VirtualBox

source: vbox/trunk/src/VBox/VMM/PDMBlkCache.cpp@ 34347

Last change on this file since 34347 was 34347, checked in by vboxsync, 14 years ago

BlkCache: Suspend the VM if an error occurs during a write

  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 84.9 KB
Line 
1/* $Id: PDMBlkCache.cpp 34347 2010-11-24 22:34:21Z 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 if (pSgBuf)
611 {
612 RTSgBufClone(&pIoXfer->SgBuf, pSgBuf);
613 RTSgBufAdvance(pSgBuf, cbData);
614 }
615
616 return pdmBlkCacheEnqueue(pBlkCache, offStart, pIoXfer);
617}
618
619/**
620 * Commit a single dirty entry to the endpoint
621 *
622 * @returns nothing
623 * @param pEntry The entry to commit.
624 */
625static void pdmBlkCacheEntryCommit(PPDMBLKCACHEENTRY pEntry)
626{
627 AssertMsg( (pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY)
628 && !(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
629 ("Invalid flags set for entry %#p\n", pEntry));
630
631 pdmBlkCacheEntryWriteToMedium(pEntry);
632}
633
634/**
635 * Commit all dirty entries for a single endpoint.
636 *
637 * @returns nothing.
638 * @param pBlkCache The endpoint cache to commit.
639 */
640static void pdmBlkCacheCommit(PPDMBLKCACHE pBlkCache)
641{
642 uint32_t cbCommitted = 0;
643 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
644
645 /* The list is moved to a new header to reduce locking overhead. */
646 RTLISTNODE ListDirtyNotCommitted;
647 RTSPINLOCKTMP Tmp;
648
649 RTListInit(&ListDirtyNotCommitted);
650 RTSpinlockAcquire(pBlkCache->LockList, &Tmp);
651 RTListMove(&ListDirtyNotCommitted, &pBlkCache->ListDirtyNotCommitted);
652 RTSpinlockRelease(pBlkCache->LockList, &Tmp);
653
654 if (!RTListIsEmpty(&ListDirtyNotCommitted))
655 {
656 PPDMBLKCACHEENTRY pEntry = RTListNodeGetFirst(&ListDirtyNotCommitted,
657 PDMBLKCACHEENTRY,
658 NodeNotCommitted);
659
660 while (!RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted))
661 {
662 PPDMBLKCACHEENTRY pNext = RTListNodeGetNext(&pEntry->NodeNotCommitted, PDMBLKCACHEENTRY,
663 NodeNotCommitted);
664 pdmBlkCacheEntryCommit(pEntry);
665 cbCommitted += pEntry->cbData;
666 RTListNodeRemove(&pEntry->NodeNotCommitted);
667 pEntry = pNext;
668 }
669
670 /* Commit the last endpoint */
671 Assert(RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted));
672 pdmBlkCacheEntryCommit(pEntry);
673 RTListNodeRemove(&pEntry->NodeNotCommitted);
674 AssertMsg(RTListIsEmpty(&ListDirtyNotCommitted),
675 ("Committed all entries but list is not empty\n"));
676 }
677
678 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
679 AssertMsg(pBlkCache->pCache->cbDirty >= cbCommitted,
680 ("Number of committed bytes exceeds number of dirty bytes\n"));
681 ASMAtomicSubU32(&pBlkCache->pCache->cbDirty, cbCommitted);
682}
683
684/**
685 * Commit all dirty entries in the cache.
686 *
687 * @returns nothing.
688 * @param pCache The global cache instance.
689 */
690static void pdmBlkCacheCommitDirtyEntries(PPDMBLKCACHEGLOBAL pCache)
691{
692 bool fCommitInProgress = ASMAtomicXchgBool(&pCache->fCommitInProgress, true);
693
694 if (!fCommitInProgress)
695 {
696 pdmBlkCacheLockEnter(pCache);
697 Assert(!RTListIsEmpty(&pCache->ListUsers));
698
699 PPDMBLKCACHE pBlkCache = RTListNodeGetFirst(&pCache->ListUsers,
700 PDMBLKCACHE,
701 NodeCacheUser);
702 AssertPtr(pBlkCache);
703
704 while (!RTListNodeIsLast(&pCache->ListUsers, &pBlkCache->NodeCacheUser))
705 {
706 pdmBlkCacheCommit(pBlkCache);
707
708 pBlkCache = RTListNodeGetNext(&pBlkCache->NodeCacheUser, PDMBLKCACHE,
709 NodeCacheUser);
710 }
711
712 /* Commit the last endpoint */
713 Assert(RTListNodeIsLast(&pCache->ListUsers, &pBlkCache->NodeCacheUser));
714 pdmBlkCacheCommit(pBlkCache);
715
716 pdmBlkCacheLockLeave(pCache);
717 ASMAtomicWriteBool(&pCache->fCommitInProgress, false);
718 }
719}
720
721/**
722 * Adds the given entry as a dirty to the cache.
723 *
724 * @returns Flag whether the amount of dirty bytes in the cache exceeds the threshold
725 * @param pBlkCache The endpoint cache the entry belongs to.
726 * @param pEntry The entry to add.
727 */
728static bool pdmBlkCacheAddDirtyEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
729{
730 bool fDirtyBytesExceeded = false;
731 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
732
733 /* If the commit timer is disabled we commit right away. */
734 if (pCache->u32CommitTimeoutMs == 0)
735 {
736 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IS_DIRTY;
737 pdmBlkCacheEntryCommit(pEntry);
738 }
739 else if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY))
740 {
741 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IS_DIRTY;
742
743 RTSPINLOCKTMP Tmp;
744 RTSpinlockAcquire(pBlkCache->LockList, &Tmp);
745 RTListAppend(&pBlkCache->ListDirtyNotCommitted, &pEntry->NodeNotCommitted);
746 RTSpinlockRelease(pBlkCache->LockList, &Tmp);
747
748 uint32_t cbDirty = ASMAtomicAddU32(&pCache->cbDirty, pEntry->cbData);
749
750 /* Prevent committing if the VM was suspended. */
751 if (RT_LIKELY(!ASMAtomicReadBool(&pCache->fIoErrorVmSuspended)))
752 fDirtyBytesExceeded = (cbDirty >= pCache->cbCommitDirtyThreshold);
753 }
754
755 return fDirtyBytesExceeded;
756}
757
758/**
759 * Commit timer callback.
760 */
761static void pdmBlkCacheCommitTimerCallback(PVM pVM, PTMTIMER pTimer, void *pvUser)
762{
763 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
764
765 LogFlowFunc(("Commit interval expired, commiting dirty entries\n"));
766
767 if ( ASMAtomicReadU32(&pCache->cbDirty) > 0
768 && !ASMAtomicReadBool(&pCache->fIoErrorVmSuspended))
769 pdmBlkCacheCommitDirtyEntries(pCache);
770
771 TMTimerSetMillies(pTimer, pCache->u32CommitTimeoutMs);
772 LogFlowFunc(("Entries committed, going to sleep\n"));
773}
774
775int pdmR3BlkCacheInit(PVM pVM)
776{
777 int rc = VINF_SUCCESS;
778 PUVM pUVM = pVM->pUVM;
779 PPDMBLKCACHEGLOBAL pBlkCacheGlobal;
780
781 LogFlowFunc((": pVM=%p\n", pVM));
782
783 VM_ASSERT_EMT(pVM);
784
785 PCFGMNODE pCfgRoot = CFGMR3GetRoot(pVM);
786 PCFGMNODE pCfgBlkCache = CFGMR3GetChild(CFGMR3GetChild(pCfgRoot, "PDM"), "BlkCache");
787
788 pBlkCacheGlobal = (PPDMBLKCACHEGLOBAL)RTMemAllocZ(sizeof(PDMBLKCACHEGLOBAL));
789 if (!pBlkCacheGlobal)
790 return VERR_NO_MEMORY;
791
792 RTListInit(&pBlkCacheGlobal->ListUsers);
793 pBlkCacheGlobal->pVM = pVM;
794 pBlkCacheGlobal->cRefs = 0;
795 pBlkCacheGlobal->cbCached = 0;
796 pBlkCacheGlobal->fCommitInProgress = false;
797
798 /* Initialize members */
799 pBlkCacheGlobal->LruRecentlyUsedIn.pHead = NULL;
800 pBlkCacheGlobal->LruRecentlyUsedIn.pTail = NULL;
801 pBlkCacheGlobal->LruRecentlyUsedIn.cbCached = 0;
802
803 pBlkCacheGlobal->LruRecentlyUsedOut.pHead = NULL;
804 pBlkCacheGlobal->LruRecentlyUsedOut.pTail = NULL;
805 pBlkCacheGlobal->LruRecentlyUsedOut.cbCached = 0;
806
807 pBlkCacheGlobal->LruFrequentlyUsed.pHead = NULL;
808 pBlkCacheGlobal->LruFrequentlyUsed.pTail = NULL;
809 pBlkCacheGlobal->LruFrequentlyUsed.cbCached = 0;
810
811 do
812 {
813 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheSize", &pBlkCacheGlobal->cbMax, 5 * _1M);
814 AssertLogRelRCBreak(rc);
815 LogFlowFunc(("Maximum number of bytes cached %u\n", pBlkCacheGlobal->cbMax));
816
817 pBlkCacheGlobal->cbRecentlyUsedInMax = (pBlkCacheGlobal->cbMax / 100) * 25; /* 25% of the buffer size */
818 pBlkCacheGlobal->cbRecentlyUsedOutMax = (pBlkCacheGlobal->cbMax / 100) * 50; /* 50% of the buffer size */
819 LogFlowFunc(("cbRecentlyUsedInMax=%u cbRecentlyUsedOutMax=%u\n",
820 pBlkCacheGlobal->cbRecentlyUsedInMax, pBlkCacheGlobal->cbRecentlyUsedOutMax));
821
822 /** @todo r=aeichner: Experiment to find optimal default values */
823 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheCommitIntervalMs", &pBlkCacheGlobal->u32CommitTimeoutMs, 10000 /* 10sec */);
824 AssertLogRelRCBreak(rc);
825 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheCommitThreshold", &pBlkCacheGlobal->cbCommitDirtyThreshold, pBlkCacheGlobal->cbMax / 2);
826 AssertLogRelRCBreak(rc);
827 } while (0);
828
829 if (RT_SUCCESS(rc))
830 {
831 STAMR3Register(pVM, &pBlkCacheGlobal->cbMax,
832 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
833 "/PDM/BlkCache/cbMax",
834 STAMUNIT_BYTES,
835 "Maximum cache size");
836 STAMR3Register(pVM, &pBlkCacheGlobal->cbCached,
837 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
838 "/PDM/BlkCache/cbCached",
839 STAMUNIT_BYTES,
840 "Currently used cache");
841 STAMR3Register(pVM, &pBlkCacheGlobal->LruRecentlyUsedIn.cbCached,
842 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
843 "/PDM/BlkCache/cbCachedMruIn",
844 STAMUNIT_BYTES,
845 "Number of bytes cached in MRU list");
846 STAMR3Register(pVM, &pBlkCacheGlobal->LruRecentlyUsedOut.cbCached,
847 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
848 "/PDM/BlkCache/cbCachedMruOut",
849 STAMUNIT_BYTES,
850 "Number of bytes cached in FRU list");
851 STAMR3Register(pVM, &pBlkCacheGlobal->LruFrequentlyUsed.cbCached,
852 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
853 "/PDM/BlkCache/cbCachedFru",
854 STAMUNIT_BYTES,
855 "Number of bytes cached in FRU ghost list");
856
857#ifdef VBOX_WITH_STATISTICS
858 STAMR3Register(pVM, &pBlkCacheGlobal->cHits,
859 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
860 "/PDM/BlkCache/CacheHits",
861 STAMUNIT_COUNT, "Number of hits in the cache");
862 STAMR3Register(pVM, &pBlkCacheGlobal->cPartialHits,
863 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
864 "/PDM/BlkCache/CachePartialHits",
865 STAMUNIT_COUNT, "Number of partial hits in the cache");
866 STAMR3Register(pVM, &pBlkCacheGlobal->cMisses,
867 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
868 "/PDM/BlkCache/CacheMisses",
869 STAMUNIT_COUNT, "Number of misses when accessing the cache");
870 STAMR3Register(pVM, &pBlkCacheGlobal->StatRead,
871 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
872 "/PDM/BlkCache/CacheRead",
873 STAMUNIT_BYTES, "Number of bytes read from the cache");
874 STAMR3Register(pVM, &pBlkCacheGlobal->StatWritten,
875 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
876 "/PDM/BlkCache/CacheWritten",
877 STAMUNIT_BYTES, "Number of bytes written to the cache");
878 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeGet,
879 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
880 "/PDM/BlkCache/CacheTreeGet",
881 STAMUNIT_TICKS_PER_CALL, "Time taken to access an entry in the tree");
882 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeInsert,
883 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
884 "/PDM/BlkCache/CacheTreeInsert",
885 STAMUNIT_TICKS_PER_CALL, "Time taken to insert an entry in the tree");
886 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeRemove,
887 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
888 "/PDM/BlkCache/CacheTreeRemove",
889 STAMUNIT_TICKS_PER_CALL, "Time taken to remove an entry an the tree");
890 STAMR3Register(pVM, &pBlkCacheGlobal->StatBuffersReused,
891 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
892 "/PDM/BlkCache/CacheBuffersReused",
893 STAMUNIT_COUNT, "Number of times a buffer could be reused");
894#endif
895
896 /* Initialize the critical section */
897 rc = RTCritSectInit(&pBlkCacheGlobal->CritSect);
898 }
899
900 if (RT_SUCCESS(rc))
901 {
902 /* Create the commit timer */
903 if (pBlkCacheGlobal->u32CommitTimeoutMs > 0)
904 rc = TMR3TimerCreateInternal(pVM, TMCLOCK_REAL,
905 pdmBlkCacheCommitTimerCallback,
906 pBlkCacheGlobal,
907 "BlkCache-Commit",
908 &pBlkCacheGlobal->pTimerCommit);
909
910 if (RT_SUCCESS(rc))
911 {
912 LogRel(("BlkCache: Cache successfully initialised. Cache size is %u bytes\n", pBlkCacheGlobal->cbMax));
913 LogRel(("BlkCache: Cache commit interval is %u ms\n", pBlkCacheGlobal->u32CommitTimeoutMs));
914 LogRel(("BlkCache: Cache commit threshold is %u bytes\n", pBlkCacheGlobal->cbCommitDirtyThreshold));
915 pUVM->pdm.s.pBlkCacheGlobal = pBlkCacheGlobal;
916 return VINF_SUCCESS;
917 }
918
919 RTCritSectDelete(&pBlkCacheGlobal->CritSect);
920 }
921
922 if (pBlkCacheGlobal)
923 RTMemFree(pBlkCacheGlobal);
924
925 LogFlowFunc((": returns rc=%Rrc\n", pVM, rc));
926 return rc;
927}
928
929void pdmR3BlkCacheTerm(PVM pVM)
930{
931 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
932
933 if (pBlkCacheGlobal)
934 {
935 /* Make sure no one else uses the cache now */
936 pdmBlkCacheLockEnter(pBlkCacheGlobal);
937
938 /* Cleanup deleting all cache entries waiting for in progress entries to finish. */
939 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruRecentlyUsedIn);
940 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruRecentlyUsedOut);
941 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruFrequentlyUsed);
942
943 pdmBlkCacheLockLeave(pBlkCacheGlobal);
944
945 RTCritSectDelete(&pBlkCacheGlobal->CritSect);
946 RTMemFree(pBlkCacheGlobal);
947 pVM->pUVM->pdm.s.pBlkCacheGlobal = NULL;
948 }
949}
950
951int pdmR3BlkCacheResume(PVM pVM)
952{
953 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
954
955 LogFlowFunc(("pVM=%#p\n", pVM));
956
957 if ( pBlkCacheGlobal
958 && ASMAtomicXchgBool(&pBlkCacheGlobal->fIoErrorVmSuspended, false))
959 {
960 /* The VM was suspended because of an I/O error, commit all dirty entries. */
961 pdmBlkCacheCommitDirtyEntries(pBlkCacheGlobal);
962 }
963
964 return VINF_SUCCESS;
965}
966
967static int pdmR3BlkCacheRetain(PVM pVM, PPPDMBLKCACHE ppBlkCache, const char *pcszId)
968{
969 int rc = VINF_SUCCESS;
970 PPDMBLKCACHE pBlkCache = NULL;
971 bool fAlreadyExists = false;
972 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
973
974 if (!pBlkCacheGlobal)
975 return VERR_NOT_SUPPORTED;
976
977 /*
978 * Check that no other user cache has the same id first,
979 * Unique id's are necessary in case the state is saved.
980 */
981 pdmBlkCacheLockEnter(pBlkCacheGlobal);
982
983 RTListForEach(&pBlkCacheGlobal->ListUsers, pBlkCache, PDMBLKCACHE, NodeCacheUser)
984 {
985 if (!RTStrCmp(pBlkCache->pszId, pcszId))
986 {
987 fAlreadyExists = true;
988 break;
989 }
990 }
991
992 if (!fAlreadyExists)
993 {
994 pBlkCache = (PPDMBLKCACHE)RTMemAllocZ(sizeof(PDMBLKCACHE));
995
996 if (pBlkCache)
997 pBlkCache->pszId = RTStrDup(pcszId);
998
999 if ( pBlkCache
1000 && pBlkCache->pszId)
1001 {
1002 pBlkCache->pCache = pBlkCacheGlobal;
1003 RTListInit(&pBlkCache->ListDirtyNotCommitted);
1004
1005 rc = RTSpinlockCreate(&pBlkCache->LockList);
1006 if (RT_SUCCESS(rc))
1007 {
1008 rc = RTSemRWCreate(&pBlkCache->SemRWEntries);
1009 if (RT_SUCCESS(rc))
1010 {
1011 pBlkCache->pTree = (PAVLRU64TREE)RTMemAllocZ(sizeof(AVLRFOFFTREE));
1012 if (pBlkCache->pTree)
1013 {
1014 /* Arm the timer if this is the first endpoint. */
1015 if ( pBlkCacheGlobal->cRefs == 1
1016 && pBlkCacheGlobal->u32CommitTimeoutMs > 0)
1017 rc = TMTimerSetMillies(pBlkCacheGlobal->pTimerCommit, pBlkCacheGlobal->u32CommitTimeoutMs);
1018
1019 if (RT_SUCCESS(rc))
1020 {
1021#ifdef VBOX_WITH_STATISTICS
1022 STAMR3RegisterF(pBlkCacheGlobal->pVM, &pBlkCache->StatWriteDeferred,
1023 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1024 STAMUNIT_COUNT, "Number of deferred writes",
1025 "/PDM/BlkCache/%s/Cache/DeferredWrites", pBlkCache->pszId);
1026#endif
1027
1028 /* Add to the list of users. */
1029 pBlkCacheGlobal->cRefs++;
1030 RTListAppend(&pBlkCacheGlobal->ListUsers, &pBlkCache->NodeCacheUser);
1031 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1032
1033 *ppBlkCache = pBlkCache;
1034 LogFlowFunc(("returns success\n"));
1035 return VINF_SUCCESS;
1036 }
1037 }
1038 else
1039 rc = VERR_NO_MEMORY;
1040
1041 RTSemRWDestroy(pBlkCache->SemRWEntries);
1042 }
1043
1044 RTSpinlockDestroy(pBlkCache->LockList);
1045 }
1046
1047 RTStrFree(pBlkCache->pszId);
1048 }
1049 else
1050 rc = VERR_NO_MEMORY;
1051
1052 if (pBlkCache)
1053 RTMemFree(pBlkCache);
1054 }
1055 else
1056 rc = VERR_ALREADY_EXISTS;
1057
1058 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1059
1060 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1061 return rc;
1062}
1063
1064VMMR3DECL(int) PDMR3BlkCacheRetainDriver(PVM pVM, PPDMDRVINS pDrvIns, PPPDMBLKCACHE ppBlkCache,
1065 PFNPDMBLKCACHEXFERCOMPLETEDRV pfnXferComplete,
1066 PFNPDMBLKCACHEXFERENQUEUEDRV pfnXferEnqueue,
1067 const char *pcszId)
1068{
1069 int rc = VINF_SUCCESS;
1070 PPDMBLKCACHE pBlkCache;
1071
1072 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1073 if (RT_SUCCESS(rc))
1074 {
1075 pBlkCache->enmType = PDMBLKCACHETYPE_DRV;
1076 pBlkCache->u.Drv.pfnXferComplete = pfnXferComplete;
1077 pBlkCache->u.Drv.pfnXferEnqueue = pfnXferEnqueue;
1078 pBlkCache->u.Drv.pDrvIns = pDrvIns;
1079 *ppBlkCache = pBlkCache;
1080 }
1081
1082 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1083 return rc;
1084}
1085
1086VMMR3DECL(int) PDMR3BlkCacheRetainDevice(PVM pVM, PPDMDEVINS pDevIns, PPPDMBLKCACHE ppBlkCache,
1087 PFNPDMBLKCACHEXFERCOMPLETEDEV pfnXferComplete,
1088 PFNPDMBLKCACHEXFERENQUEUEDEV 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_DEV;
1098 pBlkCache->u.Dev.pfnXferComplete = pfnXferComplete;
1099 pBlkCache->u.Dev.pfnXferEnqueue = pfnXferEnqueue;
1100 pBlkCache->u.Dev.pDevIns = pDevIns;
1101 *ppBlkCache = pBlkCache;
1102 }
1103
1104 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1105 return rc;
1106
1107}
1108
1109VMMR3DECL(int) PDMR3BlkCacheRetainUsb(PVM pVM, PPDMUSBINS pUsbIns, PPPDMBLKCACHE ppBlkCache,
1110 PFNPDMBLKCACHEXFERCOMPLETEUSB pfnXferComplete,
1111 PFNPDMBLKCACHEXFERENQUEUEUSB 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_USB;
1121 pBlkCache->u.Usb.pfnXferComplete = pfnXferComplete;
1122 pBlkCache->u.Usb.pfnXferEnqueue = pfnXferEnqueue;
1123 pBlkCache->u.Usb.pUsbIns = pUsbIns;
1124 *ppBlkCache = pBlkCache;
1125 }
1126
1127 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1128 return rc;
1129
1130}
1131
1132VMMR3DECL(int) PDMR3BlkCacheRetainInt(PVM pVM, void *pvUser, PPPDMBLKCACHE ppBlkCache,
1133 PFNPDMBLKCACHEXFERCOMPLETEINT pfnXferComplete,
1134 PFNPDMBLKCACHEXFERENQUEUEINT pfnXferEnqueue,
1135 const char *pcszId)
1136{
1137 int rc = VINF_SUCCESS;
1138 PPDMBLKCACHE pBlkCache;
1139
1140 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1141 if (RT_SUCCESS(rc))
1142 {
1143 pBlkCache->enmType = PDMBLKCACHETYPE_INTERNAL;
1144 pBlkCache->u.Int.pfnXferComplete = pfnXferComplete;
1145 pBlkCache->u.Int.pfnXferEnqueue = pfnXferEnqueue;
1146 pBlkCache->u.Int.pvUser = pvUser;
1147 *ppBlkCache = pBlkCache;
1148 }
1149
1150 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1151 return rc;
1152
1153}
1154
1155/**
1156 * Callback for the AVL destroy routine. Frees a cache entry for this endpoint.
1157 *
1158 * @returns IPRT status code.
1159 * @param pNode The node to destroy.
1160 * @param pvUser Opaque user data.
1161 */
1162static int pdmBlkCacheEntryDestroy(PAVLRU64NODECORE pNode, void *pvUser)
1163{
1164 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)pNode;
1165 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
1166 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
1167
1168 while (ASMAtomicReadU32(&pEntry->fFlags) & (PDMBLKCACHE_ENTRY_IO_IN_PROGRESS | PDMBLKCACHE_ENTRY_IS_DIRTY))
1169 {
1170 /* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
1171 pdmBlkCacheEntryRef(pEntry);
1172 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1173 pdmBlkCacheLockLeave(pCache);
1174
1175 RTThreadSleep(250);
1176
1177 /* Re-enter all locks */
1178 pdmBlkCacheLockEnter(pCache);
1179 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1180 pdmBlkCacheEntryRelease(pEntry);
1181 }
1182
1183 AssertMsg(!(pEntry->fFlags & (PDMBLKCACHE_ENTRY_IO_IN_PROGRESS | PDMBLKCACHE_ENTRY_IS_DIRTY)),
1184 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
1185
1186 bool fUpdateCache = pEntry->pList == &pCache->LruFrequentlyUsed
1187 || pEntry->pList == &pCache->LruRecentlyUsedIn;
1188
1189 pdmBlkCacheEntryRemoveFromList(pEntry);
1190
1191 if (fUpdateCache)
1192 pdmBlkCacheSub(pCache, pEntry->cbData);
1193
1194 RTMemPageFree(pEntry->pbData, pEntry->cbData);
1195 RTMemFree(pEntry);
1196
1197 return VINF_SUCCESS;
1198}
1199
1200/**
1201 * Destroys all cache resources used by the given endpoint.
1202 *
1203 * @returns nothing.
1204 * @param pEndpoint The endpoint to the destroy.
1205 */
1206VMMR3DECL(void) PDMR3BlkCacheRelease(PPDMBLKCACHE pBlkCache)
1207{
1208 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1209
1210 /* Make sure nobody is accessing the cache while we delete the tree. */
1211 pdmBlkCacheLockEnter(pCache);
1212 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1213 RTAvlrU64Destroy(pBlkCache->pTree, pdmBlkCacheEntryDestroy, pCache);
1214 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1215
1216 RTSpinlockDestroy(pBlkCache->LockList);
1217
1218 pCache->cRefs--;
1219 RTListNodeRemove(&pBlkCache->NodeCacheUser);
1220
1221 if ( !pCache->cRefs
1222 && pCache->u32CommitTimeoutMs > 0)
1223 TMTimerStop(pCache->pTimerCommit);
1224
1225 pdmBlkCacheLockLeave(pCache);
1226
1227 RTSemRWDestroy(pBlkCache->SemRWEntries);
1228
1229#ifdef VBOX_WITH_STATISTICS
1230 STAMR3Deregister(pCache->pVM, &pBlkCache->StatWriteDeferred);
1231#endif
1232
1233 RTStrFree(pBlkCache->pszId);
1234 RTMemFree(pBlkCache);
1235}
1236
1237VMMR3DECL(void) PDMR3BlkCacheReleaseDevice(PVM pVM, PPDMDEVINS pDevIns)
1238{
1239 LogFlow(("%s: pDevIns=%p\n", __FUNCTION__, pDevIns));
1240
1241 /*
1242 * Validate input.
1243 */
1244 if (!pDevIns)
1245 return;
1246 VM_ASSERT_EMT(pVM);
1247
1248 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1249 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1250
1251 /* Return silently if not supported. */
1252 if (!pBlkCacheGlobal)
1253 return;
1254
1255 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1256
1257 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1258 {
1259 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DEV
1260 && pBlkCache->u.Dev.pDevIns == pDevIns)
1261 PDMR3BlkCacheRelease(pBlkCache);
1262 }
1263
1264 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1265}
1266
1267VMMR3DECL(void) PDMR3BlkCacheReleaseDriver(PVM pVM, PPDMDRVINS pDrvIns)
1268{
1269 LogFlow(("%s: pDrvIns=%p\n", __FUNCTION__, pDrvIns));
1270
1271 /*
1272 * Validate input.
1273 */
1274 if (!pDrvIns)
1275 return;
1276 VM_ASSERT_EMT(pVM);
1277
1278 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1279 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1280
1281 /* Return silently if not supported. */
1282 if (!pBlkCacheGlobal)
1283 return;
1284
1285 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1286
1287 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1288 {
1289 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DRV
1290 && pBlkCache->u.Drv.pDrvIns == pDrvIns)
1291 PDMR3BlkCacheRelease(pBlkCache);
1292 }
1293
1294 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1295}
1296
1297VMMR3DECL(void) PDMR3BlkCacheReleaseUsb(PVM pVM, PPDMUSBINS pUsbIns)
1298{
1299 LogFlow(("%s: pUsbIns=%p\n", __FUNCTION__, pUsbIns));
1300
1301 /*
1302 * Validate input.
1303 */
1304 if (!pUsbIns)
1305 return;
1306 VM_ASSERT_EMT(pVM);
1307
1308 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1309 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1310
1311 /* Return silently if not supported. */
1312 if (!pBlkCacheGlobal)
1313 return;
1314
1315 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1316
1317 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1318 {
1319 if ( pBlkCache->enmType == PDMBLKCACHETYPE_USB
1320 && pBlkCache->u.Usb.pUsbIns == pUsbIns)
1321 PDMR3BlkCacheRelease(pBlkCache);
1322 }
1323
1324 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1325}
1326
1327static PPDMBLKCACHEENTRY pdmBlkCacheGetCacheEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off)
1328{
1329 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1330 PPDMBLKCACHEENTRY pEntry = NULL;
1331
1332 STAM_PROFILE_ADV_START(&pCache->StatTreeGet, Cache);
1333
1334 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1335 pEntry = (PPDMBLKCACHEENTRY)RTAvlrU64RangeGet(pBlkCache->pTree, off);
1336 if (pEntry)
1337 pdmBlkCacheEntryRef(pEntry);
1338 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1339
1340 STAM_PROFILE_ADV_STOP(&pCache->StatTreeGet, Cache);
1341
1342 return pEntry;
1343}
1344
1345/**
1346 * Return the best fit cache entries for the given offset.
1347 *
1348 * @returns nothing.
1349 * @param pBlkCache The endpoint cache.
1350 * @param off The offset.
1351 * @param pEntryAbove Where to store the pointer to the best fit entry above the
1352 * the given offset. NULL if not required.
1353 * @param pEntryBelow Where to store the pointer to the best fit entry below the
1354 * the given offset. NULL if not required.
1355 */
1356static void pdmBlkCacheGetCacheBestFitEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off,
1357 PPDMBLKCACHEENTRY *ppEntryAbove,
1358 PPDMBLKCACHEENTRY *ppEntryBelow)
1359{
1360 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1361
1362 STAM_PROFILE_ADV_START(&pCache->StatTreeGet, Cache);
1363
1364 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1365 if (ppEntryAbove)
1366 {
1367 *ppEntryAbove = (PPDMBLKCACHEENTRY)RTAvlrU64GetBestFit(pBlkCache->pTree, off, true /*fAbove*/);
1368 if (*ppEntryAbove)
1369 pdmBlkCacheEntryRef(*ppEntryAbove);
1370 }
1371
1372 if (ppEntryBelow)
1373 {
1374 *ppEntryBelow = (PPDMBLKCACHEENTRY)RTAvlrU64GetBestFit(pBlkCache->pTree, off, false /*fAbove*/);
1375 if (*ppEntryBelow)
1376 pdmBlkCacheEntryRef(*ppEntryBelow);
1377 }
1378 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1379
1380 STAM_PROFILE_ADV_STOP(&pCache->StatTreeGet, Cache);
1381}
1382
1383static void pdmBlkCacheInsertEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
1384{
1385 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1386
1387 STAM_PROFILE_ADV_START(&pCache->StatTreeInsert, Cache);
1388 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1389 bool fInserted = RTAvlrU64Insert(pBlkCache->pTree, &pEntry->Core);
1390 AssertMsg(fInserted, ("Node was not inserted into tree\n"));
1391 STAM_PROFILE_ADV_STOP(&pCache->StatTreeInsert, Cache);
1392 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1393}
1394
1395/**
1396 * Allocates and initializes a new entry for the cache.
1397 * The entry has a reference count of 1.
1398 *
1399 * @returns Pointer to the new cache entry or NULL if out of memory.
1400 * @param pBlkCache The cache the entry belongs to.
1401 * @param off Start offset.
1402 * @param cbData Size of the cache entry.
1403 * @param pbBuffer Pointer to the buffer to use.
1404 * NULL if a new buffer should be allocated.
1405 * The buffer needs to have the same size of the entry.
1406 */
1407static PPDMBLKCACHEENTRY pdmBlkCacheEntryAlloc(PPDMBLKCACHE pBlkCache,
1408 uint64_t off, size_t cbData, uint8_t *pbBuffer)
1409{
1410 PPDMBLKCACHEENTRY pEntryNew = (PPDMBLKCACHEENTRY)RTMemAllocZ(sizeof(PDMBLKCACHEENTRY));
1411
1412 if (RT_UNLIKELY(!pEntryNew))
1413 return NULL;
1414
1415 pEntryNew->Core.Key = off;
1416 pEntryNew->Core.KeyLast = off + cbData - 1;
1417 pEntryNew->pBlkCache = pBlkCache;
1418 pEntryNew->fFlags = 0;
1419 pEntryNew->cRefs = 1; /* We are using it now. */
1420 pEntryNew->pList = NULL;
1421 pEntryNew->cbData = cbData;
1422 pEntryNew->pWaitingHead = NULL;
1423 pEntryNew->pWaitingTail = NULL;
1424 if (pbBuffer)
1425 pEntryNew->pbData = pbBuffer;
1426 else
1427 pEntryNew->pbData = (uint8_t *)RTMemPageAlloc(cbData);
1428
1429 if (RT_UNLIKELY(!pEntryNew->pbData))
1430 {
1431 RTMemFree(pEntryNew);
1432 return NULL;
1433 }
1434
1435 return pEntryNew;
1436}
1437
1438/**
1439 * Checks that a set of flags is set/clear acquiring the R/W semaphore
1440 * in exclusive mode.
1441 *
1442 * @returns true if the flag in fSet is set and the one in fClear is clear.
1443 * false otherwise.
1444 * The R/W semaphore is only held if true is returned.
1445 *
1446 * @param pBlkCache The endpoint cache instance data.
1447 * @param pEntry The entry to check the flags for.
1448 * @param fSet The flag which is tested to be set.
1449 * @param fClear The flag which is tested to be clear.
1450 */
1451DECLINLINE(bool) pdmBlkCacheEntryFlagIsSetClearAcquireLock(PPDMBLKCACHE pBlkCache,
1452 PPDMBLKCACHEENTRY pEntry,
1453 uint32_t fSet, uint32_t fClear)
1454{
1455 uint32_t fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1456 bool fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1457
1458 if (fPassed)
1459 {
1460 /* Acquire the lock and check again because the completion callback might have raced us. */
1461 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1462
1463 fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1464 fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1465
1466 /* Drop the lock if we didn't passed the test. */
1467 if (!fPassed)
1468 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1469 }
1470
1471 return fPassed;
1472}
1473
1474/**
1475 * Adds a segment to the waiting list for a cache entry
1476 * which is currently in progress.
1477 *
1478 * @returns nothing.
1479 * @param pEntry The cache entry to add the segment to.
1480 * @param pSeg The segment to add.
1481 */
1482DECLINLINE(void) pdmBlkCacheEntryAddWaiter(PPDMBLKCACHEENTRY pEntry,
1483 PPDMBLKCACHEWAITER pWaiter)
1484{
1485 pWaiter->pNext = NULL;
1486
1487 if (pEntry->pWaitingHead)
1488 {
1489 AssertPtr(pEntry->pWaitingTail);
1490
1491 pEntry->pWaitingTail->pNext = pWaiter;
1492 pEntry->pWaitingTail = pWaiter;
1493 }
1494 else
1495 {
1496 Assert(!pEntry->pWaitingTail);
1497
1498 pEntry->pWaitingHead = pWaiter;
1499 pEntry->pWaitingTail = pWaiter;
1500 }
1501}
1502
1503/**
1504 * Add a buffer described by the I/O memory context
1505 * to the entry waiting for completion.
1506 *
1507 * @returns VBox status code.
1508 * @param pEntry The entry to add the buffer to.
1509 * @param pTask Task associated with the buffer.
1510 * @param pIoMemCtx The memory context to use.
1511 * @param offDiff Offset from the start of the buffer
1512 * in the entry.
1513 * @param cbData Amount of data to wait for onthis entry.
1514 * @param fWrite Flag whether the task waits because it wants to write
1515 * to the cache entry.
1516 */
1517static int pdmBlkCacheEntryWaitersAdd(PPDMBLKCACHEENTRY pEntry,
1518 PPDMBLKCACHEREQ pReq,
1519 PCRTSGBUF pSgBuf, uint64_t offDiff,
1520 size_t cbData, bool fWrite)
1521{
1522 PPDMBLKCACHEWAITER pWaiter = (PPDMBLKCACHEWAITER)RTMemAllocZ(sizeof(PDMBLKCACHEWAITER));
1523 if (!pWaiter)
1524 return VERR_NO_MEMORY;
1525
1526 ASMAtomicIncU32(&pReq->cXfersPending);
1527 pWaiter->pReq = pReq;
1528 pWaiter->offCacheEntry = offDiff;
1529 pWaiter->cbTransfer = cbData;
1530 pWaiter->fWrite = fWrite;
1531 RTSgBufClone(&pWaiter->SgBuf, pSgBuf);
1532
1533 pdmBlkCacheEntryAddWaiter(pEntry, pWaiter);
1534
1535 return VINF_SUCCESS;
1536}
1537
1538/**
1539 * Calculate aligned offset and size for a new cache entry
1540 * which do not intersect with an already existing entry and the
1541 * file end.
1542 *
1543 * @returns The number of bytes the entry can hold of the requested amount
1544 * of byte.
1545 * @param pEndpoint The endpoint.
1546 * @param pBlkCache The endpoint cache.
1547 * @param off The start offset.
1548 * @param cb The number of bytes the entry needs to hold at least.
1549 * @param uAlignment Alignment of the boundary sizes.
1550 * @param poffAligned Where to store the aligned offset.
1551 * @param pcbAligned Where to store the aligned size of the entry.
1552 */
1553static size_t pdmBlkCacheEntryBoundariesCalc(PPDMBLKCACHE pBlkCache,
1554 uint64_t off, size_t cb,
1555 unsigned uAlignment,
1556 uint64_t *poffAligned, size_t *pcbAligned)
1557{
1558 size_t cbAligned;
1559 size_t cbInEntry = 0;
1560 uint64_t offAligned;
1561 PPDMBLKCACHEENTRY pEntryAbove = NULL;
1562 PPDMBLKCACHEENTRY pEntryBelow = NULL;
1563
1564 /* Get the best fit entries around the offset */
1565 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off, &pEntryAbove, &pEntryBelow);
1566
1567 /* Log the info */
1568 LogFlow(("%sest fit entry below off=%llu (BestFit=%llu BestFitEnd=%llu BestFitSize=%u)\n",
1569 pEntryBelow ? "B" : "No b",
1570 off,
1571 pEntryBelow ? pEntryBelow->Core.Key : 0,
1572 pEntryBelow ? pEntryBelow->Core.KeyLast : 0,
1573 pEntryBelow ? pEntryBelow->cbData : 0));
1574
1575 LogFlow(("%sest fit entry above off=%llu (BestFit=%llu BestFitEnd=%llu BestFitSize=%u)\n",
1576 pEntryAbove ? "B" : "No b",
1577 off,
1578 pEntryAbove ? pEntryAbove->Core.Key : 0,
1579 pEntryAbove ? pEntryAbove->Core.KeyLast : 0,
1580 pEntryAbove ? pEntryAbove->cbData : 0));
1581
1582 /* Align the offset first. */
1583 offAligned = off & ~(uint64_t)(512-1);
1584 if ( pEntryBelow
1585 && offAligned <= pEntryBelow->Core.KeyLast)
1586 offAligned = pEntryBelow->Core.KeyLast;
1587
1588 if ( pEntryAbove
1589 && off + cb > pEntryAbove->Core.Key)
1590 {
1591 cbInEntry = pEntryAbove->Core.Key - off;
1592 cbAligned = pEntryAbove->Core.Key - offAligned;
1593 }
1594 else
1595 {
1596 cbAligned = cb;
1597 cbInEntry = cb;
1598 }
1599
1600 /* A few sanity checks */
1601 AssertMsg(!pEntryBelow || pEntryBelow->Core.KeyLast < offAligned,
1602 ("Aligned start offset intersects with another cache entry\n"));
1603 AssertMsg(!pEntryAbove || (offAligned + cbAligned) <= pEntryAbove->Core.Key,
1604 ("Aligned size intersects with another cache entry\n"));
1605 Assert(cbInEntry <= cbAligned);
1606
1607 if (pEntryBelow)
1608 pdmBlkCacheEntryRelease(pEntryBelow);
1609 if (pEntryAbove)
1610 pdmBlkCacheEntryRelease(pEntryAbove);
1611
1612 LogFlow(("offAligned=%llu cbAligned=%u\n", offAligned, cbAligned));
1613
1614 *poffAligned = offAligned;
1615 *pcbAligned = cbAligned;
1616
1617 return cbInEntry;
1618}
1619
1620/**
1621 * Create a new cache entry evicting data from the cache if required.
1622 *
1623 * @returns Pointer to the new cache entry or NULL
1624 * if not enough bytes could be evicted from the cache.
1625 * @param pEndpoint The endpoint.
1626 * @param pBlkCache The endpoint cache.
1627 * @param off The offset.
1628 * @param cb Number of bytes the cache entry should have.
1629 * @param uAlignment Alignment the size of the entry should have.
1630 * @param pcbData Where to store the number of bytes the new
1631 * entry can hold. May be lower than actually requested
1632 * due to another entry intersecting the access range.
1633 */
1634static PPDMBLKCACHEENTRY pdmBlkCacheEntryCreate(PPDMBLKCACHE pBlkCache,
1635 uint64_t off, size_t cb,
1636 unsigned uAlignment,
1637 size_t *pcbData)
1638{
1639 uint64_t offStart = 0;
1640 size_t cbEntry = 0;
1641 PPDMBLKCACHEENTRY pEntryNew = NULL;
1642 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1643 uint8_t *pbBuffer = NULL;
1644
1645 *pcbData = pdmBlkCacheEntryBoundariesCalc(pBlkCache, off, cb, uAlignment,
1646 &offStart, &cbEntry);
1647
1648 pdmBlkCacheLockEnter(pCache);
1649 bool fEnough = pdmBlkCacheReclaim(pCache, cbEntry, true, &pbBuffer);
1650
1651 if (fEnough)
1652 {
1653 LogFlow(("Evicted enough bytes (%u requested). Creating new cache entry\n", cbEntry));
1654
1655 pEntryNew = pdmBlkCacheEntryAlloc(pBlkCache, offStart, cbEntry, pbBuffer);
1656 if (RT_LIKELY(pEntryNew))
1657 {
1658 pdmBlkCacheEntryAddToList(&pCache->LruRecentlyUsedIn, pEntryNew);
1659 pdmBlkCacheAdd(pCache, cbEntry);
1660 pdmBlkCacheLockLeave(pCache);
1661
1662 pdmBlkCacheInsertEntry(pBlkCache, pEntryNew);
1663
1664 AssertMsg( (off >= pEntryNew->Core.Key)
1665 && (off + *pcbData <= pEntryNew->Core.KeyLast + 1),
1666 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
1667 off, pEntryNew->Core.Key));
1668 }
1669 else
1670 pdmBlkCacheLockLeave(pCache);
1671 }
1672 else
1673 pdmBlkCacheLockLeave(pCache);
1674
1675 return pEntryNew;
1676}
1677
1678static PPDMBLKCACHEREQ pdmBlkCacheReqAlloc(size_t cbXfer, void *pvUser)
1679{
1680 PPDMBLKCACHEREQ pReq = (PPDMBLKCACHEREQ)RTMemAlloc(sizeof(PDMBLKCACHEREQ));
1681
1682 if (RT_LIKELY(pReq))
1683 {
1684 pReq->pvUser = pvUser;
1685 pReq->cbXfer = cbXfer;
1686 pReq->rcReq = VINF_SUCCESS;
1687 pReq->cXfersPending = 0;
1688 }
1689
1690 return pReq;
1691}
1692
1693static void pdmBlkCacheReqComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq)
1694{
1695 switch (pBlkCache->enmType)
1696 {
1697 case PDMBLKCACHETYPE_DEV:
1698 {
1699 pBlkCache->u.Dev.pfnXferComplete(pBlkCache->u.Dev.pDevIns,
1700 pReq->pvUser, pReq->rcReq);
1701 break;
1702 }
1703 case PDMBLKCACHETYPE_DRV:
1704 {
1705 pBlkCache->u.Drv.pfnXferComplete(pBlkCache->u.Drv.pDrvIns,
1706 pReq->pvUser, pReq->rcReq);
1707 break;
1708 }
1709 case PDMBLKCACHETYPE_USB:
1710 {
1711 pBlkCache->u.Usb.pfnXferComplete(pBlkCache->u.Usb.pUsbIns,
1712 pReq->pvUser, pReq->rcReq);
1713 break;
1714 }
1715 case PDMBLKCACHETYPE_INTERNAL:
1716 {
1717 pBlkCache->u.Int.pfnXferComplete(pBlkCache->u.Int.pvUser,
1718 pReq->pvUser, pReq->rcReq);
1719 break;
1720 }
1721 default:
1722 AssertMsgFailed(("Unknown block cache type!\n"));
1723 }
1724
1725 RTMemFree(pReq);
1726}
1727
1728static bool pdmBlkCacheReqUpdate(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq,
1729 size_t cbComplete, int rcReq, bool fCallHandler)
1730{
1731 if (RT_FAILURE(rcReq))
1732 ASMAtomicCmpXchgS32(&pReq->rcReq, rcReq, VINF_SUCCESS);
1733
1734 AssertMsg(pReq->cbXfer >= cbComplete, ("Completed more than left\n"));
1735 AssertMsg(pReq->cXfersPending > 0, ("No transfers are pending for this request\n"));
1736 uint32_t cbOld = ASMAtomicSubU32(&pReq->cbXfer, cbComplete);
1737 uint32_t cXfersPending = ASMAtomicDecU32(&pReq->cXfersPending);
1738
1739 if ( !(cbOld - cbComplete)
1740 && !cXfersPending)
1741 {
1742 if (fCallHandler)
1743 pdmBlkCacheReqComplete(pBlkCache, pReq);
1744 return true;
1745 }
1746
1747 LogFlowFunc(("pReq=%#p cXfersPending=%u cbXfer=%u\n", pReq, cXfersPending, (cbOld - cbComplete)));
1748 return false;
1749}
1750
1751VMMR3DECL(int) PDMR3BlkCacheRead(PPDMBLKCACHE pBlkCache, uint64_t off,
1752 PCRTSGBUF pcSgBuf, size_t cbRead, void *pvUser)
1753{
1754 int rc = VINF_SUCCESS;
1755 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1756 PPDMBLKCACHEENTRY pEntry;
1757 PPDMBLKCACHEREQ pReq;
1758
1759 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pcSgBuf=%#p cbRead=%u pvUser=%#p\n",
1760 pBlkCache, pBlkCache->pszId, off, pcSgBuf, cbRead, pvUser));
1761
1762 RTSGBUF SgBuf;
1763 RTSgBufClone(&SgBuf, pcSgBuf);
1764
1765 /* Allocate new request structure. */
1766 pReq = pdmBlkCacheReqAlloc(cbRead, pvUser);
1767 if (RT_UNLIKELY(!pReq))
1768 return VERR_NO_MEMORY;
1769
1770 /* Increment data transfer counter to keep the request valid while we access it. */
1771 ASMAtomicIncU32(&pReq->cXfersPending);
1772
1773 while (cbRead)
1774 {
1775 size_t cbToRead;
1776
1777 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
1778
1779 /*
1780 * If there is no entry we try to create a new one eviciting unused pages
1781 * if the cache is full. If this is not possible we will pass the request through
1782 * and skip the caching (all entries may be still in progress so they can't
1783 * be evicted)
1784 * If we have an entry it can be in one of the LRU lists where the entry
1785 * contains data (recently used or frequently used LRU) so we can just read
1786 * the data we need and put the entry at the head of the frequently used LRU list.
1787 * In case the entry is in one of the ghost lists it doesn't contain any data.
1788 * We have to fetch it again evicting pages from either T1 or T2 to make room.
1789 */
1790 if (pEntry)
1791 {
1792 uint64_t offDiff = off - pEntry->Core.Key;
1793
1794 AssertMsg(off >= pEntry->Core.Key,
1795 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
1796 off, pEntry->Core.Key));
1797
1798 AssertPtr(pEntry->pList);
1799
1800 cbToRead = RT_MIN(pEntry->cbData - offDiff, cbRead);
1801
1802 AssertMsg(off + cbToRead <= pEntry->Core.Key + pEntry->Core.KeyLast + 1,
1803 ("Buffer of cache entry exceeded off=%llu cbToRead=%d\n",
1804 off, cbToRead));
1805
1806 cbRead -= cbToRead;
1807
1808 if (!cbRead)
1809 STAM_COUNTER_INC(&pCache->cHits);
1810 else
1811 STAM_COUNTER_INC(&pCache->cPartialHits);
1812
1813 STAM_COUNTER_ADD(&pCache->StatRead, cbToRead);
1814
1815 /* Ghost lists contain no data. */
1816 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
1817 || (pEntry->pList == &pCache->LruFrequentlyUsed))
1818 {
1819 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
1820 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
1821 PDMBLKCACHE_ENTRY_IS_DIRTY))
1822 {
1823 /* Entry didn't completed yet. Append to the list */
1824 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
1825 &SgBuf, offDiff, cbToRead,
1826 false /* fWrite */);
1827 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1828 }
1829 else
1830 {
1831 /* Read as much as we can from the entry. */
1832 RTSgBufCopyFromBuf(&SgBuf, pEntry->pbData + offDiff, cbToRead);
1833 ASMAtomicSubU32(&pReq->cbXfer, cbToRead);
1834 }
1835
1836 /* Move this entry to the top position */
1837 if (pEntry->pList == &pCache->LruFrequentlyUsed)
1838 {
1839 pdmBlkCacheLockEnter(pCache);
1840 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
1841 pdmBlkCacheLockLeave(pCache);
1842 }
1843 /* Release the entry */
1844 pdmBlkCacheEntryRelease(pEntry);
1845 }
1846 else
1847 {
1848 uint8_t *pbBuffer = NULL;
1849
1850 LogFlow(("Fetching data for ghost entry %#p from file\n", pEntry));
1851
1852 pdmBlkCacheLockEnter(pCache);
1853 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
1854 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
1855
1856 /* Move the entry to Am and fetch it to the cache. */
1857 if (fEnough)
1858 {
1859 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
1860 pdmBlkCacheAdd(pCache, pEntry->cbData);
1861 pdmBlkCacheLockLeave(pCache);
1862
1863 if (pbBuffer)
1864 pEntry->pbData = pbBuffer;
1865 else
1866 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
1867 AssertPtr(pEntry->pbData);
1868
1869 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
1870 &SgBuf, offDiff, cbToRead,
1871 false /* fWrite */);
1872 pdmBlkCacheEntryReadFromMedium(pEntry);
1873 /* Release the entry */
1874 pdmBlkCacheEntryRelease(pEntry);
1875 }
1876 else
1877 {
1878 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1879 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
1880 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
1881 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
1882 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1883
1884 pdmBlkCacheLockLeave(pCache);
1885
1886 RTMemFree(pEntry);
1887
1888 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
1889 &SgBuf, off, cbToRead,
1890 PDMBLKCACHEXFERDIR_READ);
1891 }
1892 }
1893 }
1894 else
1895 {
1896#ifdef VBOX_WITH_IO_READ_CACHE
1897 /* No entry found for this offset. Create a new entry and fetch the data to the cache. */
1898 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
1899 off, cbRead,
1900 PAGE_SIZE,
1901 &cbToRead);
1902
1903 cbRead -= cbToRead;
1904
1905 if (pEntryNew)
1906 {
1907 if (!cbRead)
1908 STAM_COUNTER_INC(&pCache->cMisses);
1909 else
1910 STAM_COUNTER_INC(&pCache->cPartialHits);
1911
1912 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
1913 &SgBuf,
1914 off - pEntryNew->Core.Key,
1915 cbToRead,
1916 false /* fWrite */);
1917 pdmBlkCacheEntryReadFromMedium(pEntryNew);
1918 pdmBlkCacheEntryRelease(pEntryNew); /* it is protected by the I/O in progress flag now. */
1919 }
1920 else
1921 {
1922 /*
1923 * There is not enough free space in the cache.
1924 * Pass the request directly to the I/O manager.
1925 */
1926 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToRead));
1927
1928 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
1929 &SgBuf, off, cbToRead,
1930 PDMBLKCACHEXFERDIR_READ);
1931 }
1932#else
1933 /* Clip read size if necessary. */
1934 PPDMBLKCACHEENTRY pEntryAbove;
1935 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off,
1936 &pEntryAbove, NULL);
1937
1938 if (pEntryAbove)
1939 {
1940 if (off + cbRead > pEntryAbove->Core.Key)
1941 cbToRead = pEntryAbove->Core.Key - off;
1942 else
1943 cbToRead = cbRead;
1944
1945 pdmBlkCacheEntryRelease(pEntryAbove);
1946 }
1947 else
1948 cbToRead = cbRead;
1949
1950 cbRead -= cbToRead;
1951 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
1952 &SgBuf, off, cbToRead,
1953 PDMBLKCACHEXFERDIR_READ);
1954#endif
1955 }
1956 off += cbToRead;
1957 }
1958
1959 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, 0, rc, false))
1960 rc = VINF_AIO_TASK_PENDING;
1961
1962 LogFlowFunc((": Leave rc=%Rrc\n", rc));
1963
1964 return rc;
1965}
1966
1967VMMR3DECL(int) PDMR3BlkCacheWrite(PPDMBLKCACHE pBlkCache, uint64_t off,
1968 PCRTSGBUF pcSgBuf, size_t cbWrite, void *pvUser)
1969{
1970 int rc = VINF_SUCCESS;
1971 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1972 PPDMBLKCACHEENTRY pEntry;
1973 PPDMBLKCACHEREQ pReq;
1974
1975 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pcSgBuf=%#p cbWrite=%u pvUser=%#p\n",
1976 pBlkCache, pBlkCache->pszId, off, pcSgBuf, cbWrite, pvUser));
1977
1978 RTSGBUF SgBuf;
1979 RTSgBufClone(&SgBuf, pcSgBuf);
1980
1981 /* Allocate new request structure. */
1982 pReq = pdmBlkCacheReqAlloc(cbWrite, pvUser);
1983 if (RT_UNLIKELY(!pReq))
1984 return VERR_NO_MEMORY;
1985
1986 /* Increment data transfer counter to keep the request valid while we access it. */
1987 ASMAtomicIncU32(&pReq->cXfersPending);
1988
1989 while (cbWrite)
1990 {
1991 size_t cbToWrite;
1992
1993 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
1994
1995 if (pEntry)
1996 {
1997 /* Write the data into the entry and mark it as dirty */
1998 AssertPtr(pEntry->pList);
1999
2000 uint64_t offDiff = off - pEntry->Core.Key;
2001
2002 AssertMsg(off >= pEntry->Core.Key,
2003 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
2004 off, pEntry->Core.Key));
2005
2006 cbToWrite = RT_MIN(pEntry->cbData - offDiff, cbWrite);
2007 cbWrite -= cbToWrite;
2008
2009 if (!cbWrite)
2010 STAM_COUNTER_INC(&pCache->cHits);
2011 else
2012 STAM_COUNTER_INC(&pCache->cPartialHits);
2013
2014 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2015
2016 /* Ghost lists contain no data. */
2017 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2018 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2019 {
2020 /* Check if the entry is dirty. */
2021 if(pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2022 PDMBLKCACHE_ENTRY_IS_DIRTY,
2023 0))
2024 {
2025 /* If it is already dirty but not in progress just update the data. */
2026 if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS))
2027 {
2028 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff,
2029 cbToWrite);
2030 ASMAtomicSubU32(&pReq->cbXfer, cbToWrite);
2031 }
2032 else
2033 {
2034 /* The data isn't written to the file yet */
2035 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2036 &SgBuf, offDiff, cbToWrite,
2037 true /* fWrite */);
2038 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2039 }
2040
2041 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2042 }
2043 else /* Dirty bit not set */
2044 {
2045 /*
2046 * Check if a read is in progress for this entry.
2047 * We have to defer processing in that case.
2048 */
2049 if(pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2050 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2051 0))
2052 {
2053 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2054 &SgBuf, offDiff, cbToWrite,
2055 true /* fWrite */);
2056 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2057 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2058 }
2059 else /* I/O in progress flag not set */
2060 {
2061 /* Write as much as we can into the entry and update the file. */
2062 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff, cbToWrite);
2063 ASMAtomicSubU32(&pReq->cbXfer, cbToWrite);
2064
2065 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2066 if (fCommit)
2067 pdmBlkCacheCommitDirtyEntries(pCache);
2068 }
2069 } /* Dirty bit not set */
2070
2071 /* Move this entry to the top position */
2072 if (pEntry->pList == &pCache->LruFrequentlyUsed)
2073 {
2074 pdmBlkCacheLockEnter(pCache);
2075 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2076 pdmBlkCacheLockLeave(pCache);
2077 }
2078
2079 pdmBlkCacheEntryRelease(pEntry);
2080 }
2081 else /* Entry is on the ghost list */
2082 {
2083 uint8_t *pbBuffer = NULL;
2084
2085 pdmBlkCacheLockEnter(pCache);
2086 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2087 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2088
2089 if (fEnough)
2090 {
2091 /* Move the entry to Am and fetch it to the cache. */
2092 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2093 pdmBlkCacheAdd(pCache, pEntry->cbData);
2094 pdmBlkCacheLockLeave(pCache);
2095
2096 if (pbBuffer)
2097 pEntry->pbData = pbBuffer;
2098 else
2099 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2100 AssertPtr(pEntry->pbData);
2101
2102 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2103 &SgBuf, offDiff, cbToWrite,
2104 true /* fWrite */);
2105 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2106 pdmBlkCacheEntryReadFromMedium(pEntry);
2107
2108 /* Release the reference. If it is still needed the I/O in progress flag should protect it now. */
2109 pdmBlkCacheEntryRelease(pEntry);
2110 }
2111 else
2112 {
2113 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2114 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2115 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2116 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2117 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2118
2119 pdmBlkCacheLockLeave(pCache);
2120
2121 RTMemFree(pEntry);
2122 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2123 &SgBuf, off, cbToWrite,
2124 PDMBLKCACHEXFERDIR_WRITE);
2125 }
2126 }
2127 }
2128 else /* No entry found */
2129 {
2130 /*
2131 * No entry found. Try to create a new cache entry to store the data in and if that fails
2132 * write directly to the file.
2133 */
2134 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
2135 off, cbWrite,
2136 512, &cbToWrite);
2137
2138 cbWrite -= cbToWrite;
2139
2140 if (pEntryNew)
2141 {
2142 uint64_t offDiff = off - pEntryNew->Core.Key;
2143
2144 STAM_COUNTER_INC(&pCache->cHits);
2145
2146 /*
2147 * Check if it is possible to just write the data without waiting
2148 * for it to get fetched first.
2149 */
2150 if (!offDiff && pEntryNew->cbData == cbToWrite)
2151 {
2152 RTSgBufCopyToBuf(&SgBuf, pEntryNew->pbData, cbToWrite);
2153 ASMAtomicSubU32(&pReq->cbXfer, cbToWrite);
2154
2155 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntryNew);
2156 if (fCommit)
2157 pdmBlkCacheCommitDirtyEntries(pCache);
2158 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2159 }
2160 else
2161 {
2162 /* Defer the write and fetch the data from the endpoint. */
2163 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
2164 &SgBuf, offDiff, cbToWrite,
2165 true /* fWrite */);
2166 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2167 pdmBlkCacheEntryReadFromMedium(pEntryNew);
2168 }
2169
2170 pdmBlkCacheEntryRelease(pEntryNew);
2171 }
2172 else
2173 {
2174 /*
2175 * There is not enough free space in the cache.
2176 * Pass the request directly to the I/O manager.
2177 */
2178 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToWrite));
2179
2180 STAM_COUNTER_INC(&pCache->cMisses);
2181
2182 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2183 &SgBuf, off, cbToWrite,
2184 PDMBLKCACHEXFERDIR_WRITE);
2185 }
2186 }
2187
2188 off += cbToWrite;
2189 }
2190
2191 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, 0, rc, false))
2192 rc = VINF_AIO_TASK_PENDING;
2193
2194 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2195
2196 return rc;
2197}
2198
2199VMMR3DECL(int) PDMR3BlkCacheFlush(PPDMBLKCACHE pBlkCache, void *pvUser)
2200{
2201 int rc = VINF_SUCCESS;
2202 PPDMBLKCACHEREQ pReq;
2203
2204 LogFlowFunc((": pBlkCache=%#p{%s}\n", pBlkCache, pBlkCache->pszId));
2205
2206 /* Commit dirty entries in the cache. */
2207 pdmBlkCacheCommit(pBlkCache);
2208
2209 /* Allocate new request structure. */
2210 pReq = pdmBlkCacheReqAlloc(0, pvUser);
2211 if (RT_UNLIKELY(!pReq))
2212 return VERR_NO_MEMORY;
2213
2214 rc = pdmBlkCacheRequestPassthrough(pBlkCache, pReq, NULL, 0, 0,
2215 PDMBLKCACHEXFERDIR_FLUSH);
2216 AssertRC(rc);
2217
2218 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2219 return VINF_AIO_TASK_PENDING;
2220}
2221
2222/**
2223 * Completes a task segment freeing all resources and completes the task handle
2224 * if everything was transferred.
2225 *
2226 * @returns Next task segment handle.
2227 * @param pTaskSeg Task segment to complete.
2228 * @param rc Status code to set.
2229 */
2230static PPDMBLKCACHEWAITER pdmBlkCacheWaiterComplete(PPDMBLKCACHE pBlkCache,
2231 PPDMBLKCACHEWAITER pWaiter,
2232 int rc)
2233{
2234 PPDMBLKCACHEWAITER pNext = pWaiter->pNext;
2235 PPDMBLKCACHEREQ pReq = pWaiter->pReq;
2236
2237 pdmBlkCacheReqUpdate(pBlkCache, pWaiter->pReq, pWaiter->cbTransfer, rc, true);
2238
2239 RTMemFree(pWaiter);
2240
2241 return pNext;
2242}
2243
2244static void pdmBlkCacheIoXferCompleteEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2245{
2246 PPDMBLKCACHEENTRY pEntry = hIoXfer->pEntry;
2247 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2248
2249 /* Reference the entry now as we are clearing the I/O in progress flag
2250 * which protected the entry till now. */
2251 pdmBlkCacheEntryRef(pEntry);
2252
2253 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2254 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
2255
2256 /* Process waiting segment list. The data in entry might have changed in-between. */
2257 bool fDirty = false;
2258 PPDMBLKCACHEWAITER pComplete = pEntry->pWaitingHead;
2259 PPDMBLKCACHEWAITER pCurr = pComplete;
2260
2261 AssertMsg((pCurr && pEntry->pWaitingTail) || (!pCurr && !pEntry->pWaitingTail),
2262 ("The list tail was not updated correctly\n"));
2263 pEntry->pWaitingTail = NULL;
2264 pEntry->pWaitingHead = NULL;
2265
2266 if (hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_WRITE)
2267 {
2268 /*
2269 * An error here is difficult to handle as the original request completed already.
2270 * The error is logged for now and the VM is paused.
2271 * If the user continues the entry is written again in the hope
2272 * the user fixed the problem and the next write succeeds.
2273 */
2274 /** @todo r=aeichner: This solution doesn't work
2275 * The user will get the message but the VM will hang afterwards
2276 * VMR3Suspend() returns when the VM is suspended but suspending
2277 * the VM will reopen the images readonly in DrvVD. They are closed first
2278 * which will close the endpoints. This will block EMT while the
2279 * I/O manager processes the close request but the IO manager is stuck
2280 * in the VMR3Suspend call and can't process the request.
2281 * Another problem is that closing the VM means flushing the cache
2282 * but the entry failed and will probably fail again.
2283 * No idea so far how to solve this problem... but the user gets informed
2284 * at least.
2285 */
2286 if (RT_FAILURE(rcIoXfer))
2287 {
2288 LogRel(("I/O cache: Error while writing entry at offset %llu (%u bytes) to medium \"%s\"\n",
2289 pEntry->Core.Key, pEntry->cbData, pBlkCache->pszId));
2290
2291 if (!ASMAtomicXchgBool(&pCache->fIoErrorVmSuspended, true))
2292 {
2293 int rc = VMSetRuntimeError(pCache->pVM, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "BLKCACHE_IOERR",
2294 N_("The I/O cache encountered an error while updating data in medium \"%s\" (rc=%Rrc). "
2295 "Make sure there is enough free space on the disk and that the disk is working properly. "
2296 "Operation can be resumed afterwards"),
2297 pBlkCache->pszId, rcIoXfer);
2298 AssertRC(rc);
2299 }
2300
2301 /*
2302 * The entry is still marked as dirty which prevents eviction.
2303 * Add the waiters to the list again.
2304 */
2305 fDirty = true;
2306
2307 if (pComplete)
2308 {
2309 pEntry->pWaitingHead = pComplete;
2310 while (pComplete->pNext)
2311 pComplete = pComplete->pNext;
2312 pEntry->pWaitingTail = pComplete;
2313 pComplete = NULL;
2314 }
2315 }
2316 else
2317 {
2318 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IS_DIRTY;
2319
2320 while (pCurr)
2321 {
2322 AssertMsg(pCurr->fWrite, ("Completed write entries should never have read tasks attached\n"));
2323
2324 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2325 fDirty = true;
2326
2327 pCurr = pCurr->pNext;
2328 }
2329 }
2330 }
2331 else
2332 {
2333 AssertMsg(hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_READ, ("Invalid transfer type\n"));
2334 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY),
2335 ("Invalid flags set\n"));
2336
2337 while (pCurr)
2338 {
2339 if (pCurr->fWrite)
2340 {
2341 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2342 fDirty = true;
2343 }
2344 else
2345 RTSgBufCopyFromBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2346
2347 pCurr = pCurr->pNext;
2348 }
2349 }
2350
2351 bool fCommit = false;
2352 if (fDirty)
2353 fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2354
2355 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2356
2357 /* Dereference so that it isn't protected anymore except we issued anyother write for it. */
2358 pdmBlkCacheEntryRelease(pEntry);
2359
2360 if (fCommit)
2361 pdmBlkCacheCommitDirtyEntries(pCache);
2362
2363 /* Complete waiters now. */
2364 while (pComplete)
2365 pComplete = pdmBlkCacheWaiterComplete(pBlkCache, pComplete, rcIoXfer);
2366}
2367
2368VMMR3DECL(void) PDMR3BlkCacheIoXferComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2369{
2370 LogFlowFunc(("pBlkCache=%#p hIoXfer=%#p rcIoXfer=%Rrc\n", pBlkCache, hIoXfer, rcIoXfer));
2371
2372 if (hIoXfer->fIoCache)
2373 pdmBlkCacheIoXferCompleteEntry(pBlkCache, hIoXfer, rcIoXfer);
2374 else
2375 pdmBlkCacheReqUpdate(pBlkCache, hIoXfer->pReq, hIoXfer->cbXfer, rcIoXfer, true);
2376}
2377
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