
924 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
925
926 if (pBlkCacheGlobal)
927 {
928 /* Make sure no one else uses the cache now */
929 pdmBlkCacheLockEnter(pBlkCacheGlobal);
930
931 /* Cleanup deleting all cache entries waiting for in progress entries to finish. */
932 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruRecentlyUsedIn);
933 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruRecentlyUsedOut);
934 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruFrequentlyUsed);
935
936 pdmBlkCacheLockLeave(pBlkCacheGlobal);
937
938 RTCritSectDelete(&pBlkCacheGlobal->CritSect);
939 RTMemFree(pBlkCacheGlobal);
940 pVM->pUVM->pdm.s.pBlkCacheGlobal = NULL;
941 }
942}
943
944static int pdmR3BlkCacheRetain(PVM pVM, PPPDMBLKCACHE ppBlkCache, const char *pcszId)
945{
946 int rc = VINF_SUCCESS;
947 PPDMBLKCACHE pBlkCache = NULL;
948 bool fAlreadyExists = false;
949 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
950
951 if (!pBlkCacheGlobal)
952 return VERR_NOT_SUPPORTED;
953
954 /*
955 * Check that no other user cache has the same id first,
956 * Unique id's are necessary in case the state is saved.
957 */
958 pdmBlkCacheLockEnter(pBlkCacheGlobal);
959
960 RTListForEach(&pBlkCacheGlobal->ListUsers, pBlkCache, PDMBLKCACHE, NodeCacheUser)
961 {
962 if (!RTStrCmp(pBlkCache->pszId, pcszId))
963 {
964 fAlreadyExists = true;
965 break;
966 }
967 }
968
969 if (!fAlreadyExists)
970 {
971 pBlkCache = (PPDMBLKCACHE)RTMemAllocZ(sizeof(PDMBLKCACHE));
972
973 if (pBlkCache)
974 pBlkCache->pszId = RTStrDup(pcszId);
975
976 if ( pBlkCache
977 && pBlkCache->pszId)
978 {
979 pBlkCache->pCache = pBlkCacheGlobal;
980 RTListInit(&pBlkCache->ListDirtyNotCommitted);
981
982 rc = RTSpinlockCreate(&pBlkCache->LockList);
983 if (RT_SUCCESS(rc))
984 {
985 rc = RTSemRWCreate(&pBlkCache->SemRWEntries);
986 if (RT_SUCCESS(rc))
987 {
988 pBlkCache->pTree = (PAVLRU64TREE)RTMemAllocZ(sizeof(AVLRFOFFTREE));
989 if (pBlkCache->pTree)
990 {
991 /* Arm the timer if this is the first endpoint. */
992 if ( pBlkCacheGlobal->cRefs == 1
993 && pBlkCacheGlobal->u32CommitTimeoutMs > 0)
994 rc = TMTimerSetMillies(pBlkCacheGlobal->pTimerCommit, pBlkCacheGlobal->u32CommitTimeoutMs);
995
996 if (RT_SUCCESS(rc))
997 {
998#ifdef VBOX_WITH_STATISTICS
999 STAMR3RegisterF(pBlkCacheGlobal->pVM, &pBlkCache->StatWriteDeferred,
1000 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1001 STAMUNIT_COUNT, "Number of deferred writes",
1002 "/PDM/BlkCache/%s/Cache/DeferredWrites", pBlkCache->pszId);
1003#endif
1004
1005 /* Add to the list of users. */
1006 pBlkCacheGlobal->cRefs++;
1007 RTListAppend(&pBlkCacheGlobal->ListUsers, &pBlkCache->NodeCacheUser);
1008 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1009
1010 *ppBlkCache = pBlkCache;
1011 LogFlowFunc(("returns success\n"));
1012 return VINF_SUCCESS;
1013 }
1014 }
1015 else
1016 rc = VERR_NO_MEMORY;
1017
1018 RTSemRWDestroy(pBlkCache->SemRWEntries);
1019 }
1020
1021 RTSpinlockDestroy(pBlkCache->LockList);
1022 }
1023
1024 RTStrFree(pBlkCache->pszId);
1025 }
1026 else
1027 rc = VERR_NO_MEMORY;
1028
1029 if (pBlkCache)
1030 RTMemFree(pBlkCache);
1031 }
1032 else
1033 rc = VERR_ALREADY_EXISTS;
1034
1035 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1036
1037 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1038 return rc;
1039}
1040
1041VMMR3DECL(int) PDMR3BlkCacheRetainDriver(PVM pVM, PPDMDRVINS pDrvIns, PPPDMBLKCACHE ppBlkCache,
1042 PFNPDMBLKCACHEXFERCOMPLETEDRV pfnXferComplete,
1043 PFNPDMBLKCACHEXFERENQUEUEDRV pfnXferEnqueue,
1044 const char *pcszId)
1045{
1046 int rc = VINF_SUCCESS;
1047 PPDMBLKCACHE pBlkCache;
1048
1049 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1050 if (RT_SUCCESS(rc))
1051 {
1052 pBlkCache->enmType = PDMBLKCACHETYPE_DRV;
1053 pBlkCache->u.Drv.pfnXferComplete = pfnXferComplete;
1054 pBlkCache->u.Drv.pfnXferEnqueue = pfnXferEnqueue;
1055 pBlkCache->u.Drv.pDrvIns = pDrvIns;
1056 *ppBlkCache = pBlkCache;
1057 }
1058
1059 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1060 return rc;
1061}
1062
1063VMMR3DECL(int) PDMR3BlkCacheRetainDevice(PVM pVM, PPDMDEVINS pDevIns, PPPDMBLKCACHE ppBlkCache,
1064 PFNPDMBLKCACHEXFERCOMPLETEDEV pfnXferComplete,
1065 PFNPDMBLKCACHEXFERENQUEUEDEV pfnXferEnqueue,
1066 const char *pcszId)
1067{
1068 int rc = VINF_SUCCESS;
1069 PPDMBLKCACHE pBlkCache;
1070
1071 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1072 if (RT_SUCCESS(rc))
1073 {
1074 pBlkCache->enmType = PDMBLKCACHETYPE_DEV;
1075 pBlkCache->u.Dev.pfnXferComplete = pfnXferComplete;
1076 pBlkCache->u.Dev.pfnXferEnqueue = pfnXferEnqueue;
1077 pBlkCache->u.Dev.pDevIns = pDevIns;
1078 *ppBlkCache = pBlkCache;
1079 }
1080
1081 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1082 return rc;
1083
1084}
1085
1086VMMR3DECL(int) PDMR3BlkCacheRetainUsb(PVM pVM, PPDMUSBINS pUsbIns, PPPDMBLKCACHE ppBlkCache,
1087 PFNPDMBLKCACHEXFERCOMPLETEUSB pfnXferComplete,
1088 PFNPDMBLKCACHEXFERENQUEUEUSB pfnXferEnqueue,
1089 const char *pcszId)
1090{
1091 int rc = VINF_SUCCESS;
1092 PPDMBLKCACHE pBlkCache;
1093
1094 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1095 if (RT_SUCCESS(rc))
1096 {
1097 pBlkCache->enmType = PDMBLKCACHETYPE_USB;
1098 pBlkCache->u.Usb.pfnXferComplete = pfnXferComplete;
1099 pBlkCache->u.Usb.pfnXferEnqueue = pfnXferEnqueue;
1100 pBlkCache->u.Usb.pUsbIns = pUsbIns;
1101 *ppBlkCache = pBlkCache;
1102 }
1103
1104 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1105 return rc;
1106
1107}
1108
1109VMMR3DECL(int) PDMR3BlkCacheRetainInt(PVM pVM, void *pvUser, PPPDMBLKCACHE ppBlkCache,
1110 PFNPDMBLKCACHEXFERCOMPLETEINT pfnXferComplete,
1111 PFNPDMBLKCACHEXFERENQUEUEINT pfnXferEnqueue,
1112 const char *pcszId)
1113{
1114 int rc = VINF_SUCCESS;
1115 PPDMBLKCACHE pBlkCache;
1116
1117 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1118 if (RT_SUCCESS(rc))
1119 {
1120 pBlkCache->enmType = PDMBLKCACHETYPE_INTERNAL;
1121 pBlkCache->u.Int.pfnXferComplete = pfnXferComplete;
1122 pBlkCache->u.Int.pfnXferEnqueue = pfnXferEnqueue;
1123 pBlkCache->u.Int.pvUser = pvUser;
1124 *ppBlkCache = pBlkCache;
1125 }
1126
1127 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1128 return rc;
1129
1130}
1131
1132/**
1133 * Callback for the AVL destroy routine. Frees a cache entry for this endpoint.
1134 *
1135 * @returns IPRT status code.
1136 * @param pNode The node to destroy.
1137 * @param pvUser Opaque user data.
1138 */
1139static int pdmBlkCacheEntryDestroy(PAVLRU64NODECORE pNode, void *pvUser)
1140{
1141 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)pNode;
1142 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
1143 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
1144
1145 while (ASMAtomicReadU32(&pEntry->fFlags) & (PDMBLKCACHE_ENTRY_IO_IN_PROGRESS | PDMBLKCACHE_ENTRY_IS_DIRTY))
1146 {
1147 /* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
1148 pdmBlkCacheEntryRef(pEntry);
1149 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1150 pdmBlkCacheLockLeave(pCache);
1151
1152 RTThreadSleep(250);
1153
1154 /* Re-enter all locks */
1155 pdmBlkCacheLockEnter(pCache);
1156 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1157 pdmBlkCacheEntryRelease(pEntry);
1158 }
1159
1160 AssertMsg(!(pEntry->fFlags & (PDMBLKCACHE_ENTRY_IO_IN_PROGRESS | PDMBLKCACHE_ENTRY_IS_DIRTY)),
1161 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
1162
1163 bool fUpdateCache = pEntry->pList == &pCache->LruFrequentlyUsed
1164 || pEntry->pList == &pCache->LruRecentlyUsedIn;
1165
1166 pdmBlkCacheEntryRemoveFromList(pEntry);
1167
1168 if (fUpdateCache)
1169 pdmBlkCacheSub(pCache, pEntry->cbData);
1170
1171 RTMemPageFree(pEntry->pbData, pEntry->cbData);
1172 RTMemFree(pEntry);
1173
1174 return VINF_SUCCESS;
1175}
1176
1177/**
1178 * Destroys all cache resources used by the given endpoint.
1179 *
1180 * @returns nothing.
1181 * @param pEndpoint The endpoint to the destroy.
1182 */
1183VMMR3DECL(void) PDMR3BlkCacheRelease(PPDMBLKCACHE pBlkCache)
1184{
1185 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1186
1187 /* Make sure nobody is accessing the cache while we delete the tree. */
1188 pdmBlkCacheLockEnter(pCache);
1189 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1190 RTAvlrU64Destroy(pBlkCache->pTree, pdmBlkCacheEntryDestroy, pCache);
1191 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1192
1193 RTSpinlockDestroy(pBlkCache->LockList);
1194
1195 pCache->cRefs--;
1196 RTListNodeRemove(&pBlkCache->NodeCacheUser);
1197
1198 if ( !pCache->cRefs
1199 && pCache->u32CommitTimeoutMs > 0)
1200 TMTimerStop(pCache->pTimerCommit);
1201
1202 pdmBlkCacheLockLeave(pCache);
1203
1204 RTSemRWDestroy(pBlkCache->SemRWEntries);
1205
1206#ifdef VBOX_WITH_STATISTICS
1207 STAMR3Deregister(pCache->pVM, &pBlkCache->StatWriteDeferred);
1208#endif
1209
1210 RTStrFree(pBlkCache->pszId);
1211 RTMemFree(pBlkCache);
1212}
1213
1214VMMR3DECL(void) PDMR3BlkCacheReleaseDevice(PVM pVM, PPDMDEVINS pDevIns)
1215{
1216 LogFlow(("%s: pDevIns=%p\n", __FUNCTION__, pDevIns));
1217
1218 /*
1219 * Validate input.
1220 */
1221 if (!pDevIns)
1222 return;
1223 VM_ASSERT_EMT(pVM);
1224
1225 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1226 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1227
1228 /* Return silently if not supported. */
1229 if (!pBlkCacheGlobal)
1230 return;
1231
1232 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1233
1234 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1235 {
1236 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DEV
1237 && pBlkCache->u.Dev.pDevIns == pDevIns)
1238 PDMR3BlkCacheRelease(pBlkCache);
1239 }
1240
1241 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1242}
1243
1244VMMR3DECL(void) PDMR3BlkCacheReleaseDriver(PVM pVM, PPDMDRVINS pDrvIns)
1245{
1246 LogFlow(("%s: pDrvIns=%p\n", __FUNCTION__, pDrvIns));
1247
1248 /*
1249 * Validate input.
1250 */
1251 if (!pDrvIns)
1252 return;
1253 VM_ASSERT_EMT(pVM);
1254
1255 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1256 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1257
1258 /* Return silently if not supported. */
1259 if (!pBlkCacheGlobal)
1260 return;
1261
1262 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1263
1264 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1265 {
1266 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DRV
1267 && pBlkCache->u.Drv.pDrvIns == pDrvIns)
1268 PDMR3BlkCacheRelease(pBlkCache);
1269 }
1270
1271 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1272}
1273
1274VMMR3DECL(void) PDMR3BlkCacheReleaseUsb(PVM pVM, PPDMUSBINS pUsbIns)
1275{
1276 LogFlow(("%s: pUsbIns=%p\n", __FUNCTION__, pUsbIns));
1277
1278 /*
1279 * Validate input.
1280 */
1281 if (!pUsbIns)
1282 return;
1283 VM_ASSERT_EMT(pVM);
1284
1285 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1286 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1287
1288 /* Return silently if not supported. */
1289 if (!pBlkCacheGlobal)
1290 return;
1291
1292 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1293
1294 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1295 {
1296 if ( pBlkCache->enmType == PDMBLKCACHETYPE_USB
1297 && pBlkCache->u.Usb.pUsbIns == pUsbIns)
1298 PDMR3BlkCacheRelease(pBlkCache);
1299 }
1300
1301 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1302}
1303
1304static PPDMBLKCACHEENTRY pdmBlkCacheGetCacheEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off)
1305{
1306 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1307 PPDMBLKCACHEENTRY pEntry = NULL;
1308
1309 STAM_PROFILE_ADV_START(&pCache->StatTreeGet, Cache);
1310
1311 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1312 pEntry = (PPDMBLKCACHEENTRY)RTAvlrU64RangeGet(pBlkCache->pTree, off);
1313 if (pEntry)
1314 pdmBlkCacheEntryRef(pEntry);
1315 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1316
1317 STAM_PROFILE_ADV_STOP(&pCache->StatTreeGet, Cache);
1318
1319 return pEntry;
1320}
1321
1322/**
1323 * Return the best fit cache entries for the given offset.
1324 *
1325 * @returns nothing.
1326 * @param pBlkCache The endpoint cache.
1327 * @param off The offset.
1328 * @param pEntryAbove Where to store the pointer to the best fit entry above the
1329 * the given offset. NULL if not required.
1330 * @param pEntryBelow Where to store the pointer to the best fit entry below the
1331 * the given offset. NULL if not required.
1332 */
1333static void pdmBlkCacheGetCacheBestFitEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off,
1334 PPDMBLKCACHEENTRY *ppEntryAbove,
1335 PPDMBLKCACHEENTRY *ppEntryBelow)
1336{
1337 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1338
1339 STAM_PROFILE_ADV_START(&pCache->StatTreeGet, Cache);
1340
1341 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1342 if (ppEntryAbove)
1343 {
1344 *ppEntryAbove = (PPDMBLKCACHEENTRY)RTAvlrU64GetBestFit(pBlkCache->pTree, off, true /*fAbove*/);
1345 if (*ppEntryAbove)
1346 pdmBlkCacheEntryRef(*ppEntryAbove);
1347 }
1348
1349 if (ppEntryBelow)
1350 {
1351 *ppEntryBelow = (PPDMBLKCACHEENTRY)RTAvlrU64GetBestFit(pBlkCache->pTree, off, false /*fAbove*/);
1352 if (*ppEntryBelow)
1353 pdmBlkCacheEntryRef(*ppEntryBelow);
1354 }
1355 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1356
1357 STAM_PROFILE_ADV_STOP(&pCache->StatTreeGet, Cache);
1358}
1359
1360static void pdmBlkCacheInsertEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
1361{
1362 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1363
1364 STAM_PROFILE_ADV_START(&pCache->StatTreeInsert, Cache);
1365 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1366 bool fInserted = RTAvlrU64Insert(pBlkCache->pTree, &pEntry->Core);
1367 AssertMsg(fInserted, ("Node was not inserted into tree\n"));
1368 STAM_PROFILE_ADV_STOP(&pCache->StatTreeInsert, Cache);
1369 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1370}
1371
1372/**
1373 * Allocates and initializes a new entry for the cache.
1374 * The entry has a reference count of 1.
1375 *
1376 * @returns Pointer to the new cache entry or NULL if out of memory.
1377 * @param pBlkCache The cache the entry belongs to.
1378 * @param off Start offset.
1379 * @param cbData Size of the cache entry.
1380 * @param pbBuffer Pointer to the buffer to use.
1381 * NULL if a new buffer should be allocated.
1382 * The buffer needs to have the same size of the entry.
1383 */
1384static PPDMBLKCACHEENTRY pdmBlkCacheEntryAlloc(PPDMBLKCACHE pBlkCache,
1385 uint64_t off, size_t cbData, uint8_t *pbBuffer)
1386{
1387 PPDMBLKCACHEENTRY pEntryNew = (PPDMBLKCACHEENTRY)RTMemAllocZ(sizeof(PDMBLKCACHEENTRY));
1388
1389 if (RT_UNLIKELY(!pEntryNew))
1390 return NULL;
1391
1392 pEntryNew->Core.Key = off;
1393 pEntryNew->Core.KeyLast = off + cbData - 1;
1394 pEntryNew->pBlkCache = pBlkCache;
1395 pEntryNew->fFlags = 0;
1396 pEntryNew->cRefs = 1; /* We are using it now. */
1397 pEntryNew->pList = NULL;
1398 pEntryNew->cbData = cbData;
1399 pEntryNew->pWaitingHead = NULL;
1400 pEntryNew->pWaitingTail = NULL;
1401 if (pbBuffer)
1402 pEntryNew->pbData = pbBuffer;
1403 else
1404 pEntryNew->pbData = (uint8_t *)RTMemPageAlloc(cbData);
1405
1406 if (RT_UNLIKELY(!pEntryNew->pbData))
1407 {
1408 RTMemFree(pEntryNew);
1409 return NULL;
1410 }
1411
1412 return pEntryNew;
1413}
1414
1415/**
1416 * Checks that a set of flags is set/clear acquiring the R/W semaphore
1417 * in exclusive mode.
1418 *
1419 * @returns true if the flag in fSet is set and the one in fClear is clear.
1420 * false otherwise.
1421 * The R/W semaphore is only held if true is returned.
1422 *
1423 * @param pBlkCache The endpoint cache instance data.
1424 * @param pEntry The entry to check the flags for.
1425 * @param fSet The flag which is tested to be set.
1426 * @param fClear The flag which is tested to be clear.
1427 */
1428DECLINLINE(bool) pdmBlkCacheEntryFlagIsSetClearAcquireLock(PPDMBLKCACHE pBlkCache,
1429 PPDMBLKCACHEENTRY pEntry,
1430 uint32_t fSet, uint32_t fClear)
1431{
1432 uint32_t fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1433 bool fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1434
1435 if (fPassed)
1436 {
1437 /* Acquire the lock and check again because the completion callback might have raced us. */
1438 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1439
1440 fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1441 fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1442
1443 /* Drop the lock if we didn't passed the test. */
1444 if (!fPassed)
1445 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1446 }
1447
1448 return fPassed;
1449}
1450
1451/**
1452 * Adds a segment to the waiting list for a cache entry
1453 * which is currently in progress.
1454 *
1455 * @returns nothing.
1456 * @param pEntry The cache entry to add the segment to.
1457 * @param pSeg The segment to add.
1458 */
1459DECLINLINE(void) pdmBlkCacheEntryAddWaiter(PPDMBLKCACHEENTRY pEntry,
1460 PPDMBLKCACHEWAITER pWaiter)
1461{
1462 pWaiter->pNext = NULL;
1463
1464 if (pEntry->pWaitingHead)
1465 {
1466 AssertPtr(pEntry->pWaitingTail);
1467
1468 pEntry->pWaitingTail->pNext = pWaiter;
1469 pEntry->pWaitingTail = pWaiter;
1470 }
1471 else
1472 {
1473 Assert(!pEntry->pWaitingTail);
1474
1475 pEntry->pWaitingHead = pWaiter;
1476 pEntry->pWaitingTail = pWaiter;
1477 }
1478}
1479
1480/**
1481 * Add a buffer described by the I/O memory context
1482 * to the entry waiting for completion.
1483 *
1484 * @returns VBox status code.
1485 * @param pEntry The entry to add the buffer to.
1486 * @param pTask Task associated with the buffer.
1487 * @param pIoMemCtx The memory context to use.
1488 * @param offDiff Offset from the start of the buffer
1489 * in the entry.
1490 * @param cbData Amount of data to wait for onthis entry.
1491 * @param fWrite Flag whether the task waits because it wants to write
1492 * to the cache entry.
1493 */
1494static int pdmBlkCacheEntryWaitersAdd(PPDMBLKCACHEENTRY pEntry,
1495 PPDMBLKCACHEREQ pReq,
1496 PCRTSGBUF pSgBuf, uint64_t offDiff,
1497 size_t cbData, bool fWrite)
1498{
1499 PPDMBLKCACHEWAITER pWaiter = (PPDMBLKCACHEWAITER)RTMemAllocZ(sizeof(PDMBLKCACHEWAITER));
1500 if (!pWaiter)
1501 return VERR_NO_MEMORY;
1502
1503 ASMAtomicIncU32(&pReq->cXfersPending);
1504 pWaiter->pReq = pReq;
1505 pWaiter->offCacheEntry = offDiff;
1506 pWaiter->cbTransfer = cbData;
1507 pWaiter->fWrite = fWrite;
1508 RTSgBufClone(&pWaiter->SgBuf, pSgBuf);
1509
1510 pdmBlkCacheEntryAddWaiter(pEntry, pWaiter);
1511
1512 return VINF_SUCCESS;
1513}
1514
1515/**
1516 * Calculate aligned offset and size for a new cache entry
1517 * which do not intersect with an already existing entry and the
1518 * file end.
1519 *
1520 * @returns The number of bytes the entry can hold of the requested amount
1521 * of byte.
1522 * @param pEndpoint The endpoint.
1523 * @param pBlkCache The endpoint cache.
1524 * @param off The start offset.
1525 * @param cb The number of bytes the entry needs to hold at least.
1526 * @param uAlignment Alignment of the boundary sizes.
1527 * @param poffAligned Where to store the aligned offset.
1528 * @param pcbAligned Where to store the aligned size of the entry.
1529 */
1530static size_t pdmBlkCacheEntryBoundariesCalc(PPDMBLKCACHE pBlkCache,
1531 uint64_t off, size_t cb,
1532 unsigned uAlignment,
1533 uint64_t *poffAligned, size_t *pcbAligned)
1534{
1535 size_t cbAligned;
1536 size_t cbInEntry = 0;
1537 uint64_t offAligned;
1538 PPDMBLKCACHEENTRY pEntryAbove = NULL;
1539 PPDMBLKCACHEENTRY pEntryBelow = NULL;
1540
1541 /* Get the best fit entries around the offset */
1542 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off, &pEntryAbove, &pEntryBelow);
1543
1544 /* Log the info */
1545 LogFlow(("%sest fit entry below off=%llu (BestFit=%llu BestFitEnd=%llu BestFitSize=%u)\n",
1546 pEntryBelow ? "B" : "No b",
1547 off,
1548 pEntryBelow ? pEntryBelow->Core.Key : 0,
1549 pEntryBelow ? pEntryBelow->Core.KeyLast : 0,
1550 pEntryBelow ? pEntryBelow->cbData : 0));
1551
1552 LogFlow(("%sest fit entry above off=%llu (BestFit=%llu BestFitEnd=%llu BestFitSize=%u)\n",
1553 pEntryAbove ? "B" : "No b",
1554 off,
1555 pEntryAbove ? pEntryAbove->Core.Key : 0,
1556 pEntryAbove ? pEntryAbove->Core.KeyLast : 0,
1557 pEntryAbove ? pEntryAbove->cbData : 0));
1558
1559 /* Align the offset first. */
1560 offAligned = off & ~(uint64_t)(512-1);
1561 if ( pEntryBelow
1562 && offAligned <= pEntryBelow->Core.KeyLast)
1563 offAligned = pEntryBelow->Core.KeyLast;
1564
1565 if ( pEntryAbove
1566 && off + cb > pEntryAbove->Core.Key)
1567 {
1568 cbInEntry = pEntryAbove->Core.Key - off;
1569 cbAligned = pEntryAbove->Core.Key - offAligned;
1570 }
1571 else
1572 {
1573 cbAligned = cb;
1574 cbInEntry = cb;
1575#if 0
1576 /*
1577 * Align the size to a 4KB boundary.
1578 * Memory size is aligned to a page boundary
1579 * and memory is wasted if the size is rather small.
1580 * (For example reads with a size of 512 bytes).
1581 */
1582 cbInEntry = cb;
1583 cbAligned = RT_ALIGN_Z(cb + (off - offAligned), uAlignment);
1584
1585 /*
1586 * Clip to file size if the original request doesn't
1587 * exceed the file (not an appending write)
1588 */
1589 uint64_t cbReq = off + cb;
1590 if (cbReq >= pEndpoint->cbFile)
1591 cbAligned = cbReq - offAligned;
1592 else
1593 cbAligned = RT_MIN(pEndpoint->cbFile - offAligned, cbAligned);
1594 if (pEntryAbove)
1595 {
1596 Assert(pEntryAbove->Core.Key >= off);
1597 cbAligned = RT_MIN(cbAligned, pEntryAbove->Core.Key - offAligned);
1598 }
1599#endif
1600 }
1601
1602 /* A few sanity checks */
1603 AssertMsg(!pEntryBelow || pEntryBelow->Core.KeyLast < offAligned,
1604 ("Aligned start offset intersects with another cache entry\n"));
1605 AssertMsg(!pEntryAbove || (offAligned + cbAligned) <= pEntryAbove->Core.Key,
1606 ("Aligned size intersects with another cache entry\n"));
1607 Assert(cbInEntry <= cbAligned);
1608
1609 if (pEntryBelow)
1610 pdmBlkCacheEntryRelease(pEntryBelow);
1611 if (pEntryAbove)
1612 pdmBlkCacheEntryRelease(pEntryAbove);
1613
1614 LogFlow(("offAligned=%llu cbAligned=%u\n", offAligned, cbAligned));
1615
1616 *poffAligned = offAligned;
1617 *pcbAligned = cbAligned;
1618
1619 return cbInEntry;
1620}
1621
1622/**
1623 * Create a new cache entry evicting data from the cache if required.
1624 *
1625 * @returns Pointer to the new cache entry or NULL
1626 * if not enough bytes could be evicted from the cache.
1627 * @param pEndpoint The endpoint.
1628 * @param pBlkCache The endpoint cache.
1629 * @param off The offset.
1630 * @param cb Number of bytes the cache entry should have.
1631 * @param uAlignment Alignment the size of the entry should have.
1632 * @param pcbData Where to store the number of bytes the new
1633 * entry can hold. May be lower than actually requested
1634 * due to another entry intersecting the access range.
1635 */
1636static PPDMBLKCACHEENTRY pdmBlkCacheEntryCreate(PPDMBLKCACHE pBlkCache,
1637 uint64_t off, size_t cb,
1638 unsigned uAlignment,
1639 size_t *pcbData)
1640{
1641 uint64_t offStart = 0;
1642 size_t cbEntry = 0;
1643 PPDMBLKCACHEENTRY pEntryNew = NULL;
1644 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1645 uint8_t *pbBuffer = NULL;
1646
1647 *pcbData = pdmBlkCacheEntryBoundariesCalc(pBlkCache, off, cb, uAlignment,
1648 &offStart, &cbEntry);
1649
1650 pdmBlkCacheLockEnter(pCache);
1651 bool fEnough = pdmBlkCacheReclaim(pCache, cbEntry, true, &pbBuffer);
1652
1653 if (fEnough)
1654 {
1655 LogFlow(("Evicted enough bytes (%u requested). Creating new cache entry\n", cbEntry));
1656
1657 pEntryNew = pdmBlkCacheEntryAlloc(pBlkCache, offStart, cbEntry, pbBuffer);
1658 if (RT_LIKELY(pEntryNew))
1659 {
1660 pdmBlkCacheEntryAddToList(&pCache->LruRecentlyUsedIn, pEntryNew);
1661 pdmBlkCacheAdd(pCache, cbEntry);
1662 pdmBlkCacheLockLeave(pCache);
1663
1664 pdmBlkCacheInsertEntry(pBlkCache, pEntryNew);
1665
1666 AssertMsg( (off >= pEntryNew->Core.Key)
1667 && (off + *pcbData <= pEntryNew->Core.KeyLast + 1),
1668 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
1669 off, pEntryNew->Core.Key));
1670 }
1671 else
1672 pdmBlkCacheLockLeave(pCache);
1673 }
1674 else
1675 pdmBlkCacheLockLeave(pCache);
1676
1677 return pEntryNew;
1678}
1679
1680static PPDMBLKCACHEREQ pdmBlkCacheReqAlloc(size_t cbXfer, void *pvUser)
1681{
1682 PPDMBLKCACHEREQ pReq = (PPDMBLKCACHEREQ)RTMemAlloc(sizeof(PDMBLKCACHEREQ));
1683
1684 if (RT_LIKELY(pReq))
1685 {
1686 pReq->pvUser = pvUser;
1687 pReq->cbXfer = cbXfer;
1688 pReq->rcReq = VINF_SUCCESS;
1689 pReq->cXfersPending = 0;
1690 }
1691
1692 return pReq;
1693}
1694
1695static void pdmBlkCacheReqComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq)
1696{
1697 switch (pBlkCache->enmType)
1698 {
1699 case PDMBLKCACHETYPE_DEV:
1700 {
1701 pBlkCache->u.Dev.pfnXferComplete(pBlkCache->u.Dev.pDevIns,
1702 pReq->pvUser, pReq->rcReq);
1703 break;
1704 }
1705 case PDMBLKCACHETYPE_DRV:
1706 {
1707 pBlkCache->u.Drv.pfnXferComplete(pBlkCache->u.Drv.pDrvIns,
1708 pReq->pvUser, pReq->rcReq);
1709 break;
1710 }
1711 case PDMBLKCACHETYPE_USB:
1712 {
1713 pBlkCache->u.Usb.pfnXferComplete(pBlkCache->u.Usb.pUsbIns,
1714 pReq->pvUser, pReq->rcReq);
1715 break;
1716 }
1717 case PDMBLKCACHETYPE_INTERNAL:
1718 {
1719 pBlkCache->u.Int.pfnXferComplete(pBlkCache->u.Int.pvUser,
1720 pReq->pvUser, pReq->rcReq);
1721 break;
1722 }
1723 default:
1724 AssertMsgFailed(("Unknown block cache type!\n"));
1725 }
1726
1727 RTMemFree(pReq);
1728}
1729
1730static bool pdmBlkCacheReqUpdate(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq,
1731 size_t cbComplete, int rcReq, bool fCallHandler)
1732{
1733 if (RT_FAILURE(rcReq))
1734 ASMAtomicCmpXchgS32(&pReq->rcReq, rcReq, VINF_SUCCESS);
1735
1736 AssertMsg(pReq->cbXfer >= cbComplete, ("Completed more than left\n"));
1737 AssertMsg(pReq->cXfersPending > 0, ("No transfers are pending for this request\n"));
1738 uint32_t cbOld = ASMAtomicSubU32(&pReq->cbXfer, cbComplete);
1739 uint32_t cXfersPending = ASMAtomicDecU32(&pReq->cXfersPending);
1740
1741 if ( !(cbOld - cbComplete)
1742 && !cXfersPending)
1743 {
1744 if (fCallHandler)
1745 pdmBlkCacheReqComplete(pBlkCache, pReq);
1746 return true;
1747 }
1748
1749 LogFlowFunc(("pReq=%#p cXfersPending=%u cbXfer=%u\n", pReq, cXfersPending, (cbOld - cbComplete)));
1750 return false;
1751}
1752
1753VMMR3DECL(int) PDMR3BlkCacheRead(PPDMBLKCACHE pBlkCache, uint64_t off,
1754 PCRTSGBUF pcSgBuf, size_t cbRead, void *pvUser)
1755{
1756 int rc = VINF_SUCCESS;
1757 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1758 PPDMBLKCACHEENTRY pEntry;
1759 PPDMBLKCACHEREQ pReq;
1760
1761 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pcSgBuf=%#p cbRead=%u pvUser=%#p\n",
1762 pBlkCache, pBlkCache->pszId, off, pcSgBuf, cbRead, pvUser));
1763
1764 RTSGBUF SgBuf;
1765 RTSgBufClone(&SgBuf, pcSgBuf);
1766
1767 /* Allocate new request structure. */
1768 pReq = pdmBlkCacheReqAlloc(cbRead, pvUser);
1769 if (RT_UNLIKELY(!pReq))
1770 return VERR_NO_MEMORY;
1771
1772 /* Increment data transfer counter to keep the request valid while we access it. */
1773 ASMAtomicIncU32(&pReq->cXfersPending);
1774
1775 while (cbRead)
1776 {
1777 size_t cbToRead;
1778
1779 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
1780
1781 /*
1782 * If there is no entry we try to create a new one eviciting unused pages
1783 * if the cache is full. If this is not possible we will pass the request through
1784 * and skip the caching (all entries may be still in progress so they can't
1785 * be evicted)
1786 * If we have an entry it can be in one of the LRU lists where the entry
1787 * contains data (recently used or frequently used LRU) so we can just read
1788 * the data we need and put the entry at the head of the frequently used LRU list.
1789 * In case the entry is in one of the ghost lists it doesn't contain any data.
1790 * We have to fetch it again evicting pages from either T1 or T2 to make room.
1791 */
1792 if (pEntry)
1793 {
1794 uint64_t offDiff = off - pEntry->Core.Key;
1795
1796 AssertMsg(off >= pEntry->Core.Key,
1797 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
1798 off, pEntry->Core.Key));
1799
1800 AssertPtr(pEntry->pList);
1801
1802 cbToRead = RT_MIN(pEntry->cbData - offDiff, cbRead);
1803
1804 AssertMsg(off + cbToRead <= pEntry->Core.Key + pEntry->Core.KeyLast + 1,
1805 ("Buffer of cache entry exceeded off=%llu cbToRead=%d\n",
1806 off, cbToRead));
1807
1808 cbRead -= cbToRead;
1809
1810 if (!cbRead)
1811 STAM_COUNTER_INC(&pCache->cHits);
1812 else
1813 STAM_COUNTER_INC(&pCache->cPartialHits);
1814
1815 STAM_COUNTER_ADD(&pCache->StatRead, cbToRead);
1816
1817 /* Ghost lists contain no data. */
1818 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
1819 || (pEntry->pList == &pCache->LruFrequentlyUsed))
1820 {
1821 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
1822 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
1823 PDMBLKCACHE_ENTRY_IS_DIRTY))
1824 {
1825 /* Entry didn't completed yet. Append to the list */
1826 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
1827 &SgBuf, offDiff, cbToRead,
1828 false /* fWrite */);
1829 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1830 }
1831 else
1832 {
1833 /* Read as much as we can from the entry. */
1834 RTSgBufCopyFromBuf(&SgBuf, pEntry->pbData + offDiff, cbToRead);
1835 ASMAtomicSubU32(&pReq->cbXfer, cbToRead);
1836 }
1837
1838 /* Move this entry to the top position */
1839 if (pEntry->pList == &pCache->LruFrequentlyUsed)
1840 {
1841 pdmBlkCacheLockEnter(pCache);
1842 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
1843 pdmBlkCacheLockLeave(pCache);
1844 }
1845 /* Release the entry */
1846 pdmBlkCacheEntryRelease(pEntry);
1847 }
1848 else
1849 {
1850 uint8_t *pbBuffer = NULL;
1851
1852 LogFlow(("Fetching data for ghost entry %#p from file\n", pEntry));
1853
1854 pdmBlkCacheLockEnter(pCache);
1855 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
1856 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
1857
1858 /* Move the entry to Am and fetch it to the cache. */
1859 if (fEnough)
1860 {
1861 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
1862 pdmBlkCacheAdd(pCache, pEntry->cbData);
1863 pdmBlkCacheLockLeave(pCache);
1864
1865 if (pbBuffer)
1866 pEntry->pbData = pbBuffer;
1867 else
1868 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
1869 AssertPtr(pEntry->pbData);
1870
1871 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
1872 &SgBuf, offDiff, cbToRead,
1873 false /* fWrite */);
1874 pdmBlkCacheEntryReadFromMedium(pEntry);
1875 /* Release the entry */
1876 pdmBlkCacheEntryRelease(pEntry);
1877 }
1878 else
1879 {
1880 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1881 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
1882 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
1883 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
1884 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1885
1886 pdmBlkCacheLockLeave(pCache);
1887
1888 RTMemFree(pEntry);
1889
1890 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
1891 &SgBuf, off, cbToRead,
1892 PDMBLKCACHEXFERDIR_READ);
1893 }
1894 }
1895 }
1896 else
1897 {
1898#ifdef VBOX_WITH_IO_READ_CACHE
1899 /* No entry found for this offset. Create a new entry and fetch the data to the cache. */
1900 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
1901 off, cbRead,
1902 PAGE_SIZE,
1903 &cbToRead);
1904
1905 cbRead -= cbToRead;
1906
1907 if (pEntryNew)
1908 {
1909 if (!cbRead)
1910 STAM_COUNTER_INC(&pCache->cMisses);
1911 else
1912 STAM_COUNTER_INC(&pCache->cPartialHits);
1913
1914 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
1915 &SgBuf,
1916 off - pEntryNew->Core.Key,
1917 cbToRead,
1918 false /* fWrite */);
1919 pdmBlkCacheEntryReadFromMedium(pEntryNew);
1920 pdmBlkCacheEntryRelease(pEntryNew); /* it is protected by the I/O in progress flag now. */
1921 }
1922 else
1923 {
1924 /*
1925 * There is not enough free space in the cache.
1926 * Pass the request directly to the I/O manager.
1927 */
1928 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToRead));
1929
1930 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
1931 &SgBuf, off, cbToRead,
1932 PDMBLKCACHEXFERDIR_READ);
1933 }
1934#else
1935 /* Clip read size if necessary. */
1936 PPDMBLKCACHEENTRY pEntryAbove;
1937 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off,
1938 &pEntryAbove, NULL);
1939
1940 if (pEntryAbove)
1941 {
1942 if (off + cbRead > pEntryAbove->Core.Key)
1943 cbToRead = pEntryAbove->Core.Key - off;
1944 else
1945 cbToRead = cbRead;
1946
1947 pdmBlkCacheEntryRelease(pEntryAbove);
1948 }
1949 else
1950 cbToRead = cbRead;
1951
1952 cbRead -= cbToRead;
1953 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
1954 &SgBuf, off, cbToRead,
1955 PDMBLKCACHEXFERDIR_READ);
1956#endif
1957 }
1958 off += cbToRead;
1959 }
1960
1961 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, 0, rc, false))
1962 rc = VINF_AIO_TASK_PENDING;
1963
1964 LogFlowFunc((": Leave rc=%Rrc\n", rc));
1965
1966 return rc;
1967}
1968
1969VMMR3DECL(int) PDMR3BlkCacheWrite(PPDMBLKCACHE pBlkCache, uint64_t off,
1970 PCRTSGBUF pcSgBuf, size_t cbWrite, void *pvUser)
1971{
1972 int rc = VINF_SUCCESS;
1973 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1974 PPDMBLKCACHEENTRY pEntry;
1975 PPDMBLKCACHEREQ pReq;
1976
1977 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pcSgBuf=%#p cbWrite=%u pvUser=%#p\n",
1978 pBlkCache, pBlkCache->pszId, off, pcSgBuf, cbWrite, pvUser));
1979
1980 RTSGBUF SgBuf;
1981 RTSgBufClone(&SgBuf, pcSgBuf);
1982
1983 /* Allocate new request structure. */
1984 pReq = pdmBlkCacheReqAlloc(cbWrite, pvUser);
1985 if (RT_UNLIKELY(!pReq))
1986 return VERR_NO_MEMORY;
1987
1988 /* Increment data transfer counter to keep the request valid while we access it. */
1989 ASMAtomicIncU32(&pReq->cXfersPending);
1990
1991 while (cbWrite)
1992 {
1993 size_t cbToWrite;
1994
1995 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
1996
1997 if (pEntry)
1998 {
1999 /* Write the data into the entry and mark it as dirty */
2000 AssertPtr(pEntry->pList);
2001
2002 uint64_t offDiff = off - pEntry->Core.Key;
2003
2004 AssertMsg(off >= pEntry->Core.Key,
2005 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
2006 off, pEntry->Core.Key));
2007
2008 cbToWrite = RT_MIN(pEntry->cbData - offDiff, cbWrite);
2009 cbWrite -= cbToWrite;
2010
2011 if (!cbWrite)
2012 STAM_COUNTER_INC(&pCache->cHits);
2013 else
2014 STAM_COUNTER_INC(&pCache->cPartialHits);
2015
2016 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2017
2018 /* Ghost lists contain no data. */
2019 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2020 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2021 {
2022 /* Check if the entry is dirty. */
2023 if(pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2024 PDMBLKCACHE_ENTRY_IS_DIRTY,
2025 0))
2026 {
2027 /* If it is already dirty but not in progress just update the data. */
2028 if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS))
2029 {
2030 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff,
2031 cbToWrite);
2032 ASMAtomicSubU32(&pReq->cbXfer, cbToWrite);
2033 }
2034 else
2035 {
2036 /* The data isn't written to the file yet */
2037 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2038 &SgBuf, offDiff, cbToWrite,
2039 true /* fWrite */);
2040 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2041 }
2042
2043 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2044 }
2045 else /* Dirty bit not set */
2046 {
2047 /*
2048 * Check if a read is in progress for this entry.
2049 * We have to defer processing in that case.
2050 */
2051 if(pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2052 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2053 0))
2054 {
2055 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2056 &SgBuf, offDiff, cbToWrite,
2057 true /* fWrite */);
2058 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2059 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2060 }
2061 else /* I/O in progress flag not set */
2062 {
2063 /* Write as much as we can into the entry and update the file. */
2064 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff, cbToWrite);
2065 ASMAtomicSubU32(&pReq->cbXfer, cbToWrite);
2066
2067 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2068 if (fCommit)
2069 pdmBlkCacheCommitDirtyEntries(pCache);
2070 }
2071 } /* Dirty bit not set */
2072
2073 /* Move this entry to the top position */
2074 if (pEntry->pList == &pCache->LruFrequentlyUsed)
2075 {
2076 pdmBlkCacheLockEnter(pCache);
2077 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2078 pdmBlkCacheLockLeave(pCache);
2079 }
2080
2081 pdmBlkCacheEntryRelease(pEntry);
2082 }
2083 else /* Entry is on the ghost list */
2084 {
2085 uint8_t *pbBuffer = NULL;
2086
2087 pdmBlkCacheLockEnter(pCache);
2088 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2089 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2090
2091 if (fEnough)
2092 {
2093 /* Move the entry to Am and fetch it to the cache. */
2094 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2095 pdmBlkCacheAdd(pCache, pEntry->cbData);
2096 pdmBlkCacheLockLeave(pCache);
2097
2098 if (pbBuffer)
2099 pEntry->pbData = pbBuffer;
2100 else
2101 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2102 AssertPtr(pEntry->pbData);
2103
2104 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2105 &SgBuf, offDiff, cbToWrite,
2106 true /* fWrite */);
2107 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2108 pdmBlkCacheEntryReadFromMedium(pEntry);
2109
2110 /* Release the reference. If it is still needed the I/O in progress flag should protect it now. */
2111 pdmBlkCacheEntryRelease(pEntry);
2112 }
2113 else
2114 {
2115 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2116 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2117 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2118 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2119 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2120
2121 pdmBlkCacheLockLeave(pCache);
2122
2123 RTMemFree(pEntry);
2124 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2125 &SgBuf, off, cbToWrite,
2126 PDMBLKCACHEXFERDIR_WRITE);
2127 }
2128 }
2129 }
2130 else /* No entry found */
2131 {
2132 /*
2133 * No entry found. Try to create a new cache entry to store the data in and if that fails
2134 * write directly to the file.
2135 */
2136 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
2137 off, cbWrite,
2138 512, &cbToWrite);
2139
2140 cbWrite -= cbToWrite;
2141
2142 if (pEntryNew)
2143 {
2144 uint64_t offDiff = off - pEntryNew->Core.Key;
2145
2146 STAM_COUNTER_INC(&pCache->cHits);
2147
2148 /*
2149 * Check if it is possible to just write the data without waiting
2150 * for it to get fetched first.
2151 */
2152 if (!offDiff && pEntryNew->cbData == cbToWrite)
2153 {
2154 RTSgBufCopyToBuf(&SgBuf, pEntryNew->pbData, cbToWrite);
2155 ASMAtomicSubU32(&pReq->cbXfer, cbToWrite);
2156
2157 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntryNew);
2158 if (fCommit)
2159 pdmBlkCacheCommitDirtyEntries(pCache);
2160 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2161 }
2162 else
2163 {
2164 /* Defer the write and fetch the data from the endpoint. */
2165 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
2166 &SgBuf, offDiff, cbToWrite,
2167 true /* fWrite */);
2168 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2169 pdmBlkCacheEntryReadFromMedium(pEntryNew);
2170 }
2171
2172 pdmBlkCacheEntryRelease(pEntryNew);
2173 }
2174 else
2175 {
2176 /*
2177 * There is not enough free space in the cache.
2178 * Pass the request directly to the I/O manager.
2179 */
2180 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToWrite));
2181
2182 STAM_COUNTER_INC(&pCache->cMisses);
2183
2184 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2185 &SgBuf, off, cbToWrite,
2186 PDMBLKCACHEXFERDIR_WRITE);
2187 }
2188 }
2189
2190 off += cbToWrite;
2191 }
2192
2193 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, 0, rc, false))
2194 rc = VINF_AIO_TASK_PENDING;
2195
2196 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2197
2198 return rc;
2199}
2200
2201VMMR3DECL(int) PDMR3BlkCacheFlush(PPDMBLKCACHE pBlkCache, void *pvUser)
2202{
2203 int rc = VINF_SUCCESS;
2204
2205 LogFlowFunc((": pBlkCache=%#p{%s}\n", pBlkCache, pBlkCache->pszId));
2206
2207 /* Commit dirty entries in the cache. */
2208 pdmBlkCacheCommit(pBlkCache);
2209
2210 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2211 return rc;
2212}
2213
2214/**
2215 * Completes a task segment freeing all resources and completes the task handle
2216 * if everything was transferred.
2217 *
2218 * @returns Next task segment handle.
2219 * @param pTaskSeg Task segment to complete.
2220 * @param rc Status code to set.
2221 */
2222static PPDMBLKCACHEWAITER pdmBlkCacheWaiterComplete(PPDMBLKCACHE pBlkCache,
2223 PPDMBLKCACHEWAITER pWaiter,
2224 int rc)
2225{
2226 PPDMBLKCACHEWAITER pNext = pWaiter->pNext;
2227 PPDMBLKCACHEREQ pReq = pWaiter->pReq;
2228
2229 pdmBlkCacheReqUpdate(pBlkCache, pWaiter->pReq, pWaiter->cbTransfer, rc, true);
2230
2231 RTMemFree(pWaiter);
2232
2233 return pNext;
2234}
2235
2236static void pdmBlkCacheIoXferCompleteEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2237{
2238 PPDMBLKCACHEENTRY pEntry = hIoXfer->pEntry;
2239 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2240
2241 /* Reference the entry now as we are clearing the I/O in progress flag
2242 * which protected the entry till now. */
2243 pdmBlkCacheEntryRef(pEntry);
2244
2245 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2246 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
2247
2248 /* Process waiting segment list. The data in entry might have changed in-between. */
2249 bool fDirty = false;
2250 PPDMBLKCACHEWAITER pComplete = pEntry->pWaitingHead;
2251 PPDMBLKCACHEWAITER pCurr = pComplete;
2252
2253 AssertMsg((pCurr && pEntry->pWaitingTail) || (!pCurr && !pEntry->pWaitingTail),
2254 ("The list tail was not updated correctly\n"));
2255 pEntry->pWaitingTail = NULL;
2256 pEntry->pWaitingHead = NULL;
2257
2258 if (hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_WRITE)
2259 {
2260 /*
2261 * An error here is difficult to handle as the original request completed already.
2262 * The error is logged for now and the VM is paused.
2263 * If the user continues the entry is written again in the hope
2264 * the user fixed the problem and the next write succeeds.
2265 */
2266 /** @todo r=aeichner: This solution doesn't work
2267 * The user will get the message but the VM will hang afterwards
2268 * VMR3Suspend() returns when the VM is suspended but suspending
2269 * the VM will reopen the images readonly in DrvVD. They are closed first
2270 * which will close the endpoints. This will block EMT while the
2271 * I/O manager processes the close request but the IO manager is stuck
2272 * in the VMR3Suspend call and can't process the request.
2273 * Another problem is that closing the VM means flushing the cache
2274 * but the entry failed and will probably fail again.
2275 * No idea so far how to solve this problem... but the user gets informed
2276 * at least.
2277 */
2278 if (RT_FAILURE(rcIoXfer))
2279 {
2280 LogRel(("I/O cache: Error while writing entry at offset %llu (%u bytes) to medium \"%s\"\n",
2281 pEntry->Core.Key, pEntry->cbData, pBlkCache->pszId));
2282
2283 int rc = VMSetRuntimeError(pCache->pVM, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "BLKCACHE_IOERR",
2284 N_("The I/O cache encountered an error while updating data in medium \"%s\" (rc=%Rrc)."
2285 "Make sure there is enough free space on the disk and that the disk is working properly."
2286 "Operation can be resumed afterwards."),
2287 pBlkCache->pszId, rcIoXfer);
2288 AssertRC(rc);
2289 }
2290 else
2291 {
2292 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IS_DIRTY;
2293
2294 while (pCurr)
2295 {
2296 AssertMsg(pCurr->fWrite, ("Completed write entries should never have read tasks attached\n"));
2297
2298 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2299 fDirty = true;
2300
2301 pCurr = pCurr->pNext;
2302 }
2303 }
2304 }
2305 else
2306 {
2307 AssertMsg(hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_READ, ("Invalid transfer type\n"));
2308 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY),
2309 ("Invalid flags set\n"));
2310
2311 while (pCurr)
2312 {
2313 if (pCurr->fWrite)
2314 {
2315 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2316 fDirty = true;
2317 }
2318 else
2319 RTSgBufCopyFromBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2320
2321 pCurr = pCurr->pNext;
2322 }
2323 }
2324
2325 bool fCommit = false;
2326 if (fDirty)
2327 fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2328
2329 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2330
2331 /* Dereference so that it isn't protected anymore except we issued anyother write for it. */
2332 pdmBlkCacheEntryRelease(pEntry);
2333
2334 if (fCommit)
2335 pdmBlkCacheCommitDirtyEntries(pCache);
2336
2337 /* Complete waiters now. */
2338 while (pComplete)
2339 pComplete = pdmBlkCacheWaiterComplete(pBlkCache, pComplete, rcIoXfer);
2340}
2341
2342VMMR3DECL(void) PDMR3BlkCacheIoXferComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2343{
2344 LogFlowFunc(("pBlkCache=%#p hIoXfer=%#p rcIoXfer=%Rrc\n", pBlkCache, hIoXfer, rcIoXfer));
2345
2346 if (hIoXfer->fIoCache)
2347 pdmBlkCacheIoXferCompleteEntry(pBlkCache, hIoXfer, rcIoXfer);
2348 else
2349 pdmBlkCacheReqUpdate(pBlkCache, hIoXfer->pReq, hIoXfer->cbXfer, rcIoXfer, true);
2350}
2351
Note: See TracBrowser for help on using the repository browser.

© 2024 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette