VirtualBox

source: vbox/trunk/src/VBox/VMM/VMMR3/PDMBlkCache.cpp@ 37423

Last change on this file since 37423 was 37330, checked in by vboxsync, 14 years ago

PDMBlkCache: Fix loading a saved state which has fewer saved entries than the current VM. This should not result in an error.
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 91.2 KB
Line 
1/* $Id: PDMBlkCache.cpp 37330 2011-06-06 16:15:52Z 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/vmm/stam.h>
33#include <VBox/vmm/uvm.h>
34#include <VBox/vmm/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#define PDM_BLK_CACHE_SAVED_STATE_VERSION 1
67
68/*******************************************************************************
69* Internal Functions *
70*******************************************************************************/
71
72static PPDMBLKCACHEENTRY pdmBlkCacheEntryAlloc(PPDMBLKCACHE pBlkCache,
73 uint64_t off, size_t cbData, uint8_t *pbBuffer);
74static bool pdmBlkCacheAddDirtyEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry);
75
76/**
77 * Decrement the reference counter of the given cache entry.
78 *
79 * @returns nothing.
80 * @param pEntry The entry to release.
81 */
82DECLINLINE(void) pdmBlkCacheEntryRelease(PPDMBLKCACHEENTRY pEntry)
83{
84 AssertMsg(pEntry->cRefs > 0, ("Trying to release a not referenced entry\n"));
85 ASMAtomicDecU32(&pEntry->cRefs);
86}
87
88/**
89 * Increment the reference counter of the given cache entry.
90 *
91 * @returns nothing.
92 * @param pEntry The entry to reference.
93 */
94DECLINLINE(void) pdmBlkCacheEntryRef(PPDMBLKCACHEENTRY pEntry)
95{
96 ASMAtomicIncU32(&pEntry->cRefs);
97}
98
99#ifdef DEBUG
100static void pdmBlkCacheValidate(PPDMBLKCACHEGLOBAL pCache)
101{
102 /* Amount of cached data should never exceed the maximum amount. */
103 AssertMsg(pCache->cbCached <= pCache->cbMax,
104 ("Current amount of cached data exceeds maximum\n"));
105
106 /* The amount of cached data in the LRU and FRU list should match cbCached */
107 AssertMsg(pCache->LruRecentlyUsedIn.cbCached + pCache->LruFrequentlyUsed.cbCached == pCache->cbCached,
108 ("Amount of cached data doesn't match\n"));
109
110 AssertMsg(pCache->LruRecentlyUsedOut.cbCached <= pCache->cbRecentlyUsedOutMax,
111 ("Paged out list exceeds maximum\n"));
112}
113#endif
114
115DECLINLINE(void) pdmBlkCacheLockEnter(PPDMBLKCACHEGLOBAL pCache)
116{
117 RTCritSectEnter(&pCache->CritSect);
118#ifdef DEBUG
119 pdmBlkCacheValidate(pCache);
120#endif
121}
122
123DECLINLINE(void) pdmBlkCacheLockLeave(PPDMBLKCACHEGLOBAL pCache)
124{
125#ifdef DEBUG
126 pdmBlkCacheValidate(pCache);
127#endif
128 RTCritSectLeave(&pCache->CritSect);
129}
130
131DECLINLINE(void) pdmBlkCacheSub(PPDMBLKCACHEGLOBAL pCache, uint32_t cbAmount)
132{
133 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
134 pCache->cbCached -= cbAmount;
135}
136
137DECLINLINE(void) pdmBlkCacheAdd(PPDMBLKCACHEGLOBAL pCache, uint32_t cbAmount)
138{
139 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
140 pCache->cbCached += cbAmount;
141}
142
143DECLINLINE(void) pdmBlkCacheListAdd(PPDMBLKLRULIST pList, uint32_t cbAmount)
144{
145 pList->cbCached += cbAmount;
146}
147
148DECLINLINE(void) pdmBlkCacheListSub(PPDMBLKLRULIST pList, uint32_t cbAmount)
149{
150 pList->cbCached -= cbAmount;
151}
152
153#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
154/**
155 * Checks consistency of a LRU list.
156 *
157 * @returns nothing
158 * @param pList The LRU list to check.
159 * @param pNotInList Element which is not allowed to occur in the list.
160 */
161static void pdmBlkCacheCheckList(PPDMBLKLRULIST pList, PPDMBLKCACHEENTRY pNotInList)
162{
163 PPDMBLKCACHEENTRY pCurr = pList->pHead;
164
165 /* Check that there are no double entries and no cycles in the list. */
166 while (pCurr)
167 {
168 PPDMBLKCACHEENTRY pNext = pCurr->pNext;
169
170 while (pNext)
171 {
172 AssertMsg(pCurr != pNext,
173 ("Entry %#p is at least two times in list %#p or there is a cycle in the list\n",
174 pCurr, pList));
175 pNext = pNext->pNext;
176 }
177
178 AssertMsg(pCurr != pNotInList, ("Not allowed entry %#p is in list\n", pCurr));
179
180 if (!pCurr->pNext)
181 AssertMsg(pCurr == pList->pTail, ("End of list reached but last element is not list tail\n"));
182
183 pCurr = pCurr->pNext;
184 }
185}
186#endif
187
188/**
189 * Unlinks a cache entry from the LRU list it is assigned to.
190 *
191 * @returns nothing.
192 * @param pEntry The entry to unlink.
193 */
194static void pdmBlkCacheEntryRemoveFromList(PPDMBLKCACHEENTRY pEntry)
195{
196 PPDMBLKLRULIST pList = pEntry->pList;
197 PPDMBLKCACHEENTRY pPrev, pNext;
198
199 LogFlowFunc((": Deleting entry %#p from list %#p\n", pEntry, pList));
200
201 AssertPtr(pList);
202
203#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
204 pdmBlkCacheCheckList(pList, NULL);
205#endif
206
207 pPrev = pEntry->pPrev;
208 pNext = pEntry->pNext;
209
210 AssertMsg(pEntry != pPrev, ("Entry links to itself as previous element\n"));
211 AssertMsg(pEntry != pNext, ("Entry links to itself as next element\n"));
212
213 if (pPrev)
214 pPrev->pNext = pNext;
215 else
216 {
217 pList->pHead = pNext;
218
219 if (pNext)
220 pNext->pPrev = NULL;
221 }
222
223 if (pNext)
224 pNext->pPrev = pPrev;
225 else
226 {
227 pList->pTail = pPrev;
228
229 if (pPrev)
230 pPrev->pNext = NULL;
231 }
232
233 pEntry->pList = NULL;
234 pEntry->pPrev = NULL;
235 pEntry->pNext = NULL;
236 pdmBlkCacheListSub(pList, pEntry->cbData);
237#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
238 pdmBlkCacheCheckList(pList, pEntry);
239#endif
240}
241
242/**
243 * Adds a cache entry to the given LRU list unlinking it from the currently
244 * assigned list if needed.
245 *
246 * @returns nothing.
247 * @param pList List to the add entry to.
248 * @param pEntry Entry to add.
249 */
250static void pdmBlkCacheEntryAddToList(PPDMBLKLRULIST pList, PPDMBLKCACHEENTRY pEntry)
251{
252 LogFlowFunc((": Adding entry %#p to list %#p\n", pEntry, pList));
253#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
254 pdmBlkCacheCheckList(pList, NULL);
255#endif
256
257 /* Remove from old list if needed */
258 if (pEntry->pList)
259 pdmBlkCacheEntryRemoveFromList(pEntry);
260
261 pEntry->pNext = pList->pHead;
262 if (pList->pHead)
263 pList->pHead->pPrev = pEntry;
264 else
265 {
266 Assert(!pList->pTail);
267 pList->pTail = pEntry;
268 }
269
270 pEntry->pPrev = NULL;
271 pList->pHead = pEntry;
272 pdmBlkCacheListAdd(pList, pEntry->cbData);
273 pEntry->pList = pList;
274#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
275 pdmBlkCacheCheckList(pList, NULL);
276#endif
277}
278
279/**
280 * Destroys a LRU list freeing all entries.
281 *
282 * @returns nothing
283 * @param pList Pointer to the LRU list to destroy.
284 *
285 * @note The caller must own the critical section of the cache.
286 */
287static void pdmBlkCacheDestroyList(PPDMBLKLRULIST pList)
288{
289 while (pList->pHead)
290 {
291 PPDMBLKCACHEENTRY pEntry = pList->pHead;
292
293 pList->pHead = pEntry->pNext;
294
295 AssertMsg(!(pEntry->fFlags & (PDMBLKCACHE_ENTRY_IO_IN_PROGRESS | PDMBLKCACHE_ENTRY_IS_DIRTY)),
296 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
297
298 RTMemPageFree(pEntry->pbData, pEntry->cbData);
299 RTMemFree(pEntry);
300 }
301}
302
303/**
304 * Tries to remove the given amount of bytes from a given list in the cache
305 * moving the entries to one of the given ghosts lists
306 *
307 * @returns Amount of data which could be freed.
308 * @param pCache Pointer to the global cache data.
309 * @param cbData The amount of the data to free.
310 * @param pListSrc The source list to evict data from.
311 * @param pGhostListSrc The ghost list removed entries should be moved to
312 * NULL if the entry should be freed.
313 * @param fReuseBuffer Flag whether a buffer should be reused if it has the same size
314 * @param ppbBuf Where to store the address of the buffer if an entry with the
315 * same size was found and fReuseBuffer is true.
316 *
317 * @note This function may return fewer bytes than requested because entries
318 * may be marked as non evictable if they are used for I/O at the
319 * moment.
320 */
321static size_t pdmBlkCacheEvictPagesFrom(PPDMBLKCACHEGLOBAL pCache, size_t cbData,
322 PPDMBLKLRULIST pListSrc, PPDMBLKLRULIST pGhostListDst,
323 bool fReuseBuffer, uint8_t **ppbBuffer)
324{
325 size_t cbEvicted = 0;
326
327 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
328
329 AssertMsg(cbData > 0, ("Evicting 0 bytes not possible\n"));
330 AssertMsg( !pGhostListDst
331 || (pGhostListDst == &pCache->LruRecentlyUsedOut),
332 ("Destination list must be NULL or the recently used but paged out list\n"));
333
334 if (fReuseBuffer)
335 {
336 AssertPtr(ppbBuffer);
337 *ppbBuffer = NULL;
338 }
339
340 /* Start deleting from the tail. */
341 PPDMBLKCACHEENTRY pEntry = pListSrc->pTail;
342
343 while ((cbEvicted < cbData) && pEntry)
344 {
345 PPDMBLKCACHEENTRY pCurr = pEntry;
346
347 pEntry = pEntry->pPrev;
348
349 /* We can't evict pages which are currently in progress or dirty but not in progress */
350 if ( !(pCurr->fFlags & PDMBLKCACHE_NOT_EVICTABLE)
351 && (ASMAtomicReadU32(&pCurr->cRefs) == 0))
352 {
353 /* Ok eviction candidate. Grab the endpoint semaphore and check again
354 * because somebody else might have raced us. */
355 PPDMBLKCACHE pBlkCache = pCurr->pBlkCache;
356 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
357
358 if (!(pCurr->fFlags & PDMBLKCACHE_NOT_EVICTABLE)
359 && (ASMAtomicReadU32(&pCurr->cRefs) == 0))
360 {
361 LogFlow(("Evicting entry %#p (%u bytes)\n", pCurr, pCurr->cbData));
362
363 if (fReuseBuffer && (pCurr->cbData == cbData))
364 {
365 STAM_COUNTER_INC(&pCache->StatBuffersReused);
366 *ppbBuffer = pCurr->pbData;
367 }
368 else if (pCurr->pbData)
369 RTMemPageFree(pCurr->pbData, pCurr->cbData);
370
371 pCurr->pbData = NULL;
372 cbEvicted += pCurr->cbData;
373
374 pdmBlkCacheEntryRemoveFromList(pCurr);
375 pdmBlkCacheSub(pCache, pCurr->cbData);
376
377 if (pGhostListDst)
378 {
379 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
380
381 PPDMBLKCACHEENTRY pGhostEntFree = pGhostListDst->pTail;
382
383 /* We have to remove the last entries from the paged out list. */
384 while ( ((pGhostListDst->cbCached + pCurr->cbData) > pCache->cbRecentlyUsedOutMax)
385 && pGhostEntFree)
386 {
387 PPDMBLKCACHEENTRY pFree = pGhostEntFree;
388 PPDMBLKCACHE pBlkCacheFree = pFree->pBlkCache;
389
390 pGhostEntFree = pGhostEntFree->pPrev;
391
392 RTSemRWRequestWrite(pBlkCacheFree->SemRWEntries, RT_INDEFINITE_WAIT);
393
394 if (ASMAtomicReadU32(&pFree->cRefs) == 0)
395 {
396 pdmBlkCacheEntryRemoveFromList(pFree);
397
398 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
399 RTAvlrU64Remove(pBlkCacheFree->pTree, pFree->Core.Key);
400 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
401
402 RTMemFree(pFree);
403 }
404
405 RTSemRWReleaseWrite(pBlkCacheFree->SemRWEntries);
406 }
407
408 if (pGhostListDst->cbCached + pCurr->cbData > pCache->cbRecentlyUsedOutMax)
409 {
410 /* Couldn't remove enough entries. Delete */
411 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
412 RTAvlrU64Remove(pCurr->pBlkCache->pTree, pCurr->Core.Key);
413 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
414
415 RTMemFree(pCurr);
416 }
417 else
418 pdmBlkCacheEntryAddToList(pGhostListDst, pCurr);
419 }
420 else
421 {
422 /* Delete the entry from the AVL tree it is assigned to. */
423 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
424 RTAvlrU64Remove(pCurr->pBlkCache->pTree, pCurr->Core.Key);
425 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
426
427 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
428 RTMemFree(pCurr);
429 }
430 }
431
432 }
433 else
434 LogFlow(("Entry %#p (%u bytes) is still in progress and can't be evicted\n", pCurr, pCurr->cbData));
435 }
436
437 return cbEvicted;
438}
439
440static bool pdmBlkCacheReclaim(PPDMBLKCACHEGLOBAL pCache, size_t cbData, bool fReuseBuffer, uint8_t **ppbBuffer)
441{
442 size_t cbRemoved = 0;
443
444 if ((pCache->cbCached + cbData) < pCache->cbMax)
445 return true;
446 else if ((pCache->LruRecentlyUsedIn.cbCached + cbData) > pCache->cbRecentlyUsedInMax)
447 {
448 /* Try to evict as many bytes as possible from A1in */
449 cbRemoved = pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruRecentlyUsedIn,
450 &pCache->LruRecentlyUsedOut, fReuseBuffer, ppbBuffer);
451
452 /*
453 * If it was not possible to remove enough entries
454 * try the frequently accessed cache.
455 */
456 if (cbRemoved < cbData)
457 {
458 Assert(!fReuseBuffer || !*ppbBuffer); /* It is not possible that we got a buffer with the correct size but we didn't freed enough data. */
459
460 /*
461 * If we removed something we can't pass the reuse buffer flag anymore because
462 * we don't need to evict that much data
463 */
464 if (!cbRemoved)
465 cbRemoved += pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruFrequentlyUsed,
466 NULL, fReuseBuffer, ppbBuffer);
467 else
468 cbRemoved += pdmBlkCacheEvictPagesFrom(pCache, cbData - cbRemoved, &pCache->LruFrequentlyUsed,
469 NULL, false, NULL);
470 }
471 }
472 else
473 {
474 /* We have to remove entries from frequently access list. */
475 cbRemoved = pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruFrequentlyUsed,
476 NULL, fReuseBuffer, ppbBuffer);
477 }
478
479 LogFlowFunc((": removed %u bytes, requested %u\n", cbRemoved, cbData));
480 return (cbRemoved >= cbData);
481}
482
483DECLINLINE(int) pdmBlkCacheEnqueue(PPDMBLKCACHE pBlkCache, uint64_t off, size_t cbXfer, PPDMBLKCACHEIOXFER pIoXfer)
484{
485 int rc = VINF_SUCCESS;
486
487 LogFlowFunc(("%s: Enqueuing hIoXfer=%#p enmXferDir=%d\n",
488 __FUNCTION__, pIoXfer, pIoXfer->enmXferDir));
489
490 switch (pBlkCache->enmType)
491 {
492 case PDMBLKCACHETYPE_DEV:
493 {
494 rc = pBlkCache->u.Dev.pfnXferEnqueue(pBlkCache->u.Dev.pDevIns,
495 pIoXfer->enmXferDir,
496 off, cbXfer,
497 &pIoXfer->SgBuf, pIoXfer);
498 break;
499 }
500 case PDMBLKCACHETYPE_DRV:
501 {
502 rc = pBlkCache->u.Drv.pfnXferEnqueue(pBlkCache->u.Drv.pDrvIns,
503 pIoXfer->enmXferDir,
504 off, cbXfer,
505 &pIoXfer->SgBuf, pIoXfer);
506 break;
507 }
508 case PDMBLKCACHETYPE_USB:
509 {
510 rc = pBlkCache->u.Usb.pfnXferEnqueue(pBlkCache->u.Usb.pUsbIns,
511 pIoXfer->enmXferDir,
512 off, cbXfer,
513 &pIoXfer->SgBuf, pIoXfer);
514 break;
515 }
516 case PDMBLKCACHETYPE_INTERNAL:
517 {
518 rc = pBlkCache->u.Int.pfnXferEnqueue(pBlkCache->u.Int.pvUser,
519 pIoXfer->enmXferDir,
520 off, cbXfer,
521 &pIoXfer->SgBuf, pIoXfer);
522 break;
523 }
524 default:
525 AssertMsgFailed(("Unknown block cache type!\n"));
526 }
527
528 LogFlowFunc(("%s: returns rc=%Rrc\n", __FUNCTION__, rc));
529 return rc;
530}
531
532/**
533 * Initiates a read I/O task for the given entry.
534 *
535 * @returns VBox status code.
536 * @param pEntry The entry to fetch the data to.
537 */
538static int pdmBlkCacheEntryReadFromMedium(PPDMBLKCACHEENTRY pEntry)
539{
540 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
541 LogFlowFunc((": Reading data into cache entry %#p\n", pEntry));
542
543 /* Make sure no one evicts the entry while it is accessed. */
544 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
545
546 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
547 if (RT_UNLIKELY(!pIoXfer))
548 return VERR_NO_MEMORY;
549
550 AssertMsg(pEntry->pbData, ("Entry is in ghost state\n"));
551
552 pIoXfer->fIoCache = true;
553 pIoXfer->pEntry = pEntry;
554 pIoXfer->SgSeg.pvSeg = pEntry->pbData;
555 pIoXfer->SgSeg.cbSeg = pEntry->cbData;
556 pIoXfer->enmXferDir = PDMBLKCACHEXFERDIR_READ;
557 RTSgBufInit(&pIoXfer->SgBuf, &pIoXfer->SgSeg, 1);
558
559 return pdmBlkCacheEnqueue(pBlkCache, pEntry->Core.Key, pEntry->cbData, pIoXfer);
560}
561
562/**
563 * Initiates a write I/O task for the given entry.
564 *
565 * @returns nothing.
566 * @param pEntry The entry to read the data from.
567 */
568static int pdmBlkCacheEntryWriteToMedium(PPDMBLKCACHEENTRY pEntry)
569{
570 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
571 LogFlowFunc((": Writing data from cache entry %#p\n", pEntry));
572
573 /* Make sure no one evicts the entry while it is accessed. */
574 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
575
576 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
577 if (RT_UNLIKELY(!pIoXfer))
578 return VERR_NO_MEMORY;
579
580 AssertMsg(pEntry->pbData, ("Entry is in ghost state\n"));
581
582 pIoXfer->fIoCache = true;
583 pIoXfer->pEntry = pEntry;
584 pIoXfer->SgSeg.pvSeg = pEntry->pbData;
585 pIoXfer->SgSeg.cbSeg = pEntry->cbData;
586 pIoXfer->enmXferDir = PDMBLKCACHEXFERDIR_WRITE;
587 RTSgBufInit(&pIoXfer->SgBuf, &pIoXfer->SgSeg, 1);
588
589 return pdmBlkCacheEnqueue(pBlkCache, pEntry->Core.Key, pEntry->cbData, pIoXfer);
590}
591
592/**
593 * Passthrough a part of a request directly to the I/O manager
594 * handling the endpoint.
595 *
596 * @returns VBox status code.
597 * @param pEndpoint The endpoint.
598 * @param pTask The task.
599 * @param pIoMemCtx The I/O memory context to use.
600 * @param offStart Offset to start transfer from.
601 * @param cbData Amount of data to transfer.
602 * @param enmTransferType The transfer type (read/write)
603 */
604static int pdmBlkCacheRequestPassthrough(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq,
605 PRTSGBUF pSgBuf, uint64_t offStart, size_t cbData,
606 PDMBLKCACHEXFERDIR enmXferDir)
607{
608
609 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
610 if (RT_UNLIKELY(!pIoXfer))
611 return VERR_NO_MEMORY;
612
613 ASMAtomicIncU32(&pReq->cXfersPending);
614 pIoXfer->fIoCache = false;
615 pIoXfer->pReq = pReq;
616 pIoXfer->enmXferDir = enmXferDir;
617 if (pSgBuf)
618 {
619 RTSgBufClone(&pIoXfer->SgBuf, pSgBuf);
620 RTSgBufAdvance(pSgBuf, cbData);
621 }
622
623 return pdmBlkCacheEnqueue(pBlkCache, offStart, cbData, pIoXfer);
624}
625
626/**
627 * Commit a single dirty entry to the endpoint
628 *
629 * @returns nothing
630 * @param pEntry The entry to commit.
631 */
632static void pdmBlkCacheEntryCommit(PPDMBLKCACHEENTRY pEntry)
633{
634 AssertMsg( (pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY)
635 && !(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
636 ("Invalid flags set for entry %#p\n", pEntry));
637
638 pdmBlkCacheEntryWriteToMedium(pEntry);
639}
640
641/**
642 * Commit all dirty entries for a single endpoint.
643 *
644 * @returns nothing.
645 * @param pBlkCache The endpoint cache to commit.
646 */
647static void pdmBlkCacheCommit(PPDMBLKCACHE pBlkCache)
648{
649 uint32_t cbCommitted = 0;
650
651 /* Return if the cache was suspended. */
652 if (pBlkCache->fSuspended)
653 return;
654
655 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
656
657 /* The list is moved to a new header to reduce locking overhead. */
658 RTLISTNODE ListDirtyNotCommitted;
659 RTSPINLOCKTMP Tmp;
660
661 RTListInit(&ListDirtyNotCommitted);
662 RTSpinlockAcquire(pBlkCache->LockList, &Tmp);
663 RTListMove(&ListDirtyNotCommitted, &pBlkCache->ListDirtyNotCommitted);
664 RTSpinlockRelease(pBlkCache->LockList, &Tmp);
665
666 if (!RTListIsEmpty(&ListDirtyNotCommitted))
667 {
668 PPDMBLKCACHEENTRY pEntry = RTListGetFirst(&ListDirtyNotCommitted, PDMBLKCACHEENTRY, NodeNotCommitted);
669
670 while (!RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted))
671 {
672 PPDMBLKCACHEENTRY pNext = RTListNodeGetNext(&pEntry->NodeNotCommitted, PDMBLKCACHEENTRY,
673 NodeNotCommitted);
674 pdmBlkCacheEntryCommit(pEntry);
675 cbCommitted += pEntry->cbData;
676 RTListNodeRemove(&pEntry->NodeNotCommitted);
677 pEntry = pNext;
678 }
679
680 /* Commit the last endpoint */
681 Assert(RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted));
682 pdmBlkCacheEntryCommit(pEntry);
683 RTListNodeRemove(&pEntry->NodeNotCommitted);
684 AssertMsg(RTListIsEmpty(&ListDirtyNotCommitted),
685 ("Committed all entries but list is not empty\n"));
686 }
687
688 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
689 AssertMsg(pBlkCache->pCache->cbDirty >= cbCommitted,
690 ("Number of committed bytes exceeds number of dirty bytes\n"));
691 uint32_t cbDirtyOld = ASMAtomicSubU32(&pBlkCache->pCache->cbDirty, cbCommitted);
692
693 /* Reset the commit timer if we don't have any dirty bits. */
694 if ( !(cbDirtyOld - cbCommitted)
695 && pBlkCache->pCache->u32CommitTimeoutMs != 0)
696 TMTimerStop(pBlkCache->pCache->pTimerCommit);
697}
698
699/**
700 * Commit all dirty entries in the cache.
701 *
702 * @returns nothing.
703 * @param pCache The global cache instance.
704 */
705static void pdmBlkCacheCommitDirtyEntries(PPDMBLKCACHEGLOBAL pCache)
706{
707 bool fCommitInProgress = ASMAtomicXchgBool(&pCache->fCommitInProgress, true);
708
709 if (!fCommitInProgress)
710 {
711 pdmBlkCacheLockEnter(pCache);
712 Assert(!RTListIsEmpty(&pCache->ListUsers));
713
714 PPDMBLKCACHE pBlkCache = RTListGetFirst(&pCache->ListUsers, PDMBLKCACHE, NodeCacheUser);
715 AssertPtr(pBlkCache);
716
717 while (!RTListNodeIsLast(&pCache->ListUsers, &pBlkCache->NodeCacheUser))
718 {
719 pdmBlkCacheCommit(pBlkCache);
720
721 pBlkCache = RTListNodeGetNext(&pBlkCache->NodeCacheUser, PDMBLKCACHE,
722 NodeCacheUser);
723 }
724
725 /* Commit the last endpoint */
726 Assert(RTListNodeIsLast(&pCache->ListUsers, &pBlkCache->NodeCacheUser));
727 pdmBlkCacheCommit(pBlkCache);
728
729 pdmBlkCacheLockLeave(pCache);
730 ASMAtomicWriteBool(&pCache->fCommitInProgress, false);
731 }
732}
733
734/**
735 * Adds the given entry as a dirty to the cache.
736 *
737 * @returns Flag whether the amount of dirty bytes in the cache exceeds the threshold
738 * @param pBlkCache The endpoint cache the entry belongs to.
739 * @param pEntry The entry to add.
740 */
741static bool pdmBlkCacheAddDirtyEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
742{
743 bool fDirtyBytesExceeded = false;
744 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
745
746 /* If the commit timer is disabled we commit right away. */
747 if (pCache->u32CommitTimeoutMs == 0)
748 {
749 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IS_DIRTY;
750 pdmBlkCacheEntryCommit(pEntry);
751 }
752 else if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY))
753 {
754 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IS_DIRTY;
755
756 RTSPINLOCKTMP Tmp;
757 RTSpinlockAcquire(pBlkCache->LockList, &Tmp);
758 RTListAppend(&pBlkCache->ListDirtyNotCommitted, &pEntry->NodeNotCommitted);
759 RTSpinlockRelease(pBlkCache->LockList, &Tmp);
760
761 uint32_t cbDirty = ASMAtomicAddU32(&pCache->cbDirty, pEntry->cbData);
762
763 /* Prevent committing if the VM was suspended. */
764 if (RT_LIKELY(!ASMAtomicReadBool(&pCache->fIoErrorVmSuspended)))
765 fDirtyBytesExceeded = (cbDirty + pEntry->cbData >= pCache->cbCommitDirtyThreshold);
766 else if (!cbDirty && pCache->u32CommitTimeoutMs > 0)
767 {
768 /* Arm the commit timer. */
769 TMTimerSetMillies(pCache->pTimerCommit, pCache->u32CommitTimeoutMs);
770 }
771 }
772
773 return fDirtyBytesExceeded;
774}
775
776static PPDMBLKCACHE pdmR3BlkCacheFindById(PPDMBLKCACHEGLOBAL pBlkCacheGlobal, const char *pcszId)
777{
778 bool fFound = false;
779 PPDMBLKCACHE pBlkCache = NULL;
780
781 RTListForEach(&pBlkCacheGlobal->ListUsers, pBlkCache, PDMBLKCACHE, NodeCacheUser)
782 {
783 if (!RTStrCmp(pBlkCache->pszId, pcszId))
784 {
785 fFound = true;
786 break;
787 }
788 }
789
790 return fFound ? pBlkCache : NULL;
791}
792
793/**
794 * Commit timer callback.
795 */
796static void pdmBlkCacheCommitTimerCallback(PVM pVM, PTMTIMER pTimer, void *pvUser)
797{
798 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
799
800 LogFlowFunc(("Commit interval expired, commiting dirty entries\n"));
801
802 if ( ASMAtomicReadU32(&pCache->cbDirty) > 0
803 && !ASMAtomicReadBool(&pCache->fIoErrorVmSuspended))
804 pdmBlkCacheCommitDirtyEntries(pCache);
805
806 LogFlowFunc(("Entries committed, going to sleep\n"));
807}
808
809static DECLCALLBACK(int) pdmR3BlkCacheSaveExec(PVM pVM, PSSMHANDLE pSSM)
810{
811 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
812
813 AssertPtr(pBlkCacheGlobal);
814
815 pdmBlkCacheLockEnter(pBlkCacheGlobal);
816
817 SSMR3PutU32(pSSM, pBlkCacheGlobal->cRefs);
818
819 /* Go through the list and save all dirty entries. */
820 PPDMBLKCACHE pBlkCache;
821 RTListForEach(&pBlkCacheGlobal->ListUsers, pBlkCache, PDMBLKCACHE, NodeCacheUser)
822 {
823 uint32_t cEntries = 0;
824 PPDMBLKCACHEENTRY pEntry;
825
826 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
827 SSMR3PutU32(pSSM, strlen(pBlkCache->pszId));
828 SSMR3PutStrZ(pSSM, pBlkCache->pszId);
829
830 /* Count the number of entries to safe. */
831 RTListForEach(&pBlkCache->ListDirtyNotCommitted, pEntry, PDMBLKCACHEENTRY, NodeNotCommitted)
832 {
833 cEntries++;
834 }
835
836 SSMR3PutU32(pSSM, cEntries);
837
838 /* Walk the list of all dirty entries and save them. */
839 RTListForEach(&pBlkCache->ListDirtyNotCommitted, pEntry, PDMBLKCACHEENTRY, NodeNotCommitted)
840 {
841 /* A few sanity checks. */
842 AssertMsg(!pEntry->cRefs, ("The entry is still referenced\n"));
843 AssertMsg(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY, ("Entry is not dirty\n"));
844 AssertMsg(!(pEntry->fFlags & ~PDMBLKCACHE_ENTRY_IS_DIRTY), ("Invalid flags set\n"));
845 AssertMsg(!pEntry->pWaitingHead && !pEntry->pWaitingTail, ("There are waiting requests\n"));
846 AssertMsg( pEntry->pList == &pBlkCacheGlobal->LruRecentlyUsedIn
847 || pEntry->pList == &pBlkCacheGlobal->LruFrequentlyUsed,
848 ("Invalid list\n"));
849 AssertMsg(pEntry->cbData == pEntry->Core.KeyLast - pEntry->Core.Key + 1,
850 ("Size and range do not match\n"));
851
852 /* Save */
853 SSMR3PutU64(pSSM, pEntry->Core.Key);
854 SSMR3PutU32(pSSM, pEntry->cbData);
855 SSMR3PutMem(pSSM, pEntry->pbData, pEntry->cbData);
856 }
857
858 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
859 }
860
861 pdmBlkCacheLockLeave(pBlkCacheGlobal);
862
863 /* Terminator */
864 return SSMR3PutU32(pSSM, UINT32_MAX);
865}
866
867static DECLCALLBACK(int) pdmR3BlkCacheLoadExec(PVM pVM, PSSMHANDLE pSSM, uint32_t uVersion, uint32_t uPass)
868{
869 int rc = VINF_SUCCESS;
870 uint32_t cRefs;
871 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
872
873 AssertPtr(pBlkCacheGlobal);
874
875 pdmBlkCacheLockEnter(pBlkCacheGlobal);
876
877 if (uVersion != PDM_BLK_CACHE_SAVED_STATE_VERSION)
878 return VERR_SSM_UNSUPPORTED_DATA_UNIT_VERSION;
879
880 SSMR3GetU32(pSSM, &cRefs);
881
882 /*
883 * Fewer users in the saved state than in the current VM are allowed
884 * because that means that there are only new ones which don't have any saved state
885 * which can get lost.
886 * More saved entries that current ones are not allowed because this could result in
887 * lost data.
888 */
889 if (cRefs <= pBlkCacheGlobal->cRefs)
890 {
891 char *pszId = NULL;
892
893 while ( cRefs > 0
894 && RT_SUCCESS(rc))
895 {
896 PPDMBLKCACHE pBlkCache = NULL;
897 uint32_t cbId = 0;
898
899 SSMR3GetU32(pSSM, &cbId);
900 Assert(cbId > 0);
901
902 cbId++; /* Include terminator */
903 pszId = (char *)RTMemAllocZ(cbId * sizeof(char));
904 if (!pszId)
905 {
906 rc = VERR_NO_MEMORY;
907 break;
908 }
909
910 rc = SSMR3GetStrZ(pSSM, pszId, cbId);
911 AssertRC(rc);
912
913 /* Search for the block cache with the provided id. */
914 pBlkCache = pdmR3BlkCacheFindById(pBlkCacheGlobal, pszId);
915 if (!pBlkCache)
916 {
917 rc = SSMR3SetCfgError(pSSM, RT_SRC_POS,
918 N_("The VM is missing a block device. Please make sure the source and target VMs have compatible storage configurations"));
919 break;
920 }
921
922 RTStrFree(pszId);
923 pszId = NULL;
924
925 /* Get the entries */
926 uint32_t cEntries;
927 SSMR3GetU32(pSSM, &cEntries);
928
929 while (cEntries > 0)
930 {
931 PPDMBLKCACHEENTRY pEntry;
932 uint64_t off;
933 uint32_t cbEntry;
934
935 SSMR3GetU64(pSSM, &off);
936 SSMR3GetU32(pSSM, &cbEntry);
937
938 pEntry = pdmBlkCacheEntryAlloc(pBlkCache, off, cbEntry, NULL);
939 if (!pEntry)
940 {
941 rc = VERR_NO_MEMORY;
942 break;
943 }
944
945 rc = SSMR3GetMem(pSSM, pEntry->pbData, cbEntry);
946 if (RT_FAILURE(rc))
947 {
948 RTMemFree(pEntry->pbData);
949 RTMemFree(pEntry);
950 break;
951 }
952
953 /* Insert into the tree. */
954 bool fInserted = RTAvlrU64Insert(pBlkCache->pTree, &pEntry->Core);
955 Assert(fInserted);
956
957 /* Add to the dirty list. */
958 pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
959 pdmBlkCacheEntryAddToList(&pBlkCacheGlobal->LruRecentlyUsedIn, pEntry);
960 pdmBlkCacheAdd(pBlkCacheGlobal, cbEntry);
961 pdmBlkCacheEntryRelease(pEntry);
962 cEntries--;
963 }
964
965 cRefs--;
966 }
967
968 if (pszId)
969 RTStrFree(pszId);
970 }
971 else
972 rc = SSMR3SetCfgError(pSSM, RT_SRC_POS,
973 N_("The VM is missing a block device. Please make sure the source and target VMs have compatible storage configurations"));
974
975 pdmBlkCacheLockLeave(pBlkCacheGlobal);
976
977 if (RT_SUCCESS(rc))
978 {
979 uint32_t u32 = 0;
980 rc = SSMR3GetU32(pSSM, &u32);
981 if (RT_SUCCESS(rc))
982 AssertMsgReturn(u32 == UINT32_MAX, ("%#x\n", u32), VERR_SSM_DATA_UNIT_FORMAT_CHANGED);
983 }
984
985 return rc;
986}
987
988int pdmR3BlkCacheInit(PVM pVM)
989{
990 int rc = VINF_SUCCESS;
991 PUVM pUVM = pVM->pUVM;
992 PPDMBLKCACHEGLOBAL pBlkCacheGlobal;
993
994 LogFlowFunc((": pVM=%p\n", pVM));
995
996 VM_ASSERT_EMT(pVM);
997
998 PCFGMNODE pCfgRoot = CFGMR3GetRoot(pVM);
999 PCFGMNODE pCfgBlkCache = CFGMR3GetChild(CFGMR3GetChild(pCfgRoot, "PDM"), "BlkCache");
1000
1001 pBlkCacheGlobal = (PPDMBLKCACHEGLOBAL)RTMemAllocZ(sizeof(PDMBLKCACHEGLOBAL));
1002 if (!pBlkCacheGlobal)
1003 return VERR_NO_MEMORY;
1004
1005 RTListInit(&pBlkCacheGlobal->ListUsers);
1006 pBlkCacheGlobal->pVM = pVM;
1007 pBlkCacheGlobal->cRefs = 0;
1008 pBlkCacheGlobal->cbCached = 0;
1009 pBlkCacheGlobal->fCommitInProgress = false;
1010
1011 /* Initialize members */
1012 pBlkCacheGlobal->LruRecentlyUsedIn.pHead = NULL;
1013 pBlkCacheGlobal->LruRecentlyUsedIn.pTail = NULL;
1014 pBlkCacheGlobal->LruRecentlyUsedIn.cbCached = 0;
1015
1016 pBlkCacheGlobal->LruRecentlyUsedOut.pHead = NULL;
1017 pBlkCacheGlobal->LruRecentlyUsedOut.pTail = NULL;
1018 pBlkCacheGlobal->LruRecentlyUsedOut.cbCached = 0;
1019
1020 pBlkCacheGlobal->LruFrequentlyUsed.pHead = NULL;
1021 pBlkCacheGlobal->LruFrequentlyUsed.pTail = NULL;
1022 pBlkCacheGlobal->LruFrequentlyUsed.cbCached = 0;
1023
1024 do
1025 {
1026 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheSize", &pBlkCacheGlobal->cbMax, 5 * _1M);
1027 AssertLogRelRCBreak(rc);
1028 LogFlowFunc(("Maximum number of bytes cached %u\n", pBlkCacheGlobal->cbMax));
1029
1030 pBlkCacheGlobal->cbRecentlyUsedInMax = (pBlkCacheGlobal->cbMax / 100) * 25; /* 25% of the buffer size */
1031 pBlkCacheGlobal->cbRecentlyUsedOutMax = (pBlkCacheGlobal->cbMax / 100) * 50; /* 50% of the buffer size */
1032 LogFlowFunc(("cbRecentlyUsedInMax=%u cbRecentlyUsedOutMax=%u\n",
1033 pBlkCacheGlobal->cbRecentlyUsedInMax, pBlkCacheGlobal->cbRecentlyUsedOutMax));
1034
1035 /** @todo r=aeichner: Experiment to find optimal default values */
1036 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheCommitIntervalMs", &pBlkCacheGlobal->u32CommitTimeoutMs, 10000 /* 10sec */);
1037 AssertLogRelRCBreak(rc);
1038 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheCommitThreshold", &pBlkCacheGlobal->cbCommitDirtyThreshold, pBlkCacheGlobal->cbMax / 2);
1039 AssertLogRelRCBreak(rc);
1040 } while (0);
1041
1042 if (RT_SUCCESS(rc))
1043 {
1044 STAMR3Register(pVM, &pBlkCacheGlobal->cbMax,
1045 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1046 "/PDM/BlkCache/cbMax",
1047 STAMUNIT_BYTES,
1048 "Maximum cache size");
1049 STAMR3Register(pVM, &pBlkCacheGlobal->cbCached,
1050 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1051 "/PDM/BlkCache/cbCached",
1052 STAMUNIT_BYTES,
1053 "Currently used cache");
1054 STAMR3Register(pVM, &pBlkCacheGlobal->LruRecentlyUsedIn.cbCached,
1055 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1056 "/PDM/BlkCache/cbCachedMruIn",
1057 STAMUNIT_BYTES,
1058 "Number of bytes cached in MRU list");
1059 STAMR3Register(pVM, &pBlkCacheGlobal->LruRecentlyUsedOut.cbCached,
1060 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1061 "/PDM/BlkCache/cbCachedMruOut",
1062 STAMUNIT_BYTES,
1063 "Number of bytes cached in FRU list");
1064 STAMR3Register(pVM, &pBlkCacheGlobal->LruFrequentlyUsed.cbCached,
1065 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1066 "/PDM/BlkCache/cbCachedFru",
1067 STAMUNIT_BYTES,
1068 "Number of bytes cached in FRU ghost list");
1069
1070#ifdef VBOX_WITH_STATISTICS
1071 STAMR3Register(pVM, &pBlkCacheGlobal->cHits,
1072 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1073 "/PDM/BlkCache/CacheHits",
1074 STAMUNIT_COUNT, "Number of hits in the cache");
1075 STAMR3Register(pVM, &pBlkCacheGlobal->cPartialHits,
1076 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1077 "/PDM/BlkCache/CachePartialHits",
1078 STAMUNIT_COUNT, "Number of partial hits in the cache");
1079 STAMR3Register(pVM, &pBlkCacheGlobal->cMisses,
1080 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1081 "/PDM/BlkCache/CacheMisses",
1082 STAMUNIT_COUNT, "Number of misses when accessing the cache");
1083 STAMR3Register(pVM, &pBlkCacheGlobal->StatRead,
1084 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1085 "/PDM/BlkCache/CacheRead",
1086 STAMUNIT_BYTES, "Number of bytes read from the cache");
1087 STAMR3Register(pVM, &pBlkCacheGlobal->StatWritten,
1088 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1089 "/PDM/BlkCache/CacheWritten",
1090 STAMUNIT_BYTES, "Number of bytes written to the cache");
1091 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeGet,
1092 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1093 "/PDM/BlkCache/CacheTreeGet",
1094 STAMUNIT_TICKS_PER_CALL, "Time taken to access an entry in the tree");
1095 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeInsert,
1096 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1097 "/PDM/BlkCache/CacheTreeInsert",
1098 STAMUNIT_TICKS_PER_CALL, "Time taken to insert an entry in the tree");
1099 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeRemove,
1100 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1101 "/PDM/BlkCache/CacheTreeRemove",
1102 STAMUNIT_TICKS_PER_CALL, "Time taken to remove an entry an the tree");
1103 STAMR3Register(pVM, &pBlkCacheGlobal->StatBuffersReused,
1104 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1105 "/PDM/BlkCache/CacheBuffersReused",
1106 STAMUNIT_COUNT, "Number of times a buffer could be reused");
1107#endif
1108
1109 /* Initialize the critical section */
1110 rc = RTCritSectInit(&pBlkCacheGlobal->CritSect);
1111 }
1112
1113 if (RT_SUCCESS(rc))
1114 {
1115 /* Create the commit timer */
1116 if (pBlkCacheGlobal->u32CommitTimeoutMs > 0)
1117 rc = TMR3TimerCreateInternal(pVM, TMCLOCK_REAL,
1118 pdmBlkCacheCommitTimerCallback,
1119 pBlkCacheGlobal,
1120 "BlkCache-Commit",
1121 &pBlkCacheGlobal->pTimerCommit);
1122
1123 if (RT_SUCCESS(rc))
1124 {
1125 /* Register saved state handler. */
1126 rc = SSMR3RegisterInternal(pVM, "pdmblkcache", 0, PDM_BLK_CACHE_SAVED_STATE_VERSION, pBlkCacheGlobal->cbMax,
1127 NULL, NULL, NULL,
1128 NULL, pdmR3BlkCacheSaveExec, NULL,
1129 NULL, pdmR3BlkCacheLoadExec, NULL);
1130 if (RT_SUCCESS(rc))
1131 {
1132 LogRel(("BlkCache: Cache successfully initialised. Cache size is %u bytes\n", pBlkCacheGlobal->cbMax));
1133 LogRel(("BlkCache: Cache commit interval is %u ms\n", pBlkCacheGlobal->u32CommitTimeoutMs));
1134 LogRel(("BlkCache: Cache commit threshold is %u bytes\n", pBlkCacheGlobal->cbCommitDirtyThreshold));
1135 pUVM->pdm.s.pBlkCacheGlobal = pBlkCacheGlobal;
1136 return VINF_SUCCESS;
1137 }
1138 }
1139
1140 RTCritSectDelete(&pBlkCacheGlobal->CritSect);
1141 }
1142
1143 if (pBlkCacheGlobal)
1144 RTMemFree(pBlkCacheGlobal);
1145
1146 LogFlowFunc((": returns rc=%Rrc\n", pVM, rc));
1147 return rc;
1148}
1149
1150void pdmR3BlkCacheTerm(PVM pVM)
1151{
1152 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1153
1154 if (pBlkCacheGlobal)
1155 {
1156 /* Make sure no one else uses the cache now */
1157 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1158
1159 /* Cleanup deleting all cache entries waiting for in progress entries to finish. */
1160 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruRecentlyUsedIn);
1161 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruRecentlyUsedOut);
1162 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruFrequentlyUsed);
1163
1164 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1165
1166 RTCritSectDelete(&pBlkCacheGlobal->CritSect);
1167 RTMemFree(pBlkCacheGlobal);
1168 pVM->pUVM->pdm.s.pBlkCacheGlobal = NULL;
1169 }
1170}
1171
1172int pdmR3BlkCacheResume(PVM pVM)
1173{
1174 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1175
1176 LogFlowFunc(("pVM=%#p\n", pVM));
1177
1178 if ( pBlkCacheGlobal
1179 && ASMAtomicXchgBool(&pBlkCacheGlobal->fIoErrorVmSuspended, false))
1180 {
1181 /* The VM was suspended because of an I/O error, commit all dirty entries. */
1182 pdmBlkCacheCommitDirtyEntries(pBlkCacheGlobal);
1183 }
1184
1185 return VINF_SUCCESS;
1186}
1187
1188static int pdmR3BlkCacheRetain(PVM pVM, PPPDMBLKCACHE ppBlkCache, const char *pcszId)
1189{
1190 int rc = VINF_SUCCESS;
1191 PPDMBLKCACHE pBlkCache = NULL;
1192 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1193
1194 if (!pBlkCacheGlobal)
1195 return VERR_NOT_SUPPORTED;
1196
1197 /*
1198 * Check that no other user cache has the same id first,
1199 * Unique id's are necessary in case the state is saved.
1200 */
1201 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1202
1203 pBlkCache = pdmR3BlkCacheFindById(pBlkCacheGlobal, pcszId);
1204
1205 if (!pBlkCache)
1206 {
1207 pBlkCache = (PPDMBLKCACHE)RTMemAllocZ(sizeof(PDMBLKCACHE));
1208
1209 if (pBlkCache)
1210 pBlkCache->pszId = RTStrDup(pcszId);
1211
1212 if ( pBlkCache
1213 && pBlkCache->pszId)
1214 {
1215 pBlkCache->fSuspended = false;
1216 pBlkCache->pCache = pBlkCacheGlobal;
1217 RTListInit(&pBlkCache->ListDirtyNotCommitted);
1218
1219 rc = RTSpinlockCreate(&pBlkCache->LockList);
1220 if (RT_SUCCESS(rc))
1221 {
1222 rc = RTSemRWCreate(&pBlkCache->SemRWEntries);
1223 if (RT_SUCCESS(rc))
1224 {
1225 pBlkCache->pTree = (PAVLRU64TREE)RTMemAllocZ(sizeof(AVLRFOFFTREE));
1226 if (pBlkCache->pTree)
1227 {
1228#ifdef VBOX_WITH_STATISTICS
1229 STAMR3RegisterF(pBlkCacheGlobal->pVM, &pBlkCache->StatWriteDeferred,
1230 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1231 STAMUNIT_COUNT, "Number of deferred writes",
1232 "/PDM/BlkCache/%s/Cache/DeferredWrites", pBlkCache->pszId);
1233#endif
1234
1235 /* Add to the list of users. */
1236 pBlkCacheGlobal->cRefs++;
1237 RTListAppend(&pBlkCacheGlobal->ListUsers, &pBlkCache->NodeCacheUser);
1238 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1239
1240 *ppBlkCache = pBlkCache;
1241 LogFlowFunc(("returns success\n"));
1242 return VINF_SUCCESS;
1243 }
1244 else
1245 rc = VERR_NO_MEMORY;
1246
1247 RTSemRWDestroy(pBlkCache->SemRWEntries);
1248 }
1249
1250 RTSpinlockDestroy(pBlkCache->LockList);
1251 }
1252
1253 RTStrFree(pBlkCache->pszId);
1254 }
1255 else
1256 rc = VERR_NO_MEMORY;
1257
1258 if (pBlkCache)
1259 RTMemFree(pBlkCache);
1260 }
1261 else
1262 rc = VERR_ALREADY_EXISTS;
1263
1264 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1265
1266 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1267 return rc;
1268}
1269
1270VMMR3DECL(int) PDMR3BlkCacheRetainDriver(PVM pVM, PPDMDRVINS pDrvIns, PPPDMBLKCACHE ppBlkCache,
1271 PFNPDMBLKCACHEXFERCOMPLETEDRV pfnXferComplete,
1272 PFNPDMBLKCACHEXFERENQUEUEDRV pfnXferEnqueue,
1273 const char *pcszId)
1274{
1275 int rc = VINF_SUCCESS;
1276 PPDMBLKCACHE pBlkCache;
1277
1278 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1279 if (RT_SUCCESS(rc))
1280 {
1281 pBlkCache->enmType = PDMBLKCACHETYPE_DRV;
1282 pBlkCache->u.Drv.pfnXferComplete = pfnXferComplete;
1283 pBlkCache->u.Drv.pfnXferEnqueue = pfnXferEnqueue;
1284 pBlkCache->u.Drv.pDrvIns = pDrvIns;
1285 *ppBlkCache = pBlkCache;
1286 }
1287
1288 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1289 return rc;
1290}
1291
1292VMMR3DECL(int) PDMR3BlkCacheRetainDevice(PVM pVM, PPDMDEVINS pDevIns, PPPDMBLKCACHE ppBlkCache,
1293 PFNPDMBLKCACHEXFERCOMPLETEDEV pfnXferComplete,
1294 PFNPDMBLKCACHEXFERENQUEUEDEV pfnXferEnqueue,
1295 const char *pcszId)
1296{
1297 int rc = VINF_SUCCESS;
1298 PPDMBLKCACHE pBlkCache;
1299
1300 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1301 if (RT_SUCCESS(rc))
1302 {
1303 pBlkCache->enmType = PDMBLKCACHETYPE_DEV;
1304 pBlkCache->u.Dev.pfnXferComplete = pfnXferComplete;
1305 pBlkCache->u.Dev.pfnXferEnqueue = pfnXferEnqueue;
1306 pBlkCache->u.Dev.pDevIns = pDevIns;
1307 *ppBlkCache = pBlkCache;
1308 }
1309
1310 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1311 return rc;
1312
1313}
1314
1315VMMR3DECL(int) PDMR3BlkCacheRetainUsb(PVM pVM, PPDMUSBINS pUsbIns, PPPDMBLKCACHE ppBlkCache,
1316 PFNPDMBLKCACHEXFERCOMPLETEUSB pfnXferComplete,
1317 PFNPDMBLKCACHEXFERENQUEUEUSB pfnXferEnqueue,
1318 const char *pcszId)
1319{
1320 int rc = VINF_SUCCESS;
1321 PPDMBLKCACHE pBlkCache;
1322
1323 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1324 if (RT_SUCCESS(rc))
1325 {
1326 pBlkCache->enmType = PDMBLKCACHETYPE_USB;
1327 pBlkCache->u.Usb.pfnXferComplete = pfnXferComplete;
1328 pBlkCache->u.Usb.pfnXferEnqueue = pfnXferEnqueue;
1329 pBlkCache->u.Usb.pUsbIns = pUsbIns;
1330 *ppBlkCache = pBlkCache;
1331 }
1332
1333 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1334 return rc;
1335
1336}
1337
1338VMMR3DECL(int) PDMR3BlkCacheRetainInt(PVM pVM, void *pvUser, PPPDMBLKCACHE ppBlkCache,
1339 PFNPDMBLKCACHEXFERCOMPLETEINT pfnXferComplete,
1340 PFNPDMBLKCACHEXFERENQUEUEINT pfnXferEnqueue,
1341 const char *pcszId)
1342{
1343 int rc = VINF_SUCCESS;
1344 PPDMBLKCACHE pBlkCache;
1345
1346 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1347 if (RT_SUCCESS(rc))
1348 {
1349 pBlkCache->enmType = PDMBLKCACHETYPE_INTERNAL;
1350 pBlkCache->u.Int.pfnXferComplete = pfnXferComplete;
1351 pBlkCache->u.Int.pfnXferEnqueue = pfnXferEnqueue;
1352 pBlkCache->u.Int.pvUser = pvUser;
1353 *ppBlkCache = pBlkCache;
1354 }
1355
1356 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1357 return rc;
1358
1359}
1360
1361/**
1362 * Callback for the AVL destroy routine. Frees a cache entry for this endpoint.
1363 *
1364 * @returns IPRT status code.
1365 * @param pNode The node to destroy.
1366 * @param pvUser Opaque user data.
1367 */
1368static int pdmBlkCacheEntryDestroy(PAVLRU64NODECORE pNode, void *pvUser)
1369{
1370 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)pNode;
1371 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
1372 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
1373
1374 while (ASMAtomicReadU32(&pEntry->fFlags) & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS)
1375 {
1376 /* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
1377 pdmBlkCacheEntryRef(pEntry);
1378 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1379 pdmBlkCacheLockLeave(pCache);
1380
1381 RTThreadSleep(250);
1382
1383 /* Re-enter all locks */
1384 pdmBlkCacheLockEnter(pCache);
1385 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1386 pdmBlkCacheEntryRelease(pEntry);
1387 }
1388
1389 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
1390 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
1391
1392 bool fUpdateCache = pEntry->pList == &pCache->LruFrequentlyUsed
1393 || pEntry->pList == &pCache->LruRecentlyUsedIn;
1394
1395 pdmBlkCacheEntryRemoveFromList(pEntry);
1396
1397 if (fUpdateCache)
1398 pdmBlkCacheSub(pCache, pEntry->cbData);
1399
1400 RTMemPageFree(pEntry->pbData, pEntry->cbData);
1401 RTMemFree(pEntry);
1402
1403 return VINF_SUCCESS;
1404}
1405
1406/**
1407 * Destroys all cache resources used by the given endpoint.
1408 *
1409 * @returns nothing.
1410 * @param pEndpoint The endpoint to the destroy.
1411 */
1412VMMR3DECL(void) PDMR3BlkCacheRelease(PPDMBLKCACHE pBlkCache)
1413{
1414 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1415
1416 /*
1417 * Commit all dirty entries now (they are waited on for completion during the
1418 * destruction of the AVL tree below).
1419 * The exception is if the VM was paused because of an I/O error before.
1420 */
1421 if (!ASMAtomicReadBool(&pCache->fIoErrorVmSuspended))
1422 pdmBlkCacheCommit(pBlkCache);
1423
1424 /* Make sure nobody is accessing the cache while we delete the tree. */
1425 pdmBlkCacheLockEnter(pCache);
1426 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1427 RTAvlrU64Destroy(pBlkCache->pTree, pdmBlkCacheEntryDestroy, pCache);
1428 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1429
1430 RTSpinlockDestroy(pBlkCache->LockList);
1431
1432 pCache->cRefs--;
1433 RTListNodeRemove(&pBlkCache->NodeCacheUser);
1434
1435 pdmBlkCacheLockLeave(pCache);
1436
1437 RTSemRWDestroy(pBlkCache->SemRWEntries);
1438
1439#ifdef VBOX_WITH_STATISTICS
1440 STAMR3Deregister(pCache->pVM, &pBlkCache->StatWriteDeferred);
1441#endif
1442
1443 RTStrFree(pBlkCache->pszId);
1444 RTMemFree(pBlkCache);
1445}
1446
1447VMMR3DECL(void) PDMR3BlkCacheReleaseDevice(PVM pVM, PPDMDEVINS pDevIns)
1448{
1449 LogFlow(("%s: pDevIns=%p\n", __FUNCTION__, pDevIns));
1450
1451 /*
1452 * Validate input.
1453 */
1454 if (!pDevIns)
1455 return;
1456 VM_ASSERT_EMT(pVM);
1457
1458 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1459 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1460
1461 /* Return silently if not supported. */
1462 if (!pBlkCacheGlobal)
1463 return;
1464
1465 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1466
1467 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1468 {
1469 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DEV
1470 && pBlkCache->u.Dev.pDevIns == pDevIns)
1471 PDMR3BlkCacheRelease(pBlkCache);
1472 }
1473
1474 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1475}
1476
1477VMMR3DECL(void) PDMR3BlkCacheReleaseDriver(PVM pVM, PPDMDRVINS pDrvIns)
1478{
1479 LogFlow(("%s: pDrvIns=%p\n", __FUNCTION__, pDrvIns));
1480
1481 /*
1482 * Validate input.
1483 */
1484 if (!pDrvIns)
1485 return;
1486 VM_ASSERT_EMT(pVM);
1487
1488 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1489 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1490
1491 /* Return silently if not supported. */
1492 if (!pBlkCacheGlobal)
1493 return;
1494
1495 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1496
1497 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1498 {
1499 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DRV
1500 && pBlkCache->u.Drv.pDrvIns == pDrvIns)
1501 PDMR3BlkCacheRelease(pBlkCache);
1502 }
1503
1504 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1505}
1506
1507VMMR3DECL(void) PDMR3BlkCacheReleaseUsb(PVM pVM, PPDMUSBINS pUsbIns)
1508{
1509 LogFlow(("%s: pUsbIns=%p\n", __FUNCTION__, pUsbIns));
1510
1511 /*
1512 * Validate input.
1513 */
1514 if (!pUsbIns)
1515 return;
1516 VM_ASSERT_EMT(pVM);
1517
1518 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1519 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1520
1521 /* Return silently if not supported. */
1522 if (!pBlkCacheGlobal)
1523 return;
1524
1525 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1526
1527 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1528 {
1529 if ( pBlkCache->enmType == PDMBLKCACHETYPE_USB
1530 && pBlkCache->u.Usb.pUsbIns == pUsbIns)
1531 PDMR3BlkCacheRelease(pBlkCache);
1532 }
1533
1534 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1535}
1536
1537static PPDMBLKCACHEENTRY pdmBlkCacheGetCacheEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off)
1538{
1539 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1540 PPDMBLKCACHEENTRY pEntry = NULL;
1541
1542 STAM_PROFILE_ADV_START(&pCache->StatTreeGet, Cache);
1543
1544 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1545 pEntry = (PPDMBLKCACHEENTRY)RTAvlrU64RangeGet(pBlkCache->pTree, off);
1546 if (pEntry)
1547 pdmBlkCacheEntryRef(pEntry);
1548 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1549
1550 STAM_PROFILE_ADV_STOP(&pCache->StatTreeGet, Cache);
1551
1552 return pEntry;
1553}
1554
1555/**
1556 * Return the best fit cache entries for the given offset.
1557 *
1558 * @returns nothing.
1559 * @param pBlkCache The endpoint cache.
1560 * @param off The offset.
1561 * @param pEntryAbove Where to store the pointer to the best fit entry above the
1562 * the given offset. NULL if not required.
1563 */
1564static void pdmBlkCacheGetCacheBestFitEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off,
1565 PPDMBLKCACHEENTRY *ppEntryAbove)
1566{
1567 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1568
1569 STAM_PROFILE_ADV_START(&pCache->StatTreeGet, Cache);
1570
1571 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1572 if (ppEntryAbove)
1573 {
1574 *ppEntryAbove = (PPDMBLKCACHEENTRY)RTAvlrU64GetBestFit(pBlkCache->pTree, off, true /*fAbove*/);
1575 if (*ppEntryAbove)
1576 pdmBlkCacheEntryRef(*ppEntryAbove);
1577 }
1578
1579 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1580
1581 STAM_PROFILE_ADV_STOP(&pCache->StatTreeGet, Cache);
1582}
1583
1584static void pdmBlkCacheInsertEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
1585{
1586 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1587
1588 STAM_PROFILE_ADV_START(&pCache->StatTreeInsert, Cache);
1589 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1590 bool fInserted = RTAvlrU64Insert(pBlkCache->pTree, &pEntry->Core);
1591 AssertMsg(fInserted, ("Node was not inserted into tree\n"));
1592 STAM_PROFILE_ADV_STOP(&pCache->StatTreeInsert, Cache);
1593 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1594}
1595
1596/**
1597 * Allocates and initializes a new entry for the cache.
1598 * The entry has a reference count of 1.
1599 *
1600 * @returns Pointer to the new cache entry or NULL if out of memory.
1601 * @param pBlkCache The cache the entry belongs to.
1602 * @param off Start offset.
1603 * @param cbData Size of the cache entry.
1604 * @param pbBuffer Pointer to the buffer to use.
1605 * NULL if a new buffer should be allocated.
1606 * The buffer needs to have the same size of the entry.
1607 */
1608static PPDMBLKCACHEENTRY pdmBlkCacheEntryAlloc(PPDMBLKCACHE pBlkCache,
1609 uint64_t off, size_t cbData, uint8_t *pbBuffer)
1610{
1611 PPDMBLKCACHEENTRY pEntryNew = (PPDMBLKCACHEENTRY)RTMemAllocZ(sizeof(PDMBLKCACHEENTRY));
1612
1613 if (RT_UNLIKELY(!pEntryNew))
1614 return NULL;
1615
1616 pEntryNew->Core.Key = off;
1617 pEntryNew->Core.KeyLast = off + cbData - 1;
1618 pEntryNew->pBlkCache = pBlkCache;
1619 pEntryNew->fFlags = 0;
1620 pEntryNew->cRefs = 1; /* We are using it now. */
1621 pEntryNew->pList = NULL;
1622 pEntryNew->cbData = cbData;
1623 pEntryNew->pWaitingHead = NULL;
1624 pEntryNew->pWaitingTail = NULL;
1625 if (pbBuffer)
1626 pEntryNew->pbData = pbBuffer;
1627 else
1628 pEntryNew->pbData = (uint8_t *)RTMemPageAlloc(cbData);
1629
1630 if (RT_UNLIKELY(!pEntryNew->pbData))
1631 {
1632 RTMemFree(pEntryNew);
1633 return NULL;
1634 }
1635
1636 return pEntryNew;
1637}
1638
1639/**
1640 * Checks that a set of flags is set/clear acquiring the R/W semaphore
1641 * in exclusive mode.
1642 *
1643 * @returns true if the flag in fSet is set and the one in fClear is clear.
1644 * false otherwise.
1645 * The R/W semaphore is only held if true is returned.
1646 *
1647 * @param pBlkCache The endpoint cache instance data.
1648 * @param pEntry The entry to check the flags for.
1649 * @param fSet The flag which is tested to be set.
1650 * @param fClear The flag which is tested to be clear.
1651 */
1652DECLINLINE(bool) pdmBlkCacheEntryFlagIsSetClearAcquireLock(PPDMBLKCACHE pBlkCache,
1653 PPDMBLKCACHEENTRY pEntry,
1654 uint32_t fSet, uint32_t fClear)
1655{
1656 uint32_t fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1657 bool fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1658
1659 if (fPassed)
1660 {
1661 /* Acquire the lock and check again because the completion callback might have raced us. */
1662 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1663
1664 fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1665 fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1666
1667 /* Drop the lock if we didn't passed the test. */
1668 if (!fPassed)
1669 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1670 }
1671
1672 return fPassed;
1673}
1674
1675/**
1676 * Adds a segment to the waiting list for a cache entry
1677 * which is currently in progress.
1678 *
1679 * @returns nothing.
1680 * @param pEntry The cache entry to add the segment to.
1681 * @param pSeg The segment to add.
1682 */
1683DECLINLINE(void) pdmBlkCacheEntryAddWaiter(PPDMBLKCACHEENTRY pEntry,
1684 PPDMBLKCACHEWAITER pWaiter)
1685{
1686 pWaiter->pNext = NULL;
1687
1688 if (pEntry->pWaitingHead)
1689 {
1690 AssertPtr(pEntry->pWaitingTail);
1691
1692 pEntry->pWaitingTail->pNext = pWaiter;
1693 pEntry->pWaitingTail = pWaiter;
1694 }
1695 else
1696 {
1697 Assert(!pEntry->pWaitingTail);
1698
1699 pEntry->pWaitingHead = pWaiter;
1700 pEntry->pWaitingTail = pWaiter;
1701 }
1702}
1703
1704/**
1705 * Add a buffer described by the I/O memory context
1706 * to the entry waiting for completion.
1707 *
1708 * @returns VBox status code.
1709 * @param pEntry The entry to add the buffer to.
1710 * @param pTask Task associated with the buffer.
1711 * @param pIoMemCtx The memory context to use.
1712 * @param offDiff Offset from the start of the buffer
1713 * in the entry.
1714 * @param cbData Amount of data to wait for onthis entry.
1715 * @param fWrite Flag whether the task waits because it wants to write
1716 * to the cache entry.
1717 */
1718static int pdmBlkCacheEntryWaitersAdd(PPDMBLKCACHEENTRY pEntry,
1719 PPDMBLKCACHEREQ pReq,
1720 PRTSGBUF pSgBuf, uint64_t offDiff,
1721 size_t cbData, bool fWrite)
1722{
1723 PPDMBLKCACHEWAITER pWaiter = (PPDMBLKCACHEWAITER)RTMemAllocZ(sizeof(PDMBLKCACHEWAITER));
1724 if (!pWaiter)
1725 return VERR_NO_MEMORY;
1726
1727 ASMAtomicIncU32(&pReq->cXfersPending);
1728 pWaiter->pReq = pReq;
1729 pWaiter->offCacheEntry = offDiff;
1730 pWaiter->cbTransfer = cbData;
1731 pWaiter->fWrite = fWrite;
1732 RTSgBufClone(&pWaiter->SgBuf, pSgBuf);
1733 RTSgBufAdvance(pSgBuf, cbData);
1734
1735 pdmBlkCacheEntryAddWaiter(pEntry, pWaiter);
1736
1737 return VINF_SUCCESS;
1738}
1739
1740/**
1741 * Calculate aligned offset and size for a new cache entry
1742 * which do not intersect with an already existing entry and the
1743 * file end.
1744 *
1745 * @returns The number of bytes the entry can hold of the requested amount
1746 * of byte.
1747 * @param pEndpoint The endpoint.
1748 * @param pBlkCache The endpoint cache.
1749 * @param off The start offset.
1750 * @param cb The number of bytes the entry needs to hold at least.
1751 * @param uAlignment Alignment of the boundary sizes.
1752 * @param poffAligned Where to store the aligned offset.
1753 * @param pcbAligned Where to store the aligned size of the entry.
1754 */
1755static size_t pdmBlkCacheEntryBoundariesCalc(PPDMBLKCACHE pBlkCache,
1756 uint64_t off, size_t cb,
1757 unsigned uAlignment,
1758 uint64_t *poffAligned, size_t *pcbAligned)
1759{
1760 size_t cbAligned;
1761 size_t cbInEntry = 0;
1762 uint64_t offAligned;
1763 PPDMBLKCACHEENTRY pEntryAbove = NULL;
1764
1765 /* Get the best fit entries around the offset */
1766 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off, &pEntryAbove);
1767
1768 /* Log the info */
1769 LogFlow(("%sest fit entry above off=%llu (BestFit=%llu BestFitEnd=%llu BestFitSize=%u)\n",
1770 pEntryAbove ? "B" : "No b",
1771 off,
1772 pEntryAbove ? pEntryAbove->Core.Key : 0,
1773 pEntryAbove ? pEntryAbove->Core.KeyLast : 0,
1774 pEntryAbove ? pEntryAbove->cbData : 0));
1775
1776 offAligned = off;
1777
1778 if ( pEntryAbove
1779 && off + cb > pEntryAbove->Core.Key)
1780 {
1781 cbInEntry = pEntryAbove->Core.Key - off;
1782 cbAligned = pEntryAbove->Core.Key - offAligned;
1783 }
1784 else
1785 {
1786 cbAligned = cb;
1787 cbInEntry = cb;
1788 }
1789
1790 /* A few sanity checks */
1791 AssertMsg(!pEntryAbove || (offAligned + cbAligned) <= pEntryAbove->Core.Key,
1792 ("Aligned size intersects with another cache entry\n"));
1793 Assert(cbInEntry <= cbAligned);
1794
1795 if (pEntryAbove)
1796 pdmBlkCacheEntryRelease(pEntryAbove);
1797
1798 LogFlow(("offAligned=%llu cbAligned=%u\n", offAligned, cbAligned));
1799
1800 *poffAligned = offAligned;
1801 *pcbAligned = cbAligned;
1802
1803 return cbInEntry;
1804}
1805
1806/**
1807 * Create a new cache entry evicting data from the cache if required.
1808 *
1809 * @returns Pointer to the new cache entry or NULL
1810 * if not enough bytes could be evicted from the cache.
1811 * @param pEndpoint The endpoint.
1812 * @param pBlkCache The endpoint cache.
1813 * @param off The offset.
1814 * @param cb Number of bytes the cache entry should have.
1815 * @param uAlignment Alignment the size of the entry should have.
1816 * @param pcbData Where to store the number of bytes the new
1817 * entry can hold. May be lower than actually requested
1818 * due to another entry intersecting the access range.
1819 */
1820static PPDMBLKCACHEENTRY pdmBlkCacheEntryCreate(PPDMBLKCACHE pBlkCache,
1821 uint64_t off, size_t cb,
1822 unsigned uAlignment,
1823 size_t *pcbData)
1824{
1825 uint64_t offStart = 0;
1826 size_t cbEntry = 0;
1827 PPDMBLKCACHEENTRY pEntryNew = NULL;
1828 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1829 uint8_t *pbBuffer = NULL;
1830
1831 *pcbData = pdmBlkCacheEntryBoundariesCalc(pBlkCache, off, cb, uAlignment,
1832 &offStart, &cbEntry);
1833
1834 pdmBlkCacheLockEnter(pCache);
1835 bool fEnough = pdmBlkCacheReclaim(pCache, cbEntry, true, &pbBuffer);
1836
1837 if (fEnough)
1838 {
1839 LogFlow(("Evicted enough bytes (%u requested). Creating new cache entry\n", cbEntry));
1840
1841 pEntryNew = pdmBlkCacheEntryAlloc(pBlkCache, offStart, cbEntry, pbBuffer);
1842 if (RT_LIKELY(pEntryNew))
1843 {
1844 pdmBlkCacheEntryAddToList(&pCache->LruRecentlyUsedIn, pEntryNew);
1845 pdmBlkCacheAdd(pCache, cbEntry);
1846 pdmBlkCacheLockLeave(pCache);
1847
1848 pdmBlkCacheInsertEntry(pBlkCache, pEntryNew);
1849
1850 AssertMsg( (off >= pEntryNew->Core.Key)
1851 && (off + *pcbData <= pEntryNew->Core.KeyLast + 1),
1852 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
1853 off, pEntryNew->Core.Key));
1854 }
1855 else
1856 pdmBlkCacheLockLeave(pCache);
1857 }
1858 else
1859 pdmBlkCacheLockLeave(pCache);
1860
1861 return pEntryNew;
1862}
1863
1864static PPDMBLKCACHEREQ pdmBlkCacheReqAlloc(void *pvUser)
1865{
1866 PPDMBLKCACHEREQ pReq = (PPDMBLKCACHEREQ)RTMemAlloc(sizeof(PDMBLKCACHEREQ));
1867
1868 if (RT_LIKELY(pReq))
1869 {
1870 pReq->pvUser = pvUser;
1871 pReq->rcReq = VINF_SUCCESS;
1872 pReq->cXfersPending = 0;
1873 }
1874
1875 return pReq;
1876}
1877
1878static void pdmBlkCacheReqComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq)
1879{
1880 switch (pBlkCache->enmType)
1881 {
1882 case PDMBLKCACHETYPE_DEV:
1883 {
1884 pBlkCache->u.Dev.pfnXferComplete(pBlkCache->u.Dev.pDevIns,
1885 pReq->pvUser, pReq->rcReq);
1886 break;
1887 }
1888 case PDMBLKCACHETYPE_DRV:
1889 {
1890 pBlkCache->u.Drv.pfnXferComplete(pBlkCache->u.Drv.pDrvIns,
1891 pReq->pvUser, pReq->rcReq);
1892 break;
1893 }
1894 case PDMBLKCACHETYPE_USB:
1895 {
1896 pBlkCache->u.Usb.pfnXferComplete(pBlkCache->u.Usb.pUsbIns,
1897 pReq->pvUser, pReq->rcReq);
1898 break;
1899 }
1900 case PDMBLKCACHETYPE_INTERNAL:
1901 {
1902 pBlkCache->u.Int.pfnXferComplete(pBlkCache->u.Int.pvUser,
1903 pReq->pvUser, pReq->rcReq);
1904 break;
1905 }
1906 default:
1907 AssertMsgFailed(("Unknown block cache type!\n"));
1908 }
1909
1910 RTMemFree(pReq);
1911}
1912
1913static bool pdmBlkCacheReqUpdate(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq,
1914 int rcReq, bool fCallHandler)
1915{
1916 if (RT_FAILURE(rcReq))
1917 ASMAtomicCmpXchgS32(&pReq->rcReq, rcReq, VINF_SUCCESS);
1918
1919 AssertMsg(pReq->cXfersPending > 0, ("No transfers are pending for this request\n"));
1920 uint32_t cXfersPending = ASMAtomicDecU32(&pReq->cXfersPending);
1921
1922 if (!cXfersPending)
1923 {
1924 if (fCallHandler)
1925 pdmBlkCacheReqComplete(pBlkCache, pReq);
1926 else
1927 RTMemFree(pReq);
1928 return true;
1929 }
1930
1931 LogFlowFunc(("pReq=%#p cXfersPending=%u\n", pReq, cXfersPending));
1932 return false;
1933}
1934
1935VMMR3DECL(int) PDMR3BlkCacheRead(PPDMBLKCACHE pBlkCache, uint64_t off,
1936 PCRTSGBUF pcSgBuf, size_t cbRead, void *pvUser)
1937{
1938 int rc = VINF_SUCCESS;
1939 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1940 PPDMBLKCACHEENTRY pEntry;
1941 PPDMBLKCACHEREQ pReq;
1942
1943 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pcSgBuf=%#p cbRead=%u pvUser=%#p\n",
1944 pBlkCache, pBlkCache->pszId, off, pcSgBuf, cbRead, pvUser));
1945
1946 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
1947 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
1948
1949 RTSGBUF SgBuf;
1950 RTSgBufClone(&SgBuf, pcSgBuf);
1951
1952 /* Allocate new request structure. */
1953 pReq = pdmBlkCacheReqAlloc(pvUser);
1954 if (RT_UNLIKELY(!pReq))
1955 return VERR_NO_MEMORY;
1956
1957 /* Increment data transfer counter to keep the request valid while we access it. */
1958 ASMAtomicIncU32(&pReq->cXfersPending);
1959
1960 while (cbRead)
1961 {
1962 size_t cbToRead;
1963
1964 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
1965
1966 /*
1967 * If there is no entry we try to create a new one eviciting unused pages
1968 * if the cache is full. If this is not possible we will pass the request through
1969 * and skip the caching (all entries may be still in progress so they can't
1970 * be evicted)
1971 * If we have an entry it can be in one of the LRU lists where the entry
1972 * contains data (recently used or frequently used LRU) so we can just read
1973 * the data we need and put the entry at the head of the frequently used LRU list.
1974 * In case the entry is in one of the ghost lists it doesn't contain any data.
1975 * We have to fetch it again evicting pages from either T1 or T2 to make room.
1976 */
1977 if (pEntry)
1978 {
1979 uint64_t offDiff = off - pEntry->Core.Key;
1980
1981 AssertMsg(off >= pEntry->Core.Key,
1982 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
1983 off, pEntry->Core.Key));
1984
1985 AssertPtr(pEntry->pList);
1986
1987 cbToRead = RT_MIN(pEntry->cbData - offDiff, cbRead);
1988
1989 AssertMsg(off + cbToRead <= pEntry->Core.Key + pEntry->Core.KeyLast + 1,
1990 ("Buffer of cache entry exceeded off=%llu cbToRead=%d\n",
1991 off, cbToRead));
1992
1993 cbRead -= cbToRead;
1994
1995 if (!cbRead)
1996 STAM_COUNTER_INC(&pCache->cHits);
1997 else
1998 STAM_COUNTER_INC(&pCache->cPartialHits);
1999
2000 STAM_COUNTER_ADD(&pCache->StatRead, cbToRead);
2001
2002 /* Ghost lists contain no data. */
2003 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2004 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2005 {
2006 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2007 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2008 PDMBLKCACHE_ENTRY_IS_DIRTY))
2009 {
2010 /* Entry didn't completed yet. Append to the list */
2011 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2012 &SgBuf, offDiff, cbToRead,
2013 false /* fWrite */);
2014 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2015 }
2016 else
2017 {
2018 /* Read as much as we can from the entry. */
2019 RTSgBufCopyFromBuf(&SgBuf, pEntry->pbData + offDiff, cbToRead);
2020 }
2021
2022 /* Move this entry to the top position */
2023 if (pEntry->pList == &pCache->LruFrequentlyUsed)
2024 {
2025 pdmBlkCacheLockEnter(pCache);
2026 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2027 pdmBlkCacheLockLeave(pCache);
2028 }
2029 /* Release the entry */
2030 pdmBlkCacheEntryRelease(pEntry);
2031 }
2032 else
2033 {
2034 uint8_t *pbBuffer = NULL;
2035
2036 LogFlow(("Fetching data for ghost entry %#p from file\n", pEntry));
2037
2038 pdmBlkCacheLockEnter(pCache);
2039 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2040 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2041
2042 /* Move the entry to Am and fetch it to the cache. */
2043 if (fEnough)
2044 {
2045 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2046 pdmBlkCacheAdd(pCache, pEntry->cbData);
2047 pdmBlkCacheLockLeave(pCache);
2048
2049 if (pbBuffer)
2050 pEntry->pbData = pbBuffer;
2051 else
2052 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2053 AssertPtr(pEntry->pbData);
2054
2055 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2056 &SgBuf, offDiff, cbToRead,
2057 false /* fWrite */);
2058 pdmBlkCacheEntryReadFromMedium(pEntry);
2059 /* Release the entry */
2060 pdmBlkCacheEntryRelease(pEntry);
2061 }
2062 else
2063 {
2064 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2065 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2066 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2067 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2068 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2069
2070 pdmBlkCacheLockLeave(pCache);
2071
2072 RTMemFree(pEntry);
2073
2074 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2075 &SgBuf, off, cbToRead,
2076 PDMBLKCACHEXFERDIR_READ);
2077 }
2078 }
2079 }
2080 else
2081 {
2082#ifdef VBOX_WITH_IO_READ_CACHE
2083 /* No entry found for this offset. Create a new entry and fetch the data to the cache. */
2084 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
2085 off, cbRead,
2086 PAGE_SIZE,
2087 &cbToRead);
2088
2089 cbRead -= cbToRead;
2090
2091 if (pEntryNew)
2092 {
2093 if (!cbRead)
2094 STAM_COUNTER_INC(&pCache->cMisses);
2095 else
2096 STAM_COUNTER_INC(&pCache->cPartialHits);
2097
2098 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
2099 &SgBuf,
2100 off - pEntryNew->Core.Key,
2101 cbToRead,
2102 false /* fWrite */);
2103 pdmBlkCacheEntryReadFromMedium(pEntryNew);
2104 pdmBlkCacheEntryRelease(pEntryNew); /* it is protected by the I/O in progress flag now. */
2105 }
2106 else
2107 {
2108 /*
2109 * There is not enough free space in the cache.
2110 * Pass the request directly to the I/O manager.
2111 */
2112 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToRead));
2113
2114 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2115 &SgBuf, off, cbToRead,
2116 PDMBLKCACHEXFERDIR_READ);
2117 }
2118#else
2119 /* Clip read size if necessary. */
2120 PPDMBLKCACHEENTRY pEntryAbove;
2121 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off, &pEntryAbove);
2122
2123 if (pEntryAbove)
2124 {
2125 if (off + cbRead > pEntryAbove->Core.Key)
2126 cbToRead = pEntryAbove->Core.Key - off;
2127 else
2128 cbToRead = cbRead;
2129
2130 pdmBlkCacheEntryRelease(pEntryAbove);
2131 }
2132 else
2133 cbToRead = cbRead;
2134
2135 cbRead -= cbToRead;
2136 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2137 &SgBuf, off, cbToRead,
2138 PDMBLKCACHEXFERDIR_READ);
2139#endif
2140 }
2141 off += cbToRead;
2142 }
2143
2144 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, false))
2145 rc = VINF_AIO_TASK_PENDING;
2146
2147 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2148
2149 return rc;
2150}
2151
2152VMMR3DECL(int) PDMR3BlkCacheWrite(PPDMBLKCACHE pBlkCache, uint64_t off,
2153 PCRTSGBUF pcSgBuf, size_t cbWrite, void *pvUser)
2154{
2155 int rc = VINF_SUCCESS;
2156 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2157 PPDMBLKCACHEENTRY pEntry;
2158 PPDMBLKCACHEREQ pReq;
2159
2160 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pcSgBuf=%#p cbWrite=%u pvUser=%#p\n",
2161 pBlkCache, pBlkCache->pszId, off, pcSgBuf, cbWrite, pvUser));
2162
2163 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2164 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
2165
2166 RTSGBUF SgBuf;
2167 RTSgBufClone(&SgBuf, pcSgBuf);
2168
2169 /* Allocate new request structure. */
2170 pReq = pdmBlkCacheReqAlloc(pvUser);
2171 if (RT_UNLIKELY(!pReq))
2172 return VERR_NO_MEMORY;
2173
2174 /* Increment data transfer counter to keep the request valid while we access it. */
2175 ASMAtomicIncU32(&pReq->cXfersPending);
2176
2177 while (cbWrite)
2178 {
2179 size_t cbToWrite;
2180
2181 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
2182
2183 if (pEntry)
2184 {
2185 /* Write the data into the entry and mark it as dirty */
2186 AssertPtr(pEntry->pList);
2187
2188 uint64_t offDiff = off - pEntry->Core.Key;
2189
2190 AssertMsg(off >= pEntry->Core.Key,
2191 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
2192 off, pEntry->Core.Key));
2193
2194 cbToWrite = RT_MIN(pEntry->cbData - offDiff, cbWrite);
2195 cbWrite -= cbToWrite;
2196
2197 if (!cbWrite)
2198 STAM_COUNTER_INC(&pCache->cHits);
2199 else
2200 STAM_COUNTER_INC(&pCache->cPartialHits);
2201
2202 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2203
2204 /* Ghost lists contain no data. */
2205 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2206 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2207 {
2208 /* Check if the entry is dirty. */
2209 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2210 PDMBLKCACHE_ENTRY_IS_DIRTY,
2211 0))
2212 {
2213 /* If it is already dirty but not in progress just update the data. */
2214 if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS))
2215 {
2216 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff,
2217 cbToWrite);
2218 }
2219 else
2220 {
2221 /* The data isn't written to the file yet */
2222 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2223 &SgBuf, offDiff, cbToWrite,
2224 true /* fWrite */);
2225 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2226 }
2227
2228 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2229 }
2230 else /* Dirty bit not set */
2231 {
2232 /*
2233 * Check if a read is in progress for this entry.
2234 * We have to defer processing in that case.
2235 */
2236 if(pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2237 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2238 0))
2239 {
2240 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2241 &SgBuf, offDiff, cbToWrite,
2242 true /* fWrite */);
2243 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2244 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2245 }
2246 else /* I/O in progress flag not set */
2247 {
2248 /* Write as much as we can into the entry and update the file. */
2249 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff, cbToWrite);
2250
2251 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2252 if (fCommit)
2253 pdmBlkCacheCommitDirtyEntries(pCache);
2254 }
2255 } /* Dirty bit not set */
2256
2257 /* Move this entry to the top position */
2258 if (pEntry->pList == &pCache->LruFrequentlyUsed)
2259 {
2260 pdmBlkCacheLockEnter(pCache);
2261 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2262 pdmBlkCacheLockLeave(pCache);
2263 }
2264
2265 pdmBlkCacheEntryRelease(pEntry);
2266 }
2267 else /* Entry is on the ghost list */
2268 {
2269 uint8_t *pbBuffer = NULL;
2270
2271 pdmBlkCacheLockEnter(pCache);
2272 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2273 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2274
2275 if (fEnough)
2276 {
2277 /* Move the entry to Am and fetch it to the cache. */
2278 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2279 pdmBlkCacheAdd(pCache, pEntry->cbData);
2280 pdmBlkCacheLockLeave(pCache);
2281
2282 if (pbBuffer)
2283 pEntry->pbData = pbBuffer;
2284 else
2285 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2286 AssertPtr(pEntry->pbData);
2287
2288 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2289 &SgBuf, offDiff, cbToWrite,
2290 true /* fWrite */);
2291 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2292 pdmBlkCacheEntryReadFromMedium(pEntry);
2293
2294 /* Release the reference. If it is still needed the I/O in progress flag should protect it now. */
2295 pdmBlkCacheEntryRelease(pEntry);
2296 }
2297 else
2298 {
2299 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2300 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2301 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2302 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2303 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2304
2305 pdmBlkCacheLockLeave(pCache);
2306
2307 RTMemFree(pEntry);
2308 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2309 &SgBuf, off, cbToWrite,
2310 PDMBLKCACHEXFERDIR_WRITE);
2311 }
2312 }
2313 }
2314 else /* No entry found */
2315 {
2316 /*
2317 * No entry found. Try to create a new cache entry to store the data in and if that fails
2318 * write directly to the file.
2319 */
2320 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
2321 off, cbWrite,
2322 512, &cbToWrite);
2323
2324 cbWrite -= cbToWrite;
2325
2326 if (pEntryNew)
2327 {
2328 uint64_t offDiff = off - pEntryNew->Core.Key;
2329
2330 STAM_COUNTER_INC(&pCache->cHits);
2331
2332 /*
2333 * Check if it is possible to just write the data without waiting
2334 * for it to get fetched first.
2335 */
2336 if (!offDiff && pEntryNew->cbData == cbToWrite)
2337 {
2338 RTSgBufCopyToBuf(&SgBuf, pEntryNew->pbData, cbToWrite);
2339
2340 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntryNew);
2341 if (fCommit)
2342 pdmBlkCacheCommitDirtyEntries(pCache);
2343 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2344 }
2345 else
2346 {
2347 /* Defer the write and fetch the data from the endpoint. */
2348 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
2349 &SgBuf, offDiff, cbToWrite,
2350 true /* fWrite */);
2351 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2352 pdmBlkCacheEntryReadFromMedium(pEntryNew);
2353 }
2354
2355 pdmBlkCacheEntryRelease(pEntryNew);
2356 }
2357 else
2358 {
2359 /*
2360 * There is not enough free space in the cache.
2361 * Pass the request directly to the I/O manager.
2362 */
2363 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToWrite));
2364
2365 STAM_COUNTER_INC(&pCache->cMisses);
2366
2367 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2368 &SgBuf, off, cbToWrite,
2369 PDMBLKCACHEXFERDIR_WRITE);
2370 }
2371 }
2372
2373 off += cbToWrite;
2374 }
2375
2376 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, false))
2377 rc = VINF_AIO_TASK_PENDING;
2378
2379 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2380
2381 return rc;
2382}
2383
2384VMMR3DECL(int) PDMR3BlkCacheFlush(PPDMBLKCACHE pBlkCache, void *pvUser)
2385{
2386 int rc = VINF_SUCCESS;
2387 PPDMBLKCACHEREQ pReq;
2388
2389 LogFlowFunc((": pBlkCache=%#p{%s}\n", pBlkCache, pBlkCache->pszId));
2390
2391 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2392 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
2393
2394 /* Commit dirty entries in the cache. */
2395 pdmBlkCacheCommit(pBlkCache);
2396
2397 /* Allocate new request structure. */
2398 pReq = pdmBlkCacheReqAlloc(pvUser);
2399 if (RT_UNLIKELY(!pReq))
2400 return VERR_NO_MEMORY;
2401
2402 rc = pdmBlkCacheRequestPassthrough(pBlkCache, pReq, NULL, 0, 0,
2403 PDMBLKCACHEXFERDIR_FLUSH);
2404 AssertRC(rc);
2405
2406 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2407 return VINF_AIO_TASK_PENDING;
2408}
2409
2410/**
2411 * Completes a task segment freeing all resources and completes the task handle
2412 * if everything was transferred.
2413 *
2414 * @returns Next task segment handle.
2415 * @param pTaskSeg Task segment to complete.
2416 * @param rc Status code to set.
2417 */
2418static PPDMBLKCACHEWAITER pdmBlkCacheWaiterComplete(PPDMBLKCACHE pBlkCache,
2419 PPDMBLKCACHEWAITER pWaiter,
2420 int rc)
2421{
2422 PPDMBLKCACHEWAITER pNext = pWaiter->pNext;
2423 PPDMBLKCACHEREQ pReq = pWaiter->pReq;
2424
2425 pdmBlkCacheReqUpdate(pBlkCache, pWaiter->pReq, rc, true);
2426
2427 RTMemFree(pWaiter);
2428
2429 return pNext;
2430}
2431
2432static void pdmBlkCacheIoXferCompleteEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2433{
2434 PPDMBLKCACHEENTRY pEntry = hIoXfer->pEntry;
2435 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2436
2437 /* Reference the entry now as we are clearing the I/O in progress flag
2438 * which protected the entry till now. */
2439 pdmBlkCacheEntryRef(pEntry);
2440
2441 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2442 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
2443
2444 /* Process waiting segment list. The data in entry might have changed in-between. */
2445 bool fDirty = false;
2446 PPDMBLKCACHEWAITER pComplete = pEntry->pWaitingHead;
2447 PPDMBLKCACHEWAITER pCurr = pComplete;
2448
2449 AssertMsg((pCurr && pEntry->pWaitingTail) || (!pCurr && !pEntry->pWaitingTail),
2450 ("The list tail was not updated correctly\n"));
2451 pEntry->pWaitingTail = NULL;
2452 pEntry->pWaitingHead = NULL;
2453
2454 if (hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_WRITE)
2455 {
2456 /*
2457 * An error here is difficult to handle as the original request completed already.
2458 * The error is logged for now and the VM is paused.
2459 * If the user continues the entry is written again in the hope
2460 * the user fixed the problem and the next write succeeds.
2461 */
2462 if (RT_FAILURE(rcIoXfer))
2463 {
2464 LogRel(("I/O cache: Error while writing entry at offset %llu (%u bytes) to medium \"%s\" (rc=%Rrc)\n",
2465 pEntry->Core.Key, pEntry->cbData, pBlkCache->pszId, rcIoXfer));
2466
2467 if (!ASMAtomicXchgBool(&pCache->fIoErrorVmSuspended, true))
2468 {
2469 int rc = VMSetRuntimeError(pCache->pVM, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "BLKCACHE_IOERR",
2470 N_("The I/O cache encountered an error while updating data in medium \"%s\" (rc=%Rrc). "
2471 "Make sure there is enough free space on the disk and that the disk is working properly. "
2472 "Operation can be resumed afterwards"),
2473 pBlkCache->pszId, rcIoXfer);
2474 AssertRC(rc);
2475 }
2476
2477 /* Mark the entry as dirty again to get it added to the list later on. */
2478 fDirty = true;
2479 }
2480
2481 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IS_DIRTY;
2482
2483 while (pCurr)
2484 {
2485 AssertMsg(pCurr->fWrite, ("Completed write entries should never have read tasks attached\n"));
2486
2487 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2488 fDirty = true;
2489 pCurr = pCurr->pNext;
2490 }
2491 }
2492 else
2493 {
2494 AssertMsg(hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_READ, ("Invalid transfer type\n"));
2495 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY),
2496 ("Invalid flags set\n"));
2497
2498 while (pCurr)
2499 {
2500 if (pCurr->fWrite)
2501 {
2502 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2503 fDirty = true;
2504 }
2505 else
2506 RTSgBufCopyFromBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2507
2508 pCurr = pCurr->pNext;
2509 }
2510 }
2511
2512 bool fCommit = false;
2513 if (fDirty)
2514 fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2515
2516 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2517
2518 /* Dereference so that it isn't protected anymore except we issued anyother write for it. */
2519 pdmBlkCacheEntryRelease(pEntry);
2520
2521 if (fCommit)
2522 pdmBlkCacheCommitDirtyEntries(pCache);
2523
2524 /* Complete waiters now. */
2525 while (pComplete)
2526 pComplete = pdmBlkCacheWaiterComplete(pBlkCache, pComplete, rcIoXfer);
2527}
2528
2529VMMR3DECL(void) PDMR3BlkCacheIoXferComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2530{
2531 LogFlowFunc(("pBlkCache=%#p hIoXfer=%#p rcIoXfer=%Rrc\n", pBlkCache, hIoXfer, rcIoXfer));
2532
2533 if (hIoXfer->fIoCache)
2534 pdmBlkCacheIoXferCompleteEntry(pBlkCache, hIoXfer, rcIoXfer);
2535 else
2536 pdmBlkCacheReqUpdate(pBlkCache, hIoXfer->pReq, rcIoXfer, true);
2537 RTMemFree(hIoXfer);
2538}
2539
2540/**
2541 * Callback for the AVL do with all routine. Waits for a cachen entry to finish any pending I/O.
2542 *
2543 * @returns IPRT status code.
2544 * @param pNode The node to destroy.
2545 * @param pvUser Opaque user data.
2546 */
2547static int pdmBlkCacheEntryQuiesce(PAVLRU64NODECORE pNode, void *pvUser)
2548{
2549 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)pNode;
2550 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
2551
2552 while (ASMAtomicReadU32(&pEntry->fFlags) & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS)
2553 {
2554 /* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
2555 pdmBlkCacheEntryRef(pEntry);
2556 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2557
2558 RTThreadSleep(1);
2559
2560 /* Re-enter all locks and drop the reference. */
2561 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2562 pdmBlkCacheEntryRelease(pEntry);
2563 }
2564
2565 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
2566 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
2567
2568 return VINF_SUCCESS;
2569}
2570
2571VMMR3DECL(int) PDMR3BlkCacheSuspend(PPDMBLKCACHE pBlkCache)
2572{
2573 int rc = VINF_SUCCESS;
2574 LogFlowFunc(("pBlkCache=%#p\n", pBlkCache));
2575
2576 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2577
2578 if (!ASMAtomicReadBool(&pBlkCache->pCache->fIoErrorVmSuspended))
2579 pdmBlkCacheCommit(pBlkCache); /* Can issue new I/O requests. */
2580 ASMAtomicXchgBool(&pBlkCache->fSuspended, true);
2581
2582 /* Wait for all I/O to complete. */
2583 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2584 rc = RTAvlrU64DoWithAll(pBlkCache->pTree, true, pdmBlkCacheEntryQuiesce, NULL);
2585 AssertRC(rc);
2586 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2587
2588 return rc;
2589}
2590
2591VMMR3DECL(int) PDMR3BlkCacheResume(PPDMBLKCACHE pBlkCache)
2592{
2593 LogFlowFunc(("pBlkCache=%#p\n", pBlkCache));
2594
2595 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2596
2597 ASMAtomicXchgBool(&pBlkCache->fSuspended, false);
2598
2599 return VINF_SUCCESS;
2600}
2601
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