VirtualBox

source: vbox/trunk/src/VBox/VMM/VMMR3/PDMBlkCache.cpp@ 86649

Last change on this file since 86649 was 86527, checked in by vboxsync, 4 years ago

VMMR3/PDMBlkCache: Get rid of duplicate doxygen comment
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 100.1 KB
Line 
1/* $Id: PDMBlkCache.cpp 86527 2020-10-11 18:41:41Z vboxsync $ */
2/** @file
3 * PDM Block Cache.
4 */
5
6/*
7 * Copyright (C) 2006-2020 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/*********************************************************************************************************************************
24* Header Files *
25*********************************************************************************************************************************/
26#define LOG_GROUP LOG_GROUP_PDM_BLK_CACHE
27#include "PDMInternal.h"
28#include <iprt/asm.h>
29#include <iprt/mem.h>
30#include <iprt/path.h>
31#include <iprt/string.h>
32#include <iprt/trace.h>
33#include <VBox/log.h>
34#include <VBox/vmm/stam.h>
35#include <VBox/vmm/uvm.h>
36#include <VBox/vmm/vm.h>
37
38#include "PDMBlkCacheInternal.h"
39
40
41/*********************************************************************************************************************************
42* Defined Constants And Macros *
43*********************************************************************************************************************************/
44#ifdef VBOX_STRICT
45# define PDMACFILECACHE_IS_CRITSECT_OWNER(Cache) \
46 do \
47 { \
48 AssertMsg(RTCritSectIsOwner(&Cache->CritSect), \
49 ("Thread does not own critical section\n"));\
50 } while (0)
51
52# define PDMACFILECACHE_EP_IS_SEMRW_WRITE_OWNER(pEpCache) \
53 do \
54 { \
55 AssertMsg(RTSemRWIsWriteOwner(pEpCache->SemRWEntries), \
56 ("Thread is not exclusive owner of the per endpoint RW semaphore\n")); \
57 } while (0)
58
59# define PDMACFILECACHE_EP_IS_SEMRW_READ_OWNER(pEpCache) \
60 do \
61 { \
62 AssertMsg(RTSemRWIsReadOwner(pEpCache->SemRWEntries), \
63 ("Thread is not read owner of the per endpoint RW semaphore\n")); \
64 } while (0)
65
66#else
67# define PDMACFILECACHE_IS_CRITSECT_OWNER(Cache) do { } while (0)
68# define PDMACFILECACHE_EP_IS_SEMRW_WRITE_OWNER(pEpCache) do { } while (0)
69# define PDMACFILECACHE_EP_IS_SEMRW_READ_OWNER(pEpCache) do { } while (0)
70#endif
71
72#define PDM_BLK_CACHE_SAVED_STATE_VERSION 1
73
74/* Enable to enable some tracing in the block cache code for investigating issues. */
75/*#define VBOX_BLKCACHE_TRACING 1*/
76
77
78/*********************************************************************************************************************************
79* Internal Functions *
80*********************************************************************************************************************************/
81
82static PPDMBLKCACHEENTRY pdmBlkCacheEntryAlloc(PPDMBLKCACHE pBlkCache,
83 uint64_t off, size_t cbData, uint8_t *pbBuffer);
84static bool pdmBlkCacheAddDirtyEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry);
85
86
87/**
88 * Add message to the VM trace buffer.
89 *
90 * @returns nothing.
91 * @param pBlkCache The block cache.
92 * @param pszFmt The format string.
93 * @param ... Additional parameters for the string formatter.
94 */
95DECLINLINE(void) pdmBlkCacheR3TraceMsgF(PPDMBLKCACHE pBlkCache, const char *pszFmt, ...)
96{
97#if defined(VBOX_BLKCACHE_TRACING)
98 va_list va;
99 va_start(va, pszFmt);
100 RTTraceBufAddMsgV(pBlkCache->pCache->pVM->CTX_SUFF(hTraceBuf), pszFmt, va);
101 va_end(va);
102#else
103 RT_NOREF2(pBlkCache, pszFmt);
104#endif
105}
106
107/**
108 * Decrement the reference counter of the given cache entry.
109 *
110 * @returns nothing.
111 * @param pEntry The entry to release.
112 */
113DECLINLINE(void) pdmBlkCacheEntryRelease(PPDMBLKCACHEENTRY pEntry)
114{
115 AssertMsg(pEntry->cRefs > 0, ("Trying to release a not referenced entry\n"));
116 ASMAtomicDecU32(&pEntry->cRefs);
117}
118
119/**
120 * Increment the reference counter of the given cache entry.
121 *
122 * @returns nothing.
123 * @param pEntry The entry to reference.
124 */
125DECLINLINE(void) pdmBlkCacheEntryRef(PPDMBLKCACHEENTRY pEntry)
126{
127 ASMAtomicIncU32(&pEntry->cRefs);
128}
129
130#ifdef VBOX_STRICT
131static void pdmBlkCacheValidate(PPDMBLKCACHEGLOBAL pCache)
132{
133 /* Amount of cached data should never exceed the maximum amount. */
134 AssertMsg(pCache->cbCached <= pCache->cbMax,
135 ("Current amount of cached data exceeds maximum\n"));
136
137 /* The amount of cached data in the LRU and FRU list should match cbCached */
138 AssertMsg(pCache->LruRecentlyUsedIn.cbCached + pCache->LruFrequentlyUsed.cbCached == pCache->cbCached,
139 ("Amount of cached data doesn't match\n"));
140
141 AssertMsg(pCache->LruRecentlyUsedOut.cbCached <= pCache->cbRecentlyUsedOutMax,
142 ("Paged out list exceeds maximum\n"));
143}
144#endif
145
146DECLINLINE(void) pdmBlkCacheLockEnter(PPDMBLKCACHEGLOBAL pCache)
147{
148 RTCritSectEnter(&pCache->CritSect);
149#ifdef VBOX_STRICT
150 pdmBlkCacheValidate(pCache);
151#endif
152}
153
154DECLINLINE(void) pdmBlkCacheLockLeave(PPDMBLKCACHEGLOBAL pCache)
155{
156#ifdef VBOX_STRICT
157 pdmBlkCacheValidate(pCache);
158#endif
159 RTCritSectLeave(&pCache->CritSect);
160}
161
162DECLINLINE(void) pdmBlkCacheSub(PPDMBLKCACHEGLOBAL pCache, uint32_t cbAmount)
163{
164 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
165 pCache->cbCached -= cbAmount;
166}
167
168DECLINLINE(void) pdmBlkCacheAdd(PPDMBLKCACHEGLOBAL pCache, uint32_t cbAmount)
169{
170 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
171 pCache->cbCached += cbAmount;
172}
173
174DECLINLINE(void) pdmBlkCacheListAdd(PPDMBLKLRULIST pList, uint32_t cbAmount)
175{
176 pList->cbCached += cbAmount;
177}
178
179DECLINLINE(void) pdmBlkCacheListSub(PPDMBLKLRULIST pList, uint32_t cbAmount)
180{
181 pList->cbCached -= cbAmount;
182}
183
184#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
185/**
186 * Checks consistency of a LRU list.
187 *
188 * @returns nothing
189 * @param pList The LRU list to check.
190 * @param pNotInList Element which is not allowed to occur in the list.
191 */
192static void pdmBlkCacheCheckList(PPDMBLKLRULIST pList, PPDMBLKCACHEENTRY pNotInList)
193{
194 PPDMBLKCACHEENTRY pCurr = pList->pHead;
195
196 /* Check that there are no double entries and no cycles in the list. */
197 while (pCurr)
198 {
199 PPDMBLKCACHEENTRY pNext = pCurr->pNext;
200
201 while (pNext)
202 {
203 AssertMsg(pCurr != pNext,
204 ("Entry %#p is at least two times in list %#p or there is a cycle in the list\n",
205 pCurr, pList));
206 pNext = pNext->pNext;
207 }
208
209 AssertMsg(pCurr != pNotInList, ("Not allowed entry %#p is in list\n", pCurr));
210
211 if (!pCurr->pNext)
212 AssertMsg(pCurr == pList->pTail, ("End of list reached but last element is not list tail\n"));
213
214 pCurr = pCurr->pNext;
215 }
216}
217#endif
218
219/**
220 * Unlinks a cache entry from the LRU list it is assigned to.
221 *
222 * @returns nothing.
223 * @param pEntry The entry to unlink.
224 */
225static void pdmBlkCacheEntryRemoveFromList(PPDMBLKCACHEENTRY pEntry)
226{
227 PPDMBLKLRULIST pList = pEntry->pList;
228 PPDMBLKCACHEENTRY pPrev, pNext;
229
230 LogFlowFunc((": Deleting entry %#p from list %#p\n", pEntry, pList));
231
232 AssertPtr(pList);
233
234#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
235 pdmBlkCacheCheckList(pList, NULL);
236#endif
237
238 pPrev = pEntry->pPrev;
239 pNext = pEntry->pNext;
240
241 AssertMsg(pEntry != pPrev, ("Entry links to itself as previous element\n"));
242 AssertMsg(pEntry != pNext, ("Entry links to itself as next element\n"));
243
244 if (pPrev)
245 pPrev->pNext = pNext;
246 else
247 {
248 pList->pHead = pNext;
249
250 if (pNext)
251 pNext->pPrev = NULL;
252 }
253
254 if (pNext)
255 pNext->pPrev = pPrev;
256 else
257 {
258 pList->pTail = pPrev;
259
260 if (pPrev)
261 pPrev->pNext = NULL;
262 }
263
264 pEntry->pList = NULL;
265 pEntry->pPrev = NULL;
266 pEntry->pNext = NULL;
267 pdmBlkCacheListSub(pList, pEntry->cbData);
268#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
269 pdmBlkCacheCheckList(pList, pEntry);
270#endif
271}
272
273/**
274 * Adds a cache entry to the given LRU list unlinking it from the currently
275 * assigned list if needed.
276 *
277 * @returns nothing.
278 * @param pList List to the add entry to.
279 * @param pEntry Entry to add.
280 */
281static void pdmBlkCacheEntryAddToList(PPDMBLKLRULIST pList, PPDMBLKCACHEENTRY pEntry)
282{
283 LogFlowFunc((": Adding entry %#p to list %#p\n", pEntry, pList));
284#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
285 pdmBlkCacheCheckList(pList, NULL);
286#endif
287
288 /* Remove from old list if needed */
289 if (pEntry->pList)
290 pdmBlkCacheEntryRemoveFromList(pEntry);
291
292 pEntry->pNext = pList->pHead;
293 if (pList->pHead)
294 pList->pHead->pPrev = pEntry;
295 else
296 {
297 Assert(!pList->pTail);
298 pList->pTail = pEntry;
299 }
300
301 pEntry->pPrev = NULL;
302 pList->pHead = pEntry;
303 pdmBlkCacheListAdd(pList, pEntry->cbData);
304 pEntry->pList = pList;
305#ifdef PDMACFILECACHE_WITH_LRULIST_CHECKS
306 pdmBlkCacheCheckList(pList, NULL);
307#endif
308}
309
310/**
311 * Destroys a LRU list freeing all entries.
312 *
313 * @returns nothing
314 * @param pList Pointer to the LRU list to destroy.
315 *
316 * @note The caller must own the critical section of the cache.
317 */
318static void pdmBlkCacheDestroyList(PPDMBLKLRULIST pList)
319{
320 while (pList->pHead)
321 {
322 PPDMBLKCACHEENTRY pEntry = pList->pHead;
323
324 pList->pHead = pEntry->pNext;
325
326 AssertMsg(!(pEntry->fFlags & (PDMBLKCACHE_ENTRY_IO_IN_PROGRESS | PDMBLKCACHE_ENTRY_IS_DIRTY)),
327 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
328
329 RTMemPageFree(pEntry->pbData, pEntry->cbData);
330 RTMemFree(pEntry);
331 }
332}
333
334/**
335 * Tries to remove the given amount of bytes from a given list in the cache
336 * moving the entries to one of the given ghosts lists
337 *
338 * @returns Amount of data which could be freed.
339 * @param pCache Pointer to the global cache data.
340 * @param cbData The amount of the data to free.
341 * @param pListSrc The source list to evict data from.
342 * @param pGhostListDst Where the ghost list removed entries should be
343 * moved to, NULL if the entry should be freed.
344 * @param fReuseBuffer Flag whether a buffer should be reused if it has
345 * the same size
346 * @param ppbBuffer Where to store the address of the buffer if an
347 * entry with the same size was found and
348 * fReuseBuffer is true.
349 *
350 * @note This function may return fewer bytes than requested because entries
351 * may be marked as non evictable if they are used for I/O at the
352 * moment.
353 */
354static size_t pdmBlkCacheEvictPagesFrom(PPDMBLKCACHEGLOBAL pCache, size_t cbData,
355 PPDMBLKLRULIST pListSrc, PPDMBLKLRULIST pGhostListDst,
356 bool fReuseBuffer, uint8_t **ppbBuffer)
357{
358 size_t cbEvicted = 0;
359
360 PDMACFILECACHE_IS_CRITSECT_OWNER(pCache);
361
362 AssertMsg(cbData > 0, ("Evicting 0 bytes not possible\n"));
363 AssertMsg( !pGhostListDst
364 || (pGhostListDst == &pCache->LruRecentlyUsedOut),
365 ("Destination list must be NULL or the recently used but paged out list\n"));
366
367 if (fReuseBuffer)
368 {
369 AssertPtr(ppbBuffer);
370 *ppbBuffer = NULL;
371 }
372
373 /* Start deleting from the tail. */
374 PPDMBLKCACHEENTRY pEntry = pListSrc->pTail;
375
376 while ((cbEvicted < cbData) && pEntry)
377 {
378 PPDMBLKCACHEENTRY pCurr = pEntry;
379
380 pEntry = pEntry->pPrev;
381
382 /* We can't evict pages which are currently in progress or dirty but not in progress */
383 if ( !(pCurr->fFlags & PDMBLKCACHE_NOT_EVICTABLE)
384 && (ASMAtomicReadU32(&pCurr->cRefs) == 0))
385 {
386 /* Ok eviction candidate. Grab the endpoint semaphore and check again
387 * because somebody else might have raced us. */
388 PPDMBLKCACHE pBlkCache = pCurr->pBlkCache;
389 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
390
391 if (!(pCurr->fFlags & PDMBLKCACHE_NOT_EVICTABLE)
392 && (ASMAtomicReadU32(&pCurr->cRefs) == 0))
393 {
394 LogFlow(("Evicting entry %#p (%u bytes)\n", pCurr, pCurr->cbData));
395
396 if (fReuseBuffer && pCurr->cbData == cbData)
397 {
398 STAM_COUNTER_INC(&pCache->StatBuffersReused);
399 *ppbBuffer = pCurr->pbData;
400 }
401 else if (pCurr->pbData)
402 RTMemPageFree(pCurr->pbData, pCurr->cbData);
403
404 pCurr->pbData = NULL;
405 cbEvicted += pCurr->cbData;
406
407 pdmBlkCacheEntryRemoveFromList(pCurr);
408 pdmBlkCacheSub(pCache, pCurr->cbData);
409
410 if (pGhostListDst)
411 {
412 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
413
414 PPDMBLKCACHEENTRY pGhostEntFree = pGhostListDst->pTail;
415
416 /* We have to remove the last entries from the paged out list. */
417 while ( pGhostListDst->cbCached + pCurr->cbData > pCache->cbRecentlyUsedOutMax
418 && pGhostEntFree)
419 {
420 PPDMBLKCACHEENTRY pFree = pGhostEntFree;
421 PPDMBLKCACHE pBlkCacheFree = pFree->pBlkCache;
422
423 pGhostEntFree = pGhostEntFree->pPrev;
424
425 RTSemRWRequestWrite(pBlkCacheFree->SemRWEntries, RT_INDEFINITE_WAIT);
426
427 if (ASMAtomicReadU32(&pFree->cRefs) == 0)
428 {
429 pdmBlkCacheEntryRemoveFromList(pFree);
430
431 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
432 RTAvlrU64Remove(pBlkCacheFree->pTree, pFree->Core.Key);
433 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
434
435 RTMemFree(pFree);
436 }
437
438 RTSemRWReleaseWrite(pBlkCacheFree->SemRWEntries);
439 }
440
441 if (pGhostListDst->cbCached + pCurr->cbData > pCache->cbRecentlyUsedOutMax)
442 {
443 /* Couldn't remove enough entries. Delete */
444 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
445 RTAvlrU64Remove(pCurr->pBlkCache->pTree, pCurr->Core.Key);
446 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
447
448 RTMemFree(pCurr);
449 }
450 else
451 pdmBlkCacheEntryAddToList(pGhostListDst, pCurr);
452 }
453 else
454 {
455 /* Delete the entry from the AVL tree it is assigned to. */
456 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
457 RTAvlrU64Remove(pCurr->pBlkCache->pTree, pCurr->Core.Key);
458 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
459
460 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
461 RTMemFree(pCurr);
462 }
463 }
464
465 }
466 else
467 LogFlow(("Entry %#p (%u bytes) is still in progress and can't be evicted\n", pCurr, pCurr->cbData));
468 }
469
470 return cbEvicted;
471}
472
473static bool pdmBlkCacheReclaim(PPDMBLKCACHEGLOBAL pCache, size_t cbData, bool fReuseBuffer, uint8_t **ppbBuffer)
474{
475 size_t cbRemoved = 0;
476
477 if ((pCache->cbCached + cbData) < pCache->cbMax)
478 return true;
479 else if ((pCache->LruRecentlyUsedIn.cbCached + cbData) > pCache->cbRecentlyUsedInMax)
480 {
481 /* Try to evict as many bytes as possible from A1in */
482 cbRemoved = pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruRecentlyUsedIn,
483 &pCache->LruRecentlyUsedOut, fReuseBuffer, ppbBuffer);
484
485 /*
486 * If it was not possible to remove enough entries
487 * try the frequently accessed cache.
488 */
489 if (cbRemoved < cbData)
490 {
491 Assert(!fReuseBuffer || !*ppbBuffer); /* It is not possible that we got a buffer with the correct size but we didn't freed enough data. */
492
493 /*
494 * If we removed something we can't pass the reuse buffer flag anymore because
495 * we don't need to evict that much data
496 */
497 if (!cbRemoved)
498 cbRemoved += pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruFrequentlyUsed,
499 NULL, fReuseBuffer, ppbBuffer);
500 else
501 cbRemoved += pdmBlkCacheEvictPagesFrom(pCache, cbData - cbRemoved, &pCache->LruFrequentlyUsed,
502 NULL, false, NULL);
503 }
504 }
505 else
506 {
507 /* We have to remove entries from frequently access list. */
508 cbRemoved = pdmBlkCacheEvictPagesFrom(pCache, cbData, &pCache->LruFrequentlyUsed,
509 NULL, fReuseBuffer, ppbBuffer);
510 }
511
512 LogFlowFunc((": removed %u bytes, requested %u\n", cbRemoved, cbData));
513 return (cbRemoved >= cbData);
514}
515
516DECLINLINE(int) pdmBlkCacheEnqueue(PPDMBLKCACHE pBlkCache, uint64_t off, size_t cbXfer, PPDMBLKCACHEIOXFER pIoXfer)
517{
518 int rc = VINF_SUCCESS;
519
520 LogFlowFunc(("%s: Enqueuing hIoXfer=%#p enmXferDir=%d\n",
521 __FUNCTION__, pIoXfer, pIoXfer->enmXferDir));
522
523 ASMAtomicIncU32(&pBlkCache->cIoXfersActive);
524 pdmBlkCacheR3TraceMsgF(pBlkCache, "BlkCache: I/O req %#p (%RTbool , %d) queued (%u now active)",
525 pIoXfer, pIoXfer->fIoCache, pIoXfer->enmXferDir, pBlkCache->cIoXfersActive);
526
527 switch (pBlkCache->enmType)
528 {
529 case PDMBLKCACHETYPE_DEV:
530 {
531 rc = pBlkCache->u.Dev.pfnXferEnqueue(pBlkCache->u.Dev.pDevIns,
532 pIoXfer->enmXferDir,
533 off, cbXfer,
534 &pIoXfer->SgBuf, pIoXfer);
535 break;
536 }
537 case PDMBLKCACHETYPE_DRV:
538 {
539 rc = pBlkCache->u.Drv.pfnXferEnqueue(pBlkCache->u.Drv.pDrvIns,
540 pIoXfer->enmXferDir,
541 off, cbXfer,
542 &pIoXfer->SgBuf, pIoXfer);
543 break;
544 }
545 case PDMBLKCACHETYPE_USB:
546 {
547 rc = pBlkCache->u.Usb.pfnXferEnqueue(pBlkCache->u.Usb.pUsbIns,
548 pIoXfer->enmXferDir,
549 off, cbXfer,
550 &pIoXfer->SgBuf, pIoXfer);
551 break;
552 }
553 case PDMBLKCACHETYPE_INTERNAL:
554 {
555 rc = pBlkCache->u.Int.pfnXferEnqueue(pBlkCache->u.Int.pvUser,
556 pIoXfer->enmXferDir,
557 off, cbXfer,
558 &pIoXfer->SgBuf, pIoXfer);
559 break;
560 }
561 default:
562 AssertMsgFailed(("Unknown block cache type!\n"));
563 }
564
565 if (RT_FAILURE(rc))
566 {
567 pdmBlkCacheR3TraceMsgF(pBlkCache, "BlkCache: Queueing I/O req %#p failed %Rrc", pIoXfer, rc);
568 ASMAtomicDecU32(&pBlkCache->cIoXfersActive);
569 }
570
571 LogFlowFunc(("%s: returns rc=%Rrc\n", __FUNCTION__, rc));
572 return rc;
573}
574
575/**
576 * Initiates a read I/O task for the given entry.
577 *
578 * @returns VBox status code.
579 * @param pEntry The entry to fetch the data to.
580 */
581static int pdmBlkCacheEntryReadFromMedium(PPDMBLKCACHEENTRY pEntry)
582{
583 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
584 LogFlowFunc((": Reading data into cache entry %#p\n", pEntry));
585
586 /* Make sure no one evicts the entry while it is accessed. */
587 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
588
589 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
590 if (RT_UNLIKELY(!pIoXfer))
591 return VERR_NO_MEMORY;
592
593 AssertMsg(pEntry->pbData, ("Entry is in ghost state\n"));
594
595 pIoXfer->fIoCache = true;
596 pIoXfer->pEntry = pEntry;
597 pIoXfer->SgSeg.pvSeg = pEntry->pbData;
598 pIoXfer->SgSeg.cbSeg = pEntry->cbData;
599 pIoXfer->enmXferDir = PDMBLKCACHEXFERDIR_READ;
600 RTSgBufInit(&pIoXfer->SgBuf, &pIoXfer->SgSeg, 1);
601
602 return pdmBlkCacheEnqueue(pBlkCache, pEntry->Core.Key, pEntry->cbData, pIoXfer);
603}
604
605/**
606 * Initiates a write I/O task for the given entry.
607 *
608 * @returns nothing.
609 * @param pEntry The entry to read the data from.
610 */
611static int pdmBlkCacheEntryWriteToMedium(PPDMBLKCACHEENTRY pEntry)
612{
613 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
614 LogFlowFunc((": Writing data from cache entry %#p\n", pEntry));
615
616 /* Make sure no one evicts the entry while it is accessed. */
617 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
618
619 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
620 if (RT_UNLIKELY(!pIoXfer))
621 return VERR_NO_MEMORY;
622
623 AssertMsg(pEntry->pbData, ("Entry is in ghost state\n"));
624
625 pIoXfer->fIoCache = true;
626 pIoXfer->pEntry = pEntry;
627 pIoXfer->SgSeg.pvSeg = pEntry->pbData;
628 pIoXfer->SgSeg.cbSeg = pEntry->cbData;
629 pIoXfer->enmXferDir = PDMBLKCACHEXFERDIR_WRITE;
630 RTSgBufInit(&pIoXfer->SgBuf, &pIoXfer->SgSeg, 1);
631
632 return pdmBlkCacheEnqueue(pBlkCache, pEntry->Core.Key, pEntry->cbData, pIoXfer);
633}
634
635/**
636 * Passthrough a part of a request directly to the I/O manager handling the
637 * endpoint.
638 *
639 * @returns VBox status code.
640 * @param pBlkCache The endpoint cache.
641 * @param pReq The request.
642 * @param pSgBuf The scatter/gather buffer.
643 * @param offStart Offset to start transfer from.
644 * @param cbData Amount of data to transfer.
645 * @param enmXferDir The transfer type (read/write)
646 */
647static int pdmBlkCacheRequestPassthrough(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq,
648 PRTSGBUF pSgBuf, uint64_t offStart, size_t cbData,
649 PDMBLKCACHEXFERDIR enmXferDir)
650{
651
652 PPDMBLKCACHEIOXFER pIoXfer = (PPDMBLKCACHEIOXFER)RTMemAllocZ(sizeof(PDMBLKCACHEIOXFER));
653 if (RT_UNLIKELY(!pIoXfer))
654 return VERR_NO_MEMORY;
655
656 ASMAtomicIncU32(&pReq->cXfersPending);
657 pIoXfer->fIoCache = false;
658 pIoXfer->pReq = pReq;
659 pIoXfer->enmXferDir = enmXferDir;
660 if (pSgBuf)
661 {
662 RTSgBufClone(&pIoXfer->SgBuf, pSgBuf);
663 RTSgBufAdvance(pSgBuf, cbData);
664 }
665
666 return pdmBlkCacheEnqueue(pBlkCache, offStart, cbData, pIoXfer);
667}
668
669/**
670 * Commit a single dirty entry to the endpoint
671 *
672 * @returns nothing
673 * @param pEntry The entry to commit.
674 */
675static void pdmBlkCacheEntryCommit(PPDMBLKCACHEENTRY pEntry)
676{
677 AssertMsg( (pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY)
678 && !(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
679 ("Invalid flags set for entry %#p\n", pEntry));
680
681 pdmBlkCacheEntryWriteToMedium(pEntry);
682}
683
684/**
685 * Commit all dirty entries for a single endpoint.
686 *
687 * @returns nothing.
688 * @param pBlkCache The endpoint cache to commit.
689 */
690static void pdmBlkCacheCommit(PPDMBLKCACHE pBlkCache)
691{
692 uint32_t cbCommitted = 0;
693
694 /* Return if the cache was suspended. */
695 if (pBlkCache->fSuspended)
696 return;
697
698 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
699
700 /* The list is moved to a new header to reduce locking overhead. */
701 RTLISTANCHOR ListDirtyNotCommitted;
702
703 RTSpinlockAcquire(pBlkCache->LockList);
704 RTListMove(&ListDirtyNotCommitted, &pBlkCache->ListDirtyNotCommitted);
705 RTSpinlockRelease(pBlkCache->LockList);
706
707 if (!RTListIsEmpty(&ListDirtyNotCommitted))
708 {
709 PPDMBLKCACHEENTRY pEntry = RTListGetFirst(&ListDirtyNotCommitted, PDMBLKCACHEENTRY, NodeNotCommitted);
710
711 while (!RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted))
712 {
713 PPDMBLKCACHEENTRY pNext = RTListNodeGetNext(&pEntry->NodeNotCommitted, PDMBLKCACHEENTRY,
714 NodeNotCommitted);
715 pdmBlkCacheEntryCommit(pEntry);
716 cbCommitted += pEntry->cbData;
717 RTListNodeRemove(&pEntry->NodeNotCommitted);
718 pEntry = pNext;
719 }
720
721 /* Commit the last endpoint */
722 Assert(RTListNodeIsLast(&ListDirtyNotCommitted, &pEntry->NodeNotCommitted));
723 pdmBlkCacheEntryCommit(pEntry);
724 cbCommitted += pEntry->cbData;
725 RTListNodeRemove(&pEntry->NodeNotCommitted);
726 AssertMsg(RTListIsEmpty(&ListDirtyNotCommitted),
727 ("Committed all entries but list is not empty\n"));
728 }
729
730 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
731 AssertMsg(pBlkCache->pCache->cbDirty >= cbCommitted,
732 ("Number of committed bytes exceeds number of dirty bytes\n"));
733 uint32_t cbDirtyOld = ASMAtomicSubU32(&pBlkCache->pCache->cbDirty, cbCommitted);
734
735 /* Reset the commit timer if we don't have any dirty bits. */
736 if ( !(cbDirtyOld - cbCommitted)
737 && pBlkCache->pCache->u32CommitTimeoutMs != 0)
738 TMTimerStop(pBlkCache->pCache->pTimerCommit);
739}
740
741/**
742 * Commit all dirty entries in the cache.
743 *
744 * @returns nothing.
745 * @param pCache The global cache instance.
746 */
747static void pdmBlkCacheCommitDirtyEntries(PPDMBLKCACHEGLOBAL pCache)
748{
749 bool fCommitInProgress = ASMAtomicXchgBool(&pCache->fCommitInProgress, true);
750
751 if (!fCommitInProgress)
752 {
753 pdmBlkCacheLockEnter(pCache);
754 Assert(!RTListIsEmpty(&pCache->ListUsers));
755
756 PPDMBLKCACHE pBlkCache = RTListGetFirst(&pCache->ListUsers, PDMBLKCACHE, NodeCacheUser);
757 AssertPtr(pBlkCache);
758
759 while (!RTListNodeIsLast(&pCache->ListUsers, &pBlkCache->NodeCacheUser))
760 {
761 pdmBlkCacheCommit(pBlkCache);
762
763 pBlkCache = RTListNodeGetNext(&pBlkCache->NodeCacheUser, PDMBLKCACHE,
764 NodeCacheUser);
765 }
766
767 /* Commit the last endpoint */
768 Assert(RTListNodeIsLast(&pCache->ListUsers, &pBlkCache->NodeCacheUser));
769 pdmBlkCacheCommit(pBlkCache);
770
771 pdmBlkCacheLockLeave(pCache);
772 ASMAtomicWriteBool(&pCache->fCommitInProgress, false);
773 }
774}
775
776/**
777 * Adds the given entry as a dirty to the cache.
778 *
779 * @returns Flag whether the amount of dirty bytes in the cache exceeds the threshold
780 * @param pBlkCache The endpoint cache the entry belongs to.
781 * @param pEntry The entry to add.
782 */
783static bool pdmBlkCacheAddDirtyEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
784{
785 bool fDirtyBytesExceeded = false;
786 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
787
788 /* If the commit timer is disabled we commit right away. */
789 if (pCache->u32CommitTimeoutMs == 0)
790 {
791 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IS_DIRTY;
792 pdmBlkCacheEntryCommit(pEntry);
793 }
794 else if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY))
795 {
796 pEntry->fFlags |= PDMBLKCACHE_ENTRY_IS_DIRTY;
797
798 RTSpinlockAcquire(pBlkCache->LockList);
799 RTListAppend(&pBlkCache->ListDirtyNotCommitted, &pEntry->NodeNotCommitted);
800 RTSpinlockRelease(pBlkCache->LockList);
801
802 uint32_t cbDirty = ASMAtomicAddU32(&pCache->cbDirty, pEntry->cbData);
803
804 /* Prevent committing if the VM was suspended. */
805 if (RT_LIKELY(!ASMAtomicReadBool(&pCache->fIoErrorVmSuspended)))
806 fDirtyBytesExceeded = (cbDirty + pEntry->cbData >= pCache->cbCommitDirtyThreshold);
807 else if (!cbDirty && pCache->u32CommitTimeoutMs > 0)
808 {
809 /* Arm the commit timer. */
810 TMTimerSetMillies(pCache->pTimerCommit, pCache->u32CommitTimeoutMs);
811 }
812 }
813
814 return fDirtyBytesExceeded;
815}
816
817static PPDMBLKCACHE pdmR3BlkCacheFindById(PPDMBLKCACHEGLOBAL pBlkCacheGlobal, const char *pcszId)
818{
819 bool fFound = false;
820
821 PPDMBLKCACHE pBlkCache;
822 RTListForEach(&pBlkCacheGlobal->ListUsers, pBlkCache, PDMBLKCACHE, NodeCacheUser)
823 {
824 if (!RTStrCmp(pBlkCache->pszId, pcszId))
825 {
826 fFound = true;
827 break;
828 }
829 }
830
831 return fFound ? pBlkCache : NULL;
832}
833
834/**
835 * Commit timer callback.
836 */
837static DECLCALLBACK(void) pdmBlkCacheCommitTimerCallback(PVM pVM, PTMTIMER pTimer, void *pvUser)
838{
839 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
840 NOREF(pVM); NOREF(pTimer);
841
842 LogFlowFunc(("Commit interval expired, commiting dirty entries\n"));
843
844 if ( ASMAtomicReadU32(&pCache->cbDirty) > 0
845 && !ASMAtomicReadBool(&pCache->fIoErrorVmSuspended))
846 pdmBlkCacheCommitDirtyEntries(pCache);
847
848 LogFlowFunc(("Entries committed, going to sleep\n"));
849}
850
851static DECLCALLBACK(int) pdmR3BlkCacheSaveExec(PVM pVM, PSSMHANDLE pSSM)
852{
853 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
854
855 AssertPtr(pBlkCacheGlobal);
856
857 pdmBlkCacheLockEnter(pBlkCacheGlobal);
858
859 SSMR3PutU32(pSSM, pBlkCacheGlobal->cRefs);
860
861 /* Go through the list and save all dirty entries. */
862 PPDMBLKCACHE pBlkCache;
863 RTListForEach(&pBlkCacheGlobal->ListUsers, pBlkCache, PDMBLKCACHE, NodeCacheUser)
864 {
865 uint32_t cEntries = 0;
866 PPDMBLKCACHEENTRY pEntry;
867
868 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
869 SSMR3PutU32(pSSM, (uint32_t)strlen(pBlkCache->pszId));
870 SSMR3PutStrZ(pSSM, pBlkCache->pszId);
871
872 /* Count the number of entries to safe. */
873 RTListForEach(&pBlkCache->ListDirtyNotCommitted, pEntry, PDMBLKCACHEENTRY, NodeNotCommitted)
874 {
875 cEntries++;
876 }
877
878 SSMR3PutU32(pSSM, cEntries);
879
880 /* Walk the list of all dirty entries and save them. */
881 RTListForEach(&pBlkCache->ListDirtyNotCommitted, pEntry, PDMBLKCACHEENTRY, NodeNotCommitted)
882 {
883 /* A few sanity checks. */
884 AssertMsg(!pEntry->cRefs, ("The entry is still referenced\n"));
885 AssertMsg(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY, ("Entry is not dirty\n"));
886 AssertMsg(!(pEntry->fFlags & ~PDMBLKCACHE_ENTRY_IS_DIRTY), ("Invalid flags set\n"));
887 AssertMsg(!pEntry->pWaitingHead && !pEntry->pWaitingTail, ("There are waiting requests\n"));
888 AssertMsg( pEntry->pList == &pBlkCacheGlobal->LruRecentlyUsedIn
889 || pEntry->pList == &pBlkCacheGlobal->LruFrequentlyUsed,
890 ("Invalid list\n"));
891 AssertMsg(pEntry->cbData == pEntry->Core.KeyLast - pEntry->Core.Key + 1,
892 ("Size and range do not match\n"));
893
894 /* Save */
895 SSMR3PutU64(pSSM, pEntry->Core.Key);
896 SSMR3PutU32(pSSM, pEntry->cbData);
897 SSMR3PutMem(pSSM, pEntry->pbData, pEntry->cbData);
898 }
899
900 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
901 }
902
903 pdmBlkCacheLockLeave(pBlkCacheGlobal);
904
905 /* Terminator */
906 return SSMR3PutU32(pSSM, UINT32_MAX);
907}
908
909static DECLCALLBACK(int) pdmR3BlkCacheLoadExec(PVM pVM, PSSMHANDLE pSSM, uint32_t uVersion, uint32_t uPass)
910{
911 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
912 uint32_t cRefs;
913
914 NOREF(uPass);
915 AssertPtr(pBlkCacheGlobal);
916
917 pdmBlkCacheLockEnter(pBlkCacheGlobal);
918
919 if (uVersion != PDM_BLK_CACHE_SAVED_STATE_VERSION)
920 return VERR_SSM_UNSUPPORTED_DATA_UNIT_VERSION;
921
922 SSMR3GetU32(pSSM, &cRefs);
923
924 /*
925 * Fewer users in the saved state than in the current VM are allowed
926 * because that means that there are only new ones which don't have any saved state
927 * which can get lost.
928 * More saved state entries than registered cache users are only allowed if the
929 * missing users don't have any data saved in the cache.
930 */
931 int rc = VINF_SUCCESS;
932 char *pszId = NULL;
933
934 while ( cRefs > 0
935 && RT_SUCCESS(rc))
936 {
937 PPDMBLKCACHE pBlkCache = NULL;
938 uint32_t cbId = 0;
939
940 SSMR3GetU32(pSSM, &cbId);
941 Assert(cbId > 0);
942
943 cbId++; /* Include terminator */
944 pszId = (char *)RTMemAllocZ(cbId * sizeof(char));
945 if (!pszId)
946 {
947 rc = VERR_NO_MEMORY;
948 break;
949 }
950
951 rc = SSMR3GetStrZ(pSSM, pszId, cbId);
952 AssertRC(rc);
953
954 /* Search for the block cache with the provided id. */
955 pBlkCache = pdmR3BlkCacheFindById(pBlkCacheGlobal, pszId);
956
957 /* Get the entries */
958 uint32_t cEntries;
959 SSMR3GetU32(pSSM, &cEntries);
960
961 if (!pBlkCache && (cEntries > 0))
962 {
963 rc = SSMR3SetCfgError(pSSM, RT_SRC_POS,
964 N_("The VM is missing a block device and there is data in the cache. Please make sure the source and target VMs have compatible storage configurations"));
965 break;
966 }
967
968 RTMemFree(pszId);
969 pszId = NULL;
970
971 while (cEntries > 0)
972 {
973 PPDMBLKCACHEENTRY pEntry;
974 uint64_t off;
975 uint32_t cbEntry;
976
977 SSMR3GetU64(pSSM, &off);
978 SSMR3GetU32(pSSM, &cbEntry);
979
980 pEntry = pdmBlkCacheEntryAlloc(pBlkCache, off, cbEntry, NULL);
981 if (!pEntry)
982 {
983 rc = VERR_NO_MEMORY;
984 break;
985 }
986
987 rc = SSMR3GetMem(pSSM, pEntry->pbData, cbEntry);
988 if (RT_FAILURE(rc))
989 {
990 RTMemFree(pEntry->pbData);
991 RTMemFree(pEntry);
992 break;
993 }
994
995 /* Insert into the tree. */
996 bool fInserted = RTAvlrU64Insert(pBlkCache->pTree, &pEntry->Core);
997 Assert(fInserted); NOREF(fInserted);
998
999 /* Add to the dirty list. */
1000 pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
1001 pdmBlkCacheEntryAddToList(&pBlkCacheGlobal->LruRecentlyUsedIn, pEntry);
1002 pdmBlkCacheAdd(pBlkCacheGlobal, cbEntry);
1003 pdmBlkCacheEntryRelease(pEntry);
1004 cEntries--;
1005 }
1006
1007 cRefs--;
1008 }
1009
1010 if (pszId)
1011 RTMemFree(pszId);
1012
1013 if (cRefs && RT_SUCCESS(rc))
1014 rc = SSMR3SetCfgError(pSSM, RT_SRC_POS,
1015 N_("Unexpected error while restoring state. Please make sure the source and target VMs have compatible storage configurations"));
1016
1017 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1018
1019 if (RT_SUCCESS(rc))
1020 {
1021 uint32_t u32 = 0;
1022 rc = SSMR3GetU32(pSSM, &u32);
1023 if (RT_SUCCESS(rc))
1024 AssertMsgReturn(u32 == UINT32_MAX, ("%#x\n", u32), VERR_SSM_DATA_UNIT_FORMAT_CHANGED);
1025 }
1026
1027 return rc;
1028}
1029
1030int pdmR3BlkCacheInit(PVM pVM)
1031{
1032 int rc = VINF_SUCCESS;
1033 PUVM pUVM = pVM->pUVM;
1034 PPDMBLKCACHEGLOBAL pBlkCacheGlobal;
1035
1036 LogFlowFunc((": pVM=%p\n", pVM));
1037
1038 VM_ASSERT_EMT(pVM);
1039
1040 PCFGMNODE pCfgRoot = CFGMR3GetRoot(pVM);
1041 PCFGMNODE pCfgBlkCache = CFGMR3GetChild(CFGMR3GetChild(pCfgRoot, "PDM"), "BlkCache");
1042
1043 pBlkCacheGlobal = (PPDMBLKCACHEGLOBAL)RTMemAllocZ(sizeof(PDMBLKCACHEGLOBAL));
1044 if (!pBlkCacheGlobal)
1045 return VERR_NO_MEMORY;
1046
1047 RTListInit(&pBlkCacheGlobal->ListUsers);
1048 pBlkCacheGlobal->pVM = pVM;
1049 pBlkCacheGlobal->cRefs = 0;
1050 pBlkCacheGlobal->cbCached = 0;
1051 pBlkCacheGlobal->fCommitInProgress = false;
1052
1053 /* Initialize members */
1054 pBlkCacheGlobal->LruRecentlyUsedIn.pHead = NULL;
1055 pBlkCacheGlobal->LruRecentlyUsedIn.pTail = NULL;
1056 pBlkCacheGlobal->LruRecentlyUsedIn.cbCached = 0;
1057
1058 pBlkCacheGlobal->LruRecentlyUsedOut.pHead = NULL;
1059 pBlkCacheGlobal->LruRecentlyUsedOut.pTail = NULL;
1060 pBlkCacheGlobal->LruRecentlyUsedOut.cbCached = 0;
1061
1062 pBlkCacheGlobal->LruFrequentlyUsed.pHead = NULL;
1063 pBlkCacheGlobal->LruFrequentlyUsed.pTail = NULL;
1064 pBlkCacheGlobal->LruFrequentlyUsed.cbCached = 0;
1065
1066 do
1067 {
1068 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheSize", &pBlkCacheGlobal->cbMax, 5 * _1M);
1069 AssertLogRelRCBreak(rc);
1070 LogFlowFunc(("Maximum number of bytes cached %u\n", pBlkCacheGlobal->cbMax));
1071
1072 pBlkCacheGlobal->cbRecentlyUsedInMax = (pBlkCacheGlobal->cbMax / 100) * 25; /* 25% of the buffer size */
1073 pBlkCacheGlobal->cbRecentlyUsedOutMax = (pBlkCacheGlobal->cbMax / 100) * 50; /* 50% of the buffer size */
1074 LogFlowFunc(("cbRecentlyUsedInMax=%u cbRecentlyUsedOutMax=%u\n",
1075 pBlkCacheGlobal->cbRecentlyUsedInMax, pBlkCacheGlobal->cbRecentlyUsedOutMax));
1076
1077 /** @todo r=aeichner: Experiment to find optimal default values */
1078 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheCommitIntervalMs", &pBlkCacheGlobal->u32CommitTimeoutMs, 10000 /* 10sec */);
1079 AssertLogRelRCBreak(rc);
1080 rc = CFGMR3QueryU32Def(pCfgBlkCache, "CacheCommitThreshold", &pBlkCacheGlobal->cbCommitDirtyThreshold, pBlkCacheGlobal->cbMax / 2);
1081 AssertLogRelRCBreak(rc);
1082 } while (0);
1083
1084 if (RT_SUCCESS(rc))
1085 {
1086 STAMR3Register(pVM, &pBlkCacheGlobal->cbMax,
1087 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1088 "/PDM/BlkCache/cbMax",
1089 STAMUNIT_BYTES,
1090 "Maximum cache size");
1091 STAMR3Register(pVM, &pBlkCacheGlobal->cbCached,
1092 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1093 "/PDM/BlkCache/cbCached",
1094 STAMUNIT_BYTES,
1095 "Currently used cache");
1096 STAMR3Register(pVM, &pBlkCacheGlobal->LruRecentlyUsedIn.cbCached,
1097 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1098 "/PDM/BlkCache/cbCachedMruIn",
1099 STAMUNIT_BYTES,
1100 "Number of bytes cached in MRU list");
1101 STAMR3Register(pVM, &pBlkCacheGlobal->LruRecentlyUsedOut.cbCached,
1102 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1103 "/PDM/BlkCache/cbCachedMruOut",
1104 STAMUNIT_BYTES,
1105 "Number of bytes cached in FRU list");
1106 STAMR3Register(pVM, &pBlkCacheGlobal->LruFrequentlyUsed.cbCached,
1107 STAMTYPE_U32, STAMVISIBILITY_ALWAYS,
1108 "/PDM/BlkCache/cbCachedFru",
1109 STAMUNIT_BYTES,
1110 "Number of bytes cached in FRU ghost list");
1111
1112#ifdef VBOX_WITH_STATISTICS
1113 STAMR3Register(pVM, &pBlkCacheGlobal->cHits,
1114 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1115 "/PDM/BlkCache/CacheHits",
1116 STAMUNIT_COUNT, "Number of hits in the cache");
1117 STAMR3Register(pVM, &pBlkCacheGlobal->cPartialHits,
1118 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1119 "/PDM/BlkCache/CachePartialHits",
1120 STAMUNIT_COUNT, "Number of partial hits in the cache");
1121 STAMR3Register(pVM, &pBlkCacheGlobal->cMisses,
1122 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1123 "/PDM/BlkCache/CacheMisses",
1124 STAMUNIT_COUNT, "Number of misses when accessing the cache");
1125 STAMR3Register(pVM, &pBlkCacheGlobal->StatRead,
1126 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1127 "/PDM/BlkCache/CacheRead",
1128 STAMUNIT_BYTES, "Number of bytes read from the cache");
1129 STAMR3Register(pVM, &pBlkCacheGlobal->StatWritten,
1130 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1131 "/PDM/BlkCache/CacheWritten",
1132 STAMUNIT_BYTES, "Number of bytes written to the cache");
1133 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeGet,
1134 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1135 "/PDM/BlkCache/CacheTreeGet",
1136 STAMUNIT_TICKS_PER_CALL, "Time taken to access an entry in the tree");
1137 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeInsert,
1138 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1139 "/PDM/BlkCache/CacheTreeInsert",
1140 STAMUNIT_TICKS_PER_CALL, "Time taken to insert an entry in the tree");
1141 STAMR3Register(pVM, &pBlkCacheGlobal->StatTreeRemove,
1142 STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1143 "/PDM/BlkCache/CacheTreeRemove",
1144 STAMUNIT_TICKS_PER_CALL, "Time taken to remove an entry an the tree");
1145 STAMR3Register(pVM, &pBlkCacheGlobal->StatBuffersReused,
1146 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1147 "/PDM/BlkCache/CacheBuffersReused",
1148 STAMUNIT_COUNT, "Number of times a buffer could be reused");
1149#endif
1150
1151 /* Initialize the critical section */
1152 rc = RTCritSectInit(&pBlkCacheGlobal->CritSect);
1153 }
1154
1155 if (RT_SUCCESS(rc))
1156 {
1157 /* Create the commit timer */
1158 if (pBlkCacheGlobal->u32CommitTimeoutMs > 0)
1159 rc = TMR3TimerCreateInternal(pVM, TMCLOCK_REAL,
1160 pdmBlkCacheCommitTimerCallback,
1161 pBlkCacheGlobal,
1162 "BlkCache-Commit",
1163 &pBlkCacheGlobal->pTimerCommit);
1164
1165 if (RT_SUCCESS(rc))
1166 {
1167 /* Register saved state handler. */
1168 rc = SSMR3RegisterInternal(pVM, "pdmblkcache", 0, PDM_BLK_CACHE_SAVED_STATE_VERSION, pBlkCacheGlobal->cbMax,
1169 NULL, NULL, NULL,
1170 NULL, pdmR3BlkCacheSaveExec, NULL,
1171 NULL, pdmR3BlkCacheLoadExec, NULL);
1172 if (RT_SUCCESS(rc))
1173 {
1174 LogRel(("BlkCache: Cache successfully initialized. Cache size is %u bytes\n", pBlkCacheGlobal->cbMax));
1175 LogRel(("BlkCache: Cache commit interval is %u ms\n", pBlkCacheGlobal->u32CommitTimeoutMs));
1176 LogRel(("BlkCache: Cache commit threshold is %u bytes\n", pBlkCacheGlobal->cbCommitDirtyThreshold));
1177 pUVM->pdm.s.pBlkCacheGlobal = pBlkCacheGlobal;
1178 return VINF_SUCCESS;
1179 }
1180 }
1181
1182 RTCritSectDelete(&pBlkCacheGlobal->CritSect);
1183 }
1184
1185 if (pBlkCacheGlobal)
1186 RTMemFree(pBlkCacheGlobal);
1187
1188 LogFlowFunc((": returns rc=%Rrc\n", rc));
1189 return rc;
1190}
1191
1192void pdmR3BlkCacheTerm(PVM pVM)
1193{
1194 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1195
1196 if (pBlkCacheGlobal)
1197 {
1198 /* Make sure no one else uses the cache now */
1199 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1200
1201 /* Cleanup deleting all cache entries waiting for in progress entries to finish. */
1202 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruRecentlyUsedIn);
1203 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruRecentlyUsedOut);
1204 pdmBlkCacheDestroyList(&pBlkCacheGlobal->LruFrequentlyUsed);
1205
1206 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1207
1208 RTCritSectDelete(&pBlkCacheGlobal->CritSect);
1209 RTMemFree(pBlkCacheGlobal);
1210 pVM->pUVM->pdm.s.pBlkCacheGlobal = NULL;
1211 }
1212}
1213
1214int pdmR3BlkCacheResume(PVM pVM)
1215{
1216 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1217
1218 LogFlowFunc(("pVM=%#p\n", pVM));
1219
1220 if ( pBlkCacheGlobal
1221 && ASMAtomicXchgBool(&pBlkCacheGlobal->fIoErrorVmSuspended, false))
1222 {
1223 /* The VM was suspended because of an I/O error, commit all dirty entries. */
1224 pdmBlkCacheCommitDirtyEntries(pBlkCacheGlobal);
1225 }
1226
1227 return VINF_SUCCESS;
1228}
1229
1230static int pdmR3BlkCacheRetain(PVM pVM, PPPDMBLKCACHE ppBlkCache, const char *pcszId)
1231{
1232 int rc = VINF_SUCCESS;
1233 PPDMBLKCACHE pBlkCache = NULL;
1234 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1235
1236 if (!pBlkCacheGlobal)
1237 return VERR_NOT_SUPPORTED;
1238
1239 /*
1240 * Check that no other user cache has the same id first,
1241 * Unique id's are necessary in case the state is saved.
1242 */
1243 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1244
1245 pBlkCache = pdmR3BlkCacheFindById(pBlkCacheGlobal, pcszId);
1246
1247 if (!pBlkCache)
1248 {
1249 pBlkCache = (PPDMBLKCACHE)RTMemAllocZ(sizeof(PDMBLKCACHE));
1250
1251 if (pBlkCache)
1252 pBlkCache->pszId = RTStrDup(pcszId);
1253
1254 if ( pBlkCache
1255 && pBlkCache->pszId)
1256 {
1257 pBlkCache->fSuspended = false;
1258 pBlkCache->cIoXfersActive = 0;
1259 pBlkCache->pCache = pBlkCacheGlobal;
1260 RTListInit(&pBlkCache->ListDirtyNotCommitted);
1261
1262 rc = RTSpinlockCreate(&pBlkCache->LockList, RTSPINLOCK_FLAGS_INTERRUPT_UNSAFE, "pdmR3BlkCacheRetain");
1263 if (RT_SUCCESS(rc))
1264 {
1265 rc = RTSemRWCreate(&pBlkCache->SemRWEntries);
1266 if (RT_SUCCESS(rc))
1267 {
1268 pBlkCache->pTree = (PAVLRU64TREE)RTMemAllocZ(sizeof(AVLRFOFFTREE));
1269 if (pBlkCache->pTree)
1270 {
1271#ifdef VBOX_WITH_STATISTICS
1272 STAMR3RegisterF(pBlkCacheGlobal->pVM, &pBlkCache->StatWriteDeferred,
1273 STAMTYPE_COUNTER, STAMVISIBILITY_ALWAYS,
1274 STAMUNIT_COUNT, "Number of deferred writes",
1275 "/PDM/BlkCache/%s/Cache/DeferredWrites", pBlkCache->pszId);
1276#endif
1277
1278 /* Add to the list of users. */
1279 pBlkCacheGlobal->cRefs++;
1280 RTListAppend(&pBlkCacheGlobal->ListUsers, &pBlkCache->NodeCacheUser);
1281 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1282
1283 *ppBlkCache = pBlkCache;
1284 LogFlowFunc(("returns success\n"));
1285 return VINF_SUCCESS;
1286 }
1287
1288 rc = VERR_NO_MEMORY;
1289 RTSemRWDestroy(pBlkCache->SemRWEntries);
1290 }
1291
1292 RTSpinlockDestroy(pBlkCache->LockList);
1293 }
1294
1295 RTStrFree(pBlkCache->pszId);
1296 }
1297 else
1298 rc = VERR_NO_MEMORY;
1299
1300 if (pBlkCache)
1301 RTMemFree(pBlkCache);
1302 }
1303 else
1304 rc = VERR_ALREADY_EXISTS;
1305
1306 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1307
1308 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1309 return rc;
1310}
1311
1312VMMR3DECL(int) PDMR3BlkCacheRetainDriver(PVM pVM, PPDMDRVINS pDrvIns, PPPDMBLKCACHE ppBlkCache,
1313 PFNPDMBLKCACHEXFERCOMPLETEDRV pfnXferComplete,
1314 PFNPDMBLKCACHEXFERENQUEUEDRV pfnXferEnqueue,
1315 PFNPDMBLKCACHEXFERENQUEUEDISCARDDRV pfnXferEnqueueDiscard,
1316 const char *pcszId)
1317{
1318 int rc = VINF_SUCCESS;
1319 PPDMBLKCACHE pBlkCache;
1320
1321 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1322 if (RT_SUCCESS(rc))
1323 {
1324 pBlkCache->enmType = PDMBLKCACHETYPE_DRV;
1325 pBlkCache->u.Drv.pfnXferComplete = pfnXferComplete;
1326 pBlkCache->u.Drv.pfnXferEnqueue = pfnXferEnqueue;
1327 pBlkCache->u.Drv.pfnXferEnqueueDiscard = pfnXferEnqueueDiscard;
1328 pBlkCache->u.Drv.pDrvIns = pDrvIns;
1329 *ppBlkCache = pBlkCache;
1330 }
1331
1332 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1333 return rc;
1334}
1335
1336VMMR3DECL(int) PDMR3BlkCacheRetainDevice(PVM pVM, PPDMDEVINS pDevIns, PPPDMBLKCACHE ppBlkCache,
1337 PFNPDMBLKCACHEXFERCOMPLETEDEV pfnXferComplete,
1338 PFNPDMBLKCACHEXFERENQUEUEDEV pfnXferEnqueue,
1339 PFNPDMBLKCACHEXFERENQUEUEDISCARDDEV pfnXferEnqueueDiscard,
1340 const char *pcszId)
1341{
1342 int rc = VINF_SUCCESS;
1343 PPDMBLKCACHE pBlkCache;
1344
1345 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1346 if (RT_SUCCESS(rc))
1347 {
1348 pBlkCache->enmType = PDMBLKCACHETYPE_DEV;
1349 pBlkCache->u.Dev.pfnXferComplete = pfnXferComplete;
1350 pBlkCache->u.Dev.pfnXferEnqueue = pfnXferEnqueue;
1351 pBlkCache->u.Dev.pfnXferEnqueueDiscard = pfnXferEnqueueDiscard;
1352 pBlkCache->u.Dev.pDevIns = pDevIns;
1353 *ppBlkCache = pBlkCache;
1354 }
1355
1356 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1357 return rc;
1358
1359}
1360
1361VMMR3DECL(int) PDMR3BlkCacheRetainUsb(PVM pVM, PPDMUSBINS pUsbIns, PPPDMBLKCACHE ppBlkCache,
1362 PFNPDMBLKCACHEXFERCOMPLETEUSB pfnXferComplete,
1363 PFNPDMBLKCACHEXFERENQUEUEUSB pfnXferEnqueue,
1364 PFNPDMBLKCACHEXFERENQUEUEDISCARDUSB pfnXferEnqueueDiscard,
1365 const char *pcszId)
1366{
1367 int rc = VINF_SUCCESS;
1368 PPDMBLKCACHE pBlkCache;
1369
1370 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1371 if (RT_SUCCESS(rc))
1372 {
1373 pBlkCache->enmType = PDMBLKCACHETYPE_USB;
1374 pBlkCache->u.Usb.pfnXferComplete = pfnXferComplete;
1375 pBlkCache->u.Usb.pfnXferEnqueue = pfnXferEnqueue;
1376 pBlkCache->u.Usb.pfnXferEnqueueDiscard = pfnXferEnqueueDiscard;
1377 pBlkCache->u.Usb.pUsbIns = pUsbIns;
1378 *ppBlkCache = pBlkCache;
1379 }
1380
1381 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1382 return rc;
1383
1384}
1385
1386VMMR3DECL(int) PDMR3BlkCacheRetainInt(PVM pVM, void *pvUser, PPPDMBLKCACHE ppBlkCache,
1387 PFNPDMBLKCACHEXFERCOMPLETEINT pfnXferComplete,
1388 PFNPDMBLKCACHEXFERENQUEUEINT pfnXferEnqueue,
1389 PFNPDMBLKCACHEXFERENQUEUEDISCARDINT pfnXferEnqueueDiscard,
1390 const char *pcszId)
1391{
1392 int rc = VINF_SUCCESS;
1393 PPDMBLKCACHE pBlkCache;
1394
1395 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1396 if (RT_SUCCESS(rc))
1397 {
1398 pBlkCache->enmType = PDMBLKCACHETYPE_INTERNAL;
1399 pBlkCache->u.Int.pfnXferComplete = pfnXferComplete;
1400 pBlkCache->u.Int.pfnXferEnqueue = pfnXferEnqueue;
1401 pBlkCache->u.Int.pfnXferEnqueueDiscard = pfnXferEnqueueDiscard;
1402 pBlkCache->u.Int.pvUser = pvUser;
1403 *ppBlkCache = pBlkCache;
1404 }
1405
1406 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1407 return rc;
1408
1409}
1410
1411/**
1412 * Callback for the AVL destroy routine. Frees a cache entry for this endpoint.
1413 *
1414 * @returns IPRT status code.
1415 * @param pNode The node to destroy.
1416 * @param pvUser Opaque user data.
1417 */
1418static DECLCALLBACK(int) pdmBlkCacheEntryDestroy(PAVLRU64NODECORE pNode, void *pvUser)
1419{
1420 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)pNode;
1421 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
1422 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
1423
1424 while (ASMAtomicReadU32(&pEntry->fFlags) & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS)
1425 {
1426 /* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
1427 pdmBlkCacheEntryRef(pEntry);
1428 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1429 pdmBlkCacheLockLeave(pCache);
1430
1431 RTThreadSleep(250);
1432
1433 /* Re-enter all locks */
1434 pdmBlkCacheLockEnter(pCache);
1435 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1436 pdmBlkCacheEntryRelease(pEntry);
1437 }
1438
1439 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
1440 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
1441
1442 bool fUpdateCache = pEntry->pList == &pCache->LruFrequentlyUsed
1443 || pEntry->pList == &pCache->LruRecentlyUsedIn;
1444
1445 pdmBlkCacheEntryRemoveFromList(pEntry);
1446
1447 if (fUpdateCache)
1448 pdmBlkCacheSub(pCache, pEntry->cbData);
1449
1450 RTMemPageFree(pEntry->pbData, pEntry->cbData);
1451 RTMemFree(pEntry);
1452
1453 return VINF_SUCCESS;
1454}
1455
1456VMMR3DECL(void) PDMR3BlkCacheRelease(PPDMBLKCACHE pBlkCache)
1457{
1458 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1459
1460 /*
1461 * Commit all dirty entries now (they are waited on for completion during the
1462 * destruction of the AVL tree below).
1463 * The exception is if the VM was paused because of an I/O error before.
1464 */
1465 if (!ASMAtomicReadBool(&pCache->fIoErrorVmSuspended))
1466 pdmBlkCacheCommit(pBlkCache);
1467
1468 /* Make sure nobody is accessing the cache while we delete the tree. */
1469 pdmBlkCacheLockEnter(pCache);
1470 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1471 RTAvlrU64Destroy(pBlkCache->pTree, pdmBlkCacheEntryDestroy, pCache);
1472 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1473
1474 RTSpinlockDestroy(pBlkCache->LockList);
1475
1476 pCache->cRefs--;
1477 RTListNodeRemove(&pBlkCache->NodeCacheUser);
1478
1479 pdmBlkCacheLockLeave(pCache);
1480
1481 RTMemFree(pBlkCache->pTree);
1482 pBlkCache->pTree = NULL;
1483 RTSemRWDestroy(pBlkCache->SemRWEntries);
1484
1485#ifdef VBOX_WITH_STATISTICS
1486 STAMR3DeregisterF(pCache->pVM->pUVM, "/PDM/BlkCache/%s/Cache/DeferredWrites", pBlkCache->pszId);
1487#endif
1488
1489 RTStrFree(pBlkCache->pszId);
1490 RTMemFree(pBlkCache);
1491}
1492
1493VMMR3DECL(void) PDMR3BlkCacheReleaseDevice(PVM pVM, PPDMDEVINS pDevIns)
1494{
1495 LogFlow(("%s: pDevIns=%p\n", __FUNCTION__, pDevIns));
1496
1497 /*
1498 * Validate input.
1499 */
1500 if (!pDevIns)
1501 return;
1502 VM_ASSERT_EMT(pVM);
1503
1504 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1505 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1506
1507 /* Return silently if not supported. */
1508 if (!pBlkCacheGlobal)
1509 return;
1510
1511 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1512
1513 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1514 {
1515 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DEV
1516 && pBlkCache->u.Dev.pDevIns == pDevIns)
1517 PDMR3BlkCacheRelease(pBlkCache);
1518 }
1519
1520 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1521}
1522
1523VMMR3DECL(void) PDMR3BlkCacheReleaseDriver(PVM pVM, PPDMDRVINS pDrvIns)
1524{
1525 LogFlow(("%s: pDrvIns=%p\n", __FUNCTION__, pDrvIns));
1526
1527 /*
1528 * Validate input.
1529 */
1530 if (!pDrvIns)
1531 return;
1532 VM_ASSERT_EMT(pVM);
1533
1534 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1535 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1536
1537 /* Return silently if not supported. */
1538 if (!pBlkCacheGlobal)
1539 return;
1540
1541 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1542
1543 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1544 {
1545 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DRV
1546 && pBlkCache->u.Drv.pDrvIns == pDrvIns)
1547 PDMR3BlkCacheRelease(pBlkCache);
1548 }
1549
1550 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1551}
1552
1553VMMR3DECL(void) PDMR3BlkCacheReleaseUsb(PVM pVM, PPDMUSBINS pUsbIns)
1554{
1555 LogFlow(("%s: pUsbIns=%p\n", __FUNCTION__, pUsbIns));
1556
1557 /*
1558 * Validate input.
1559 */
1560 if (!pUsbIns)
1561 return;
1562 VM_ASSERT_EMT(pVM);
1563
1564 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1565 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1566
1567 /* Return silently if not supported. */
1568 if (!pBlkCacheGlobal)
1569 return;
1570
1571 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1572
1573 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1574 {
1575 if ( pBlkCache->enmType == PDMBLKCACHETYPE_USB
1576 && pBlkCache->u.Usb.pUsbIns == pUsbIns)
1577 PDMR3BlkCacheRelease(pBlkCache);
1578 }
1579
1580 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1581}
1582
1583static PPDMBLKCACHEENTRY pdmBlkCacheGetCacheEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off)
1584{
1585 STAM_PROFILE_ADV_START(&pBlkCache->pCache->StatTreeGet, Cache);
1586
1587 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1588 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)RTAvlrU64RangeGet(pBlkCache->pTree, off);
1589 if (pEntry)
1590 pdmBlkCacheEntryRef(pEntry);
1591 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1592
1593 STAM_PROFILE_ADV_STOP(&pBlkCache->pCache->StatTreeGet, Cache);
1594
1595 return pEntry;
1596}
1597
1598/**
1599 * Return the best fit cache entries for the given offset.
1600 *
1601 * @returns nothing.
1602 * @param pBlkCache The endpoint cache.
1603 * @param off The offset.
1604 * @param ppEntryAbove Where to store the pointer to the best fit entry above
1605 * the given offset. NULL if not required.
1606 */
1607static void pdmBlkCacheGetCacheBestFitEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off, PPDMBLKCACHEENTRY *ppEntryAbove)
1608{
1609 STAM_PROFILE_ADV_START(&pBlkCache->pCache->StatTreeGet, Cache);
1610
1611 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1612 if (ppEntryAbove)
1613 {
1614 *ppEntryAbove = (PPDMBLKCACHEENTRY)RTAvlrU64GetBestFit(pBlkCache->pTree, off, true /*fAbove*/);
1615 if (*ppEntryAbove)
1616 pdmBlkCacheEntryRef(*ppEntryAbove);
1617 }
1618
1619 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1620
1621 STAM_PROFILE_ADV_STOP(&pBlkCache->pCache->StatTreeGet, Cache);
1622}
1623
1624static void pdmBlkCacheInsertEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
1625{
1626 STAM_PROFILE_ADV_START(&pBlkCache->pCache->StatTreeInsert, Cache);
1627 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1628 bool fInserted = RTAvlrU64Insert(pBlkCache->pTree, &pEntry->Core);
1629 AssertMsg(fInserted, ("Node was not inserted into tree\n")); NOREF(fInserted);
1630 STAM_PROFILE_ADV_STOP(&pBlkCache->pCache->StatTreeInsert, Cache);
1631 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1632}
1633
1634/**
1635 * Allocates and initializes a new entry for the cache.
1636 * The entry has a reference count of 1.
1637 *
1638 * @returns Pointer to the new cache entry or NULL if out of memory.
1639 * @param pBlkCache The cache the entry belongs to.
1640 * @param off Start offset.
1641 * @param cbData Size of the cache entry.
1642 * @param pbBuffer Pointer to the buffer to use.
1643 * NULL if a new buffer should be allocated.
1644 * The buffer needs to have the same size of the entry.
1645 */
1646static PPDMBLKCACHEENTRY pdmBlkCacheEntryAlloc(PPDMBLKCACHE pBlkCache, uint64_t off, size_t cbData, uint8_t *pbBuffer)
1647{
1648 AssertReturn(cbData <= UINT32_MAX, NULL);
1649 PPDMBLKCACHEENTRY pEntryNew = (PPDMBLKCACHEENTRY)RTMemAllocZ(sizeof(PDMBLKCACHEENTRY));
1650
1651 if (RT_UNLIKELY(!pEntryNew))
1652 return NULL;
1653
1654 pEntryNew->Core.Key = off;
1655 pEntryNew->Core.KeyLast = off + cbData - 1;
1656 pEntryNew->pBlkCache = pBlkCache;
1657 pEntryNew->fFlags = 0;
1658 pEntryNew->cRefs = 1; /* We are using it now. */
1659 pEntryNew->pList = NULL;
1660 pEntryNew->cbData = (uint32_t)cbData;
1661 pEntryNew->pWaitingHead = NULL;
1662 pEntryNew->pWaitingTail = NULL;
1663 if (pbBuffer)
1664 pEntryNew->pbData = pbBuffer;
1665 else
1666 pEntryNew->pbData = (uint8_t *)RTMemPageAlloc(cbData);
1667
1668 if (RT_UNLIKELY(!pEntryNew->pbData))
1669 {
1670 RTMemFree(pEntryNew);
1671 return NULL;
1672 }
1673
1674 return pEntryNew;
1675}
1676
1677/**
1678 * Checks that a set of flags is set/clear acquiring the R/W semaphore
1679 * in exclusive mode.
1680 *
1681 * @returns true if the flag in fSet is set and the one in fClear is clear.
1682 * false otherwise.
1683 * The R/W semaphore is only held if true is returned.
1684 *
1685 * @param pBlkCache The endpoint cache instance data.
1686 * @param pEntry The entry to check the flags for.
1687 * @param fSet The flag which is tested to be set.
1688 * @param fClear The flag which is tested to be clear.
1689 */
1690DECLINLINE(bool) pdmBlkCacheEntryFlagIsSetClearAcquireLock(PPDMBLKCACHE pBlkCache,
1691 PPDMBLKCACHEENTRY pEntry,
1692 uint32_t fSet, uint32_t fClear)
1693{
1694 uint32_t fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1695 bool fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1696
1697 if (fPassed)
1698 {
1699 /* Acquire the lock and check again because the completion callback might have raced us. */
1700 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1701
1702 fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1703 fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1704
1705 /* Drop the lock if we didn't passed the test. */
1706 if (!fPassed)
1707 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1708 }
1709
1710 return fPassed;
1711}
1712
1713/**
1714 * Adds a segment to the waiting list for a cache entry
1715 * which is currently in progress.
1716 *
1717 * @returns nothing.
1718 * @param pEntry The cache entry to add the segment to.
1719 * @param pWaiter The waiter entry to add.
1720 */
1721DECLINLINE(void) pdmBlkCacheEntryAddWaiter(PPDMBLKCACHEENTRY pEntry,
1722 PPDMBLKCACHEWAITER pWaiter)
1723{
1724 pWaiter->pNext = NULL;
1725
1726 if (pEntry->pWaitingHead)
1727 {
1728 AssertPtr(pEntry->pWaitingTail);
1729
1730 pEntry->pWaitingTail->pNext = pWaiter;
1731 pEntry->pWaitingTail = pWaiter;
1732 }
1733 else
1734 {
1735 Assert(!pEntry->pWaitingTail);
1736
1737 pEntry->pWaitingHead = pWaiter;
1738 pEntry->pWaitingTail = pWaiter;
1739 }
1740}
1741
1742/**
1743 * Add a buffer described by the I/O memory context
1744 * to the entry waiting for completion.
1745 *
1746 * @returns VBox status code.
1747 * @param pEntry The entry to add the buffer to.
1748 * @param pReq The request.
1749 * @param pSgBuf The scatter/gather buffer. Will be advanced by cbData.
1750 * @param offDiff Offset from the start of the buffer in the entry.
1751 * @param cbData Amount of data to wait for onthis entry.
1752 * @param fWrite Flag whether the task waits because it wants to write to
1753 * the cache entry.
1754 */
1755static int pdmBlkCacheEntryWaitersAdd(PPDMBLKCACHEENTRY pEntry, PPDMBLKCACHEREQ pReq,
1756 PRTSGBUF pSgBuf, uint64_t offDiff, size_t cbData, bool fWrite)
1757{
1758 PPDMBLKCACHEWAITER pWaiter = (PPDMBLKCACHEWAITER)RTMemAllocZ(sizeof(PDMBLKCACHEWAITER));
1759 if (!pWaiter)
1760 return VERR_NO_MEMORY;
1761
1762 ASMAtomicIncU32(&pReq->cXfersPending);
1763 pWaiter->pReq = pReq;
1764 pWaiter->offCacheEntry = offDiff;
1765 pWaiter->cbTransfer = cbData;
1766 pWaiter->fWrite = fWrite;
1767 RTSgBufClone(&pWaiter->SgBuf, pSgBuf);
1768 RTSgBufAdvance(pSgBuf, cbData);
1769
1770 pdmBlkCacheEntryAddWaiter(pEntry, pWaiter);
1771
1772 return VINF_SUCCESS;
1773}
1774
1775/**
1776 * Calculate aligned offset and size for a new cache entry which do not
1777 * intersect with an already existing entry and the file end.
1778 *
1779 * @returns The number of bytes the entry can hold of the requested amount
1780 * of bytes.
1781 * @param pBlkCache The endpoint cache.
1782 * @param off The start offset.
1783 * @param cb The number of bytes the entry needs to hold at
1784 * least.
1785 * @param pcbEntry Where to store the number of bytes the entry can hold.
1786 * Can be less than given because of other entries.
1787 */
1788static uint32_t pdmBlkCacheEntryBoundariesCalc(PPDMBLKCACHE pBlkCache,
1789 uint64_t off, uint32_t cb,
1790 uint32_t *pcbEntry)
1791{
1792 /* Get the best fit entries around the offset */
1793 PPDMBLKCACHEENTRY pEntryAbove = NULL;
1794 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off, &pEntryAbove);
1795
1796 /* Log the info */
1797 LogFlow(("%sest fit entry above off=%llu (BestFit=%llu BestFitEnd=%llu BestFitSize=%u)\n",
1798 pEntryAbove ? "B" : "No b",
1799 off,
1800 pEntryAbove ? pEntryAbove->Core.Key : 0,
1801 pEntryAbove ? pEntryAbove->Core.KeyLast : 0,
1802 pEntryAbove ? pEntryAbove->cbData : 0));
1803
1804 uint32_t cbNext;
1805 uint32_t cbInEntry;
1806 if ( pEntryAbove
1807 && off + cb > pEntryAbove->Core.Key)
1808 {
1809 cbInEntry = (uint32_t)(pEntryAbove->Core.Key - off);
1810 cbNext = (uint32_t)(pEntryAbove->Core.Key - off);
1811 }
1812 else
1813 {
1814 cbInEntry = cb;
1815 cbNext = cb;
1816 }
1817
1818 /* A few sanity checks */
1819 AssertMsg(!pEntryAbove || off + cbNext <= pEntryAbove->Core.Key,
1820 ("Aligned size intersects with another cache entry\n"));
1821 Assert(cbInEntry <= cbNext);
1822
1823 if (pEntryAbove)
1824 pdmBlkCacheEntryRelease(pEntryAbove);
1825
1826 LogFlow(("off=%llu cbNext=%u\n", off, cbNext));
1827
1828 *pcbEntry = cbNext;
1829
1830 return cbInEntry;
1831}
1832
1833/**
1834 * Create a new cache entry evicting data from the cache if required.
1835 *
1836 * @returns Pointer to the new cache entry or NULL
1837 * if not enough bytes could be evicted from the cache.
1838 * @param pBlkCache The endpoint cache.
1839 * @param off The offset.
1840 * @param cb Number of bytes the cache entry should have.
1841 * @param pcbData Where to store the number of bytes the new
1842 * entry can hold. May be lower than actually
1843 * requested due to another entry intersecting the
1844 * access range.
1845 */
1846static PPDMBLKCACHEENTRY pdmBlkCacheEntryCreate(PPDMBLKCACHE pBlkCache, uint64_t off, size_t cb, size_t *pcbData)
1847{
1848 uint32_t cbEntry = 0;
1849
1850 *pcbData = pdmBlkCacheEntryBoundariesCalc(pBlkCache, off, (uint32_t)cb, &cbEntry);
1851 AssertReturn(cb <= UINT32_MAX, NULL);
1852
1853 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1854 pdmBlkCacheLockEnter(pCache);
1855
1856 PPDMBLKCACHEENTRY pEntryNew = NULL;
1857 uint8_t *pbBuffer = NULL;
1858 bool fEnough = pdmBlkCacheReclaim(pCache, cbEntry, true, &pbBuffer);
1859 if (fEnough)
1860 {
1861 LogFlow(("Evicted enough bytes (%u requested). Creating new cache entry\n", cbEntry));
1862
1863 pEntryNew = pdmBlkCacheEntryAlloc(pBlkCache, off, cbEntry, pbBuffer);
1864 if (RT_LIKELY(pEntryNew))
1865 {
1866 pdmBlkCacheEntryAddToList(&pCache->LruRecentlyUsedIn, pEntryNew);
1867 pdmBlkCacheAdd(pCache, cbEntry);
1868 pdmBlkCacheLockLeave(pCache);
1869
1870 pdmBlkCacheInsertEntry(pBlkCache, pEntryNew);
1871
1872 AssertMsg( (off >= pEntryNew->Core.Key)
1873 && (off + *pcbData <= pEntryNew->Core.KeyLast + 1),
1874 ("Overflow in calculation off=%llu\n", off));
1875 }
1876 else
1877 pdmBlkCacheLockLeave(pCache);
1878 }
1879 else
1880 pdmBlkCacheLockLeave(pCache);
1881
1882 return pEntryNew;
1883}
1884
1885static PPDMBLKCACHEREQ pdmBlkCacheReqAlloc(void *pvUser)
1886{
1887 PPDMBLKCACHEREQ pReq = (PPDMBLKCACHEREQ)RTMemAlloc(sizeof(PDMBLKCACHEREQ));
1888
1889 if (RT_LIKELY(pReq))
1890 {
1891 pReq->pvUser = pvUser;
1892 pReq->rcReq = VINF_SUCCESS;
1893 pReq->cXfersPending = 0;
1894 }
1895
1896 return pReq;
1897}
1898
1899static void pdmBlkCacheReqComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq)
1900{
1901 switch (pBlkCache->enmType)
1902 {
1903 case PDMBLKCACHETYPE_DEV:
1904 {
1905 pBlkCache->u.Dev.pfnXferComplete(pBlkCache->u.Dev.pDevIns,
1906 pReq->pvUser, pReq->rcReq);
1907 break;
1908 }
1909 case PDMBLKCACHETYPE_DRV:
1910 {
1911 pBlkCache->u.Drv.pfnXferComplete(pBlkCache->u.Drv.pDrvIns,
1912 pReq->pvUser, pReq->rcReq);
1913 break;
1914 }
1915 case PDMBLKCACHETYPE_USB:
1916 {
1917 pBlkCache->u.Usb.pfnXferComplete(pBlkCache->u.Usb.pUsbIns,
1918 pReq->pvUser, pReq->rcReq);
1919 break;
1920 }
1921 case PDMBLKCACHETYPE_INTERNAL:
1922 {
1923 pBlkCache->u.Int.pfnXferComplete(pBlkCache->u.Int.pvUser,
1924 pReq->pvUser, pReq->rcReq);
1925 break;
1926 }
1927 default:
1928 AssertMsgFailed(("Unknown block cache type!\n"));
1929 }
1930
1931 RTMemFree(pReq);
1932}
1933
1934static bool pdmBlkCacheReqUpdate(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq,
1935 int rcReq, bool fCallHandler)
1936{
1937 if (RT_FAILURE(rcReq))
1938 ASMAtomicCmpXchgS32(&pReq->rcReq, rcReq, VINF_SUCCESS);
1939
1940 AssertMsg(pReq->cXfersPending > 0, ("No transfers are pending for this request\n"));
1941 uint32_t cXfersPending = ASMAtomicDecU32(&pReq->cXfersPending);
1942
1943 if (!cXfersPending)
1944 {
1945 if (fCallHandler)
1946 pdmBlkCacheReqComplete(pBlkCache, pReq);
1947 return true;
1948 }
1949
1950 LogFlowFunc(("pReq=%#p cXfersPending=%u\n", pReq, cXfersPending));
1951 return false;
1952}
1953
1954VMMR3DECL(int) PDMR3BlkCacheRead(PPDMBLKCACHE pBlkCache, uint64_t off,
1955 PCRTSGBUF pSgBuf, size_t cbRead, void *pvUser)
1956{
1957 int rc = VINF_SUCCESS;
1958 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1959 PPDMBLKCACHEENTRY pEntry;
1960 PPDMBLKCACHEREQ pReq;
1961
1962 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pSgBuf=%#p cbRead=%u pvUser=%#p\n",
1963 pBlkCache, pBlkCache->pszId, off, pSgBuf, cbRead, pvUser));
1964
1965 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
1966 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
1967
1968 RTSGBUF SgBuf;
1969 RTSgBufClone(&SgBuf, pSgBuf);
1970
1971 /* Allocate new request structure. */
1972 pReq = pdmBlkCacheReqAlloc(pvUser);
1973 if (RT_UNLIKELY(!pReq))
1974 return VERR_NO_MEMORY;
1975
1976 /* Increment data transfer counter to keep the request valid while we access it. */
1977 ASMAtomicIncU32(&pReq->cXfersPending);
1978
1979 while (cbRead)
1980 {
1981 size_t cbToRead;
1982
1983 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
1984
1985 /*
1986 * If there is no entry we try to create a new one eviciting unused pages
1987 * if the cache is full. If this is not possible we will pass the request through
1988 * and skip the caching (all entries may be still in progress so they can't
1989 * be evicted)
1990 * If we have an entry it can be in one of the LRU lists where the entry
1991 * contains data (recently used or frequently used LRU) so we can just read
1992 * the data we need and put the entry at the head of the frequently used LRU list.
1993 * In case the entry is in one of the ghost lists it doesn't contain any data.
1994 * We have to fetch it again evicting pages from either T1 or T2 to make room.
1995 */
1996 if (pEntry)
1997 {
1998 uint64_t offDiff = off - pEntry->Core.Key;
1999
2000 AssertMsg(off >= pEntry->Core.Key,
2001 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
2002 off, pEntry->Core.Key));
2003
2004 AssertPtr(pEntry->pList);
2005
2006 cbToRead = RT_MIN(pEntry->cbData - offDiff, cbRead);
2007
2008 AssertMsg(off + cbToRead <= pEntry->Core.Key + pEntry->Core.KeyLast + 1,
2009 ("Buffer of cache entry exceeded off=%llu cbToRead=%d\n",
2010 off, cbToRead));
2011
2012 cbRead -= cbToRead;
2013
2014 if (!cbRead)
2015 STAM_COUNTER_INC(&pCache->cHits);
2016 else
2017 STAM_COUNTER_INC(&pCache->cPartialHits);
2018
2019 STAM_COUNTER_ADD(&pCache->StatRead, cbToRead);
2020
2021 /* Ghost lists contain no data. */
2022 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2023 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2024 {
2025 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2026 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2027 PDMBLKCACHE_ENTRY_IS_DIRTY))
2028 {
2029 /* Entry didn't completed yet. Append to the list */
2030 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2031 &SgBuf, offDiff, cbToRead,
2032 false /* fWrite */);
2033 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2034 }
2035 else
2036 {
2037 /* Read as much as we can from the entry. */
2038 RTSgBufCopyFromBuf(&SgBuf, pEntry->pbData + offDiff, cbToRead);
2039 }
2040
2041 /* Move this entry to the top position */
2042 if (pEntry->pList == &pCache->LruFrequentlyUsed)
2043 {
2044 pdmBlkCacheLockEnter(pCache);
2045 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2046 pdmBlkCacheLockLeave(pCache);
2047 }
2048 /* Release the entry */
2049 pdmBlkCacheEntryRelease(pEntry);
2050 }
2051 else
2052 {
2053 uint8_t *pbBuffer = NULL;
2054
2055 LogFlow(("Fetching data for ghost entry %#p from file\n", pEntry));
2056
2057 pdmBlkCacheLockEnter(pCache);
2058 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2059 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2060
2061 /* Move the entry to Am and fetch it to the cache. */
2062 if (fEnough)
2063 {
2064 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2065 pdmBlkCacheAdd(pCache, pEntry->cbData);
2066 pdmBlkCacheLockLeave(pCache);
2067
2068 if (pbBuffer)
2069 pEntry->pbData = pbBuffer;
2070 else
2071 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2072 AssertPtr(pEntry->pbData);
2073
2074 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2075 &SgBuf, offDiff, cbToRead,
2076 false /* fWrite */);
2077 pdmBlkCacheEntryReadFromMedium(pEntry);
2078 /* Release the entry */
2079 pdmBlkCacheEntryRelease(pEntry);
2080 }
2081 else
2082 {
2083 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2084 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2085 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2086 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2087 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2088
2089 pdmBlkCacheLockLeave(pCache);
2090
2091 RTMemFree(pEntry);
2092
2093 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2094 &SgBuf, off, cbToRead,
2095 PDMBLKCACHEXFERDIR_READ);
2096 }
2097 }
2098 }
2099 else
2100 {
2101#ifdef VBOX_WITH_IO_READ_CACHE
2102 /* No entry found for this offset. Create a new entry and fetch the data to the cache. */
2103 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
2104 off, cbRead,
2105 &cbToRead);
2106
2107 cbRead -= cbToRead;
2108
2109 if (pEntryNew)
2110 {
2111 if (!cbRead)
2112 STAM_COUNTER_INC(&pCache->cMisses);
2113 else
2114 STAM_COUNTER_INC(&pCache->cPartialHits);
2115
2116 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
2117 &SgBuf,
2118 off - pEntryNew->Core.Key,
2119 cbToRead,
2120 false /* fWrite */);
2121 pdmBlkCacheEntryReadFromMedium(pEntryNew);
2122 pdmBlkCacheEntryRelease(pEntryNew); /* it is protected by the I/O in progress flag now. */
2123 }
2124 else
2125 {
2126 /*
2127 * There is not enough free space in the cache.
2128 * Pass the request directly to the I/O manager.
2129 */
2130 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToRead));
2131
2132 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2133 &SgBuf, off, cbToRead,
2134 PDMBLKCACHEXFERDIR_READ);
2135 }
2136#else
2137 /* Clip read size if necessary. */
2138 PPDMBLKCACHEENTRY pEntryAbove;
2139 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off, &pEntryAbove);
2140
2141 if (pEntryAbove)
2142 {
2143 if (off + cbRead > pEntryAbove->Core.Key)
2144 cbToRead = pEntryAbove->Core.Key - off;
2145 else
2146 cbToRead = cbRead;
2147
2148 pdmBlkCacheEntryRelease(pEntryAbove);
2149 }
2150 else
2151 cbToRead = cbRead;
2152
2153 cbRead -= cbToRead;
2154 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2155 &SgBuf, off, cbToRead,
2156 PDMBLKCACHEXFERDIR_READ);
2157#endif
2158 }
2159 off += cbToRead;
2160 }
2161
2162 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, false))
2163 rc = VINF_AIO_TASK_PENDING;
2164 else
2165 {
2166 rc = pReq->rcReq;
2167 RTMemFree(pReq);
2168 }
2169
2170 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2171
2172 return rc;
2173}
2174
2175VMMR3DECL(int) PDMR3BlkCacheWrite(PPDMBLKCACHE pBlkCache, uint64_t off, PCRTSGBUF pSgBuf, size_t cbWrite, void *pvUser)
2176{
2177 int rc = VINF_SUCCESS;
2178 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2179 PPDMBLKCACHEENTRY pEntry;
2180 PPDMBLKCACHEREQ pReq;
2181
2182 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pSgBuf=%#p cbWrite=%u pvUser=%#p\n",
2183 pBlkCache, pBlkCache->pszId, off, pSgBuf, cbWrite, pvUser));
2184
2185 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2186 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
2187
2188 RTSGBUF SgBuf;
2189 RTSgBufClone(&SgBuf, pSgBuf);
2190
2191 /* Allocate new request structure. */
2192 pReq = pdmBlkCacheReqAlloc(pvUser);
2193 if (RT_UNLIKELY(!pReq))
2194 return VERR_NO_MEMORY;
2195
2196 /* Increment data transfer counter to keep the request valid while we access it. */
2197 ASMAtomicIncU32(&pReq->cXfersPending);
2198
2199 while (cbWrite)
2200 {
2201 size_t cbToWrite;
2202
2203 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
2204 if (pEntry)
2205 {
2206 /* Write the data into the entry and mark it as dirty */
2207 AssertPtr(pEntry->pList);
2208
2209 uint64_t offDiff = off - pEntry->Core.Key;
2210 AssertMsg(off >= pEntry->Core.Key, ("Overflow in calculation off=%llu OffsetAligned=%llu\n", off, pEntry->Core.Key));
2211
2212 cbToWrite = RT_MIN(pEntry->cbData - offDiff, cbWrite);
2213 cbWrite -= cbToWrite;
2214
2215 if (!cbWrite)
2216 STAM_COUNTER_INC(&pCache->cHits);
2217 else
2218 STAM_COUNTER_INC(&pCache->cPartialHits);
2219
2220 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2221
2222 /* Ghost lists contain no data. */
2223 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2224 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2225 {
2226 /* Check if the entry is dirty. */
2227 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2228 PDMBLKCACHE_ENTRY_IS_DIRTY,
2229 0))
2230 {
2231 /* If it is already dirty but not in progress just update the data. */
2232 if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS))
2233 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff, cbToWrite);
2234 else
2235 {
2236 /* The data isn't written to the file yet */
2237 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2238 &SgBuf, offDiff, cbToWrite,
2239 true /* fWrite */);
2240 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2241 }
2242
2243 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2244 }
2245 else /* Dirty bit not set */
2246 {
2247 /*
2248 * Check if a read is in progress for this entry.
2249 * We have to defer processing in that case.
2250 */
2251 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2252 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2253 0))
2254 {
2255 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2256 &SgBuf, offDiff, cbToWrite,
2257 true /* fWrite */);
2258 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2259 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2260 }
2261 else /* I/O in progress flag not set */
2262 {
2263 /* Write as much as we can into the entry and update the file. */
2264 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff, cbToWrite);
2265
2266 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2267 if (fCommit)
2268 pdmBlkCacheCommitDirtyEntries(pCache);
2269 }
2270 } /* Dirty bit not set */
2271
2272 /* Move this entry to the top position */
2273 if (pEntry->pList == &pCache->LruFrequentlyUsed)
2274 {
2275 pdmBlkCacheLockEnter(pCache);
2276 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2277 pdmBlkCacheLockLeave(pCache);
2278 }
2279
2280 pdmBlkCacheEntryRelease(pEntry);
2281 }
2282 else /* Entry is on the ghost list */
2283 {
2284 uint8_t *pbBuffer = NULL;
2285
2286 pdmBlkCacheLockEnter(pCache);
2287 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2288 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2289
2290 if (fEnough)
2291 {
2292 /* Move the entry to Am and fetch it to the cache. */
2293 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2294 pdmBlkCacheAdd(pCache, pEntry->cbData);
2295 pdmBlkCacheLockLeave(pCache);
2296
2297 if (pbBuffer)
2298 pEntry->pbData = pbBuffer;
2299 else
2300 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2301 AssertPtr(pEntry->pbData);
2302
2303 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2304 &SgBuf, offDiff, cbToWrite,
2305 true /* fWrite */);
2306 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2307 pdmBlkCacheEntryReadFromMedium(pEntry);
2308
2309 /* Release the reference. If it is still needed the I/O in progress flag should protect it now. */
2310 pdmBlkCacheEntryRelease(pEntry);
2311 }
2312 else
2313 {
2314 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2315 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2316 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2317 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2318 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2319
2320 pdmBlkCacheLockLeave(pCache);
2321
2322 RTMemFree(pEntry);
2323 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2324 &SgBuf, off, cbToWrite,
2325 PDMBLKCACHEXFERDIR_WRITE);
2326 }
2327 }
2328 }
2329 else /* No entry found */
2330 {
2331 /*
2332 * No entry found. Try to create a new cache entry to store the data in and if that fails
2333 * write directly to the file.
2334 */
2335 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
2336 off, cbWrite,
2337 &cbToWrite);
2338
2339 cbWrite -= cbToWrite;
2340
2341 if (pEntryNew)
2342 {
2343 uint64_t offDiff = off - pEntryNew->Core.Key;
2344
2345 STAM_COUNTER_INC(&pCache->cHits);
2346
2347 /*
2348 * Check if it is possible to just write the data without waiting
2349 * for it to get fetched first.
2350 */
2351 if (!offDiff && pEntryNew->cbData == cbToWrite)
2352 {
2353 RTSgBufCopyToBuf(&SgBuf, pEntryNew->pbData, cbToWrite);
2354
2355 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntryNew);
2356 if (fCommit)
2357 pdmBlkCacheCommitDirtyEntries(pCache);
2358 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2359 }
2360 else
2361 {
2362 /* Defer the write and fetch the data from the endpoint. */
2363 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
2364 &SgBuf, offDiff, cbToWrite,
2365 true /* fWrite */);
2366 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2367 pdmBlkCacheEntryReadFromMedium(pEntryNew);
2368 }
2369
2370 pdmBlkCacheEntryRelease(pEntryNew);
2371 }
2372 else
2373 {
2374 /*
2375 * There is not enough free space in the cache.
2376 * Pass the request directly to the I/O manager.
2377 */
2378 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToWrite));
2379
2380 STAM_COUNTER_INC(&pCache->cMisses);
2381
2382 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2383 &SgBuf, off, cbToWrite,
2384 PDMBLKCACHEXFERDIR_WRITE);
2385 }
2386 }
2387
2388 off += cbToWrite;
2389 }
2390
2391 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, false))
2392 rc = VINF_AIO_TASK_PENDING;
2393 else
2394 {
2395 rc = pReq->rcReq;
2396 RTMemFree(pReq);
2397 }
2398
2399 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2400
2401 return rc;
2402}
2403
2404VMMR3DECL(int) PDMR3BlkCacheFlush(PPDMBLKCACHE pBlkCache, void *pvUser)
2405{
2406 int rc = VINF_SUCCESS;
2407 PPDMBLKCACHEREQ pReq;
2408
2409 LogFlowFunc((": pBlkCache=%#p{%s}\n", pBlkCache, pBlkCache->pszId));
2410
2411 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2412 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
2413
2414 /* Commit dirty entries in the cache. */
2415 pdmBlkCacheCommit(pBlkCache);
2416
2417 /* Allocate new request structure. */
2418 pReq = pdmBlkCacheReqAlloc(pvUser);
2419 if (RT_UNLIKELY(!pReq))
2420 return VERR_NO_MEMORY;
2421
2422 rc = pdmBlkCacheRequestPassthrough(pBlkCache, pReq, NULL, 0, 0,
2423 PDMBLKCACHEXFERDIR_FLUSH);
2424 AssertRC(rc);
2425
2426 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2427 return VINF_AIO_TASK_PENDING;
2428}
2429
2430VMMR3DECL(int) PDMR3BlkCacheDiscard(PPDMBLKCACHE pBlkCache, PCRTRANGE paRanges,
2431 unsigned cRanges, void *pvUser)
2432{
2433 int rc = VINF_SUCCESS;
2434 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2435 PPDMBLKCACHEENTRY pEntry;
2436 PPDMBLKCACHEREQ pReq;
2437
2438 LogFlowFunc((": pBlkCache=%#p{%s} paRanges=%#p cRanges=%u pvUser=%#p\n",
2439 pBlkCache, pBlkCache->pszId, paRanges, cRanges, pvUser));
2440
2441 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2442 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
2443
2444 /* Allocate new request structure. */
2445 pReq = pdmBlkCacheReqAlloc(pvUser);
2446 if (RT_UNLIKELY(!pReq))
2447 return VERR_NO_MEMORY;
2448
2449 /* Increment data transfer counter to keep the request valid while we access it. */
2450 ASMAtomicIncU32(&pReq->cXfersPending);
2451
2452 for (unsigned i = 0; i < cRanges; i++)
2453 {
2454 uint64_t offCur = paRanges[i].offStart;
2455 size_t cbLeft = paRanges[i].cbRange;
2456
2457 while (cbLeft)
2458 {
2459 size_t cbThisDiscard = 0;
2460
2461 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, offCur);
2462
2463 if (pEntry)
2464 {
2465 /* Write the data into the entry and mark it as dirty */
2466 AssertPtr(pEntry->pList);
2467
2468 uint64_t offDiff = offCur - pEntry->Core.Key;
2469
2470 AssertMsg(offCur >= pEntry->Core.Key,
2471 ("Overflow in calculation offCur=%llu OffsetAligned=%llu\n",
2472 offCur, pEntry->Core.Key));
2473
2474 cbThisDiscard = RT_MIN(pEntry->cbData - offDiff, cbLeft);
2475
2476 /* Ghost lists contain no data. */
2477 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2478 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2479 {
2480 /* Check if the entry is dirty. */
2481 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2482 PDMBLKCACHE_ENTRY_IS_DIRTY,
2483 0))
2484 {
2485 /* If it is dirty but not yet in progress remove it. */
2486 if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS))
2487 {
2488 pdmBlkCacheLockEnter(pCache);
2489 pdmBlkCacheEntryRemoveFromList(pEntry);
2490
2491 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2492 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2493 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2494
2495 pdmBlkCacheLockLeave(pCache);
2496
2497 RTMemFree(pEntry);
2498 }
2499 else
2500 {
2501#if 0
2502 /* The data isn't written to the file yet */
2503 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2504 &SgBuf, offDiff, cbToWrite,
2505 true /* fWrite */);
2506 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2507#endif
2508 }
2509
2510 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2511 pdmBlkCacheEntryRelease(pEntry);
2512 }
2513 else /* Dirty bit not set */
2514 {
2515 /*
2516 * Check if a read is in progress for this entry.
2517 * We have to defer processing in that case.
2518 */
2519 if(pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2520 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2521 0))
2522 {
2523#if 0
2524 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2525 &SgBuf, offDiff, cbToWrite,
2526 true /* fWrite */);
2527#endif
2528 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2529 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2530 pdmBlkCacheEntryRelease(pEntry);
2531 }
2532 else /* I/O in progress flag not set */
2533 {
2534 pdmBlkCacheLockEnter(pCache);
2535 pdmBlkCacheEntryRemoveFromList(pEntry);
2536
2537 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2538 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2539 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2540 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2541 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2542
2543 pdmBlkCacheLockLeave(pCache);
2544
2545 RTMemFree(pEntry);
2546 }
2547 } /* Dirty bit not set */
2548 }
2549 else /* Entry is on the ghost list just remove cache entry. */
2550 {
2551 pdmBlkCacheLockEnter(pCache);
2552 pdmBlkCacheEntryRemoveFromList(pEntry);
2553
2554 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2555 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2556 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2557 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2558 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2559
2560 pdmBlkCacheLockLeave(pCache);
2561
2562 RTMemFree(pEntry);
2563 }
2564 }
2565 /* else: no entry found. */
2566
2567 offCur += cbThisDiscard;
2568 cbLeft -= cbThisDiscard;
2569 }
2570 }
2571
2572 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, false))
2573 rc = VINF_AIO_TASK_PENDING;
2574 else
2575 {
2576 rc = pReq->rcReq;
2577 RTMemFree(pReq);
2578 }
2579
2580 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2581
2582 return rc;
2583}
2584
2585/**
2586 * Completes a task segment freeing all resources and completes the task handle
2587 * if everything was transferred.
2588 *
2589 * @returns Next task segment handle.
2590 * @param pBlkCache The endpoint block cache.
2591 * @param pWaiter Task segment to complete.
2592 * @param rc Status code to set.
2593 */
2594static PPDMBLKCACHEWAITER pdmBlkCacheWaiterComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEWAITER pWaiter, int rc)
2595{
2596 PPDMBLKCACHEWAITER pNext = pWaiter->pNext;
2597 PPDMBLKCACHEREQ pReq = pWaiter->pReq;
2598
2599 pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, true);
2600
2601 RTMemFree(pWaiter);
2602
2603 return pNext;
2604}
2605
2606static void pdmBlkCacheIoXferCompleteEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2607{
2608 PPDMBLKCACHEENTRY pEntry = hIoXfer->pEntry;
2609 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2610
2611 /* Reference the entry now as we are clearing the I/O in progress flag
2612 * which protected the entry till now. */
2613 pdmBlkCacheEntryRef(pEntry);
2614
2615 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2616 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
2617
2618 /* Process waiting segment list. The data in entry might have changed in-between. */
2619 bool fDirty = false;
2620 PPDMBLKCACHEWAITER pComplete = pEntry->pWaitingHead;
2621 PPDMBLKCACHEWAITER pCurr = pComplete;
2622
2623 AssertMsg((pCurr && pEntry->pWaitingTail) || (!pCurr && !pEntry->pWaitingTail),
2624 ("The list tail was not updated correctly\n"));
2625 pEntry->pWaitingTail = NULL;
2626 pEntry->pWaitingHead = NULL;
2627
2628 if (hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_WRITE)
2629 {
2630 /*
2631 * An error here is difficult to handle as the original request completed already.
2632 * The error is logged for now and the VM is paused.
2633 * If the user continues the entry is written again in the hope
2634 * the user fixed the problem and the next write succeeds.
2635 */
2636 if (RT_FAILURE(rcIoXfer))
2637 {
2638 LogRel(("I/O cache: Error while writing entry at offset %llu (%u bytes) to medium \"%s\" (rc=%Rrc)\n",
2639 pEntry->Core.Key, pEntry->cbData, pBlkCache->pszId, rcIoXfer));
2640
2641 if (!ASMAtomicXchgBool(&pCache->fIoErrorVmSuspended, true))
2642 {
2643 int rc = VMSetRuntimeError(pCache->pVM, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "BLKCACHE_IOERR",
2644 N_("The I/O cache encountered an error while updating data in medium \"%s\" (rc=%Rrc). "
2645 "Make sure there is enough free space on the disk and that the disk is working properly. "
2646 "Operation can be resumed afterwards"),
2647 pBlkCache->pszId, rcIoXfer);
2648 AssertRC(rc);
2649 }
2650
2651 /* Mark the entry as dirty again to get it added to the list later on. */
2652 fDirty = true;
2653 }
2654
2655 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IS_DIRTY;
2656
2657 while (pCurr)
2658 {
2659 AssertMsg(pCurr->fWrite, ("Completed write entries should never have read tasks attached\n"));
2660
2661 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2662 fDirty = true;
2663 pCurr = pCurr->pNext;
2664 }
2665 }
2666 else
2667 {
2668 AssertMsg(hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_READ, ("Invalid transfer type\n"));
2669 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY),
2670 ("Invalid flags set\n"));
2671
2672 while (pCurr)
2673 {
2674 if (pCurr->fWrite)
2675 {
2676 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2677 fDirty = true;
2678 }
2679 else
2680 RTSgBufCopyFromBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2681
2682 pCurr = pCurr->pNext;
2683 }
2684 }
2685
2686 bool fCommit = false;
2687 if (fDirty)
2688 fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2689
2690 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2691
2692 /* Dereference so that it isn't protected anymore except we issued anyother write for it. */
2693 pdmBlkCacheEntryRelease(pEntry);
2694
2695 if (fCommit)
2696 pdmBlkCacheCommitDirtyEntries(pCache);
2697
2698 /* Complete waiters now. */
2699 while (pComplete)
2700 pComplete = pdmBlkCacheWaiterComplete(pBlkCache, pComplete, rcIoXfer);
2701}
2702
2703VMMR3DECL(void) PDMR3BlkCacheIoXferComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2704{
2705 LogFlowFunc(("pBlkCache=%#p hIoXfer=%#p rcIoXfer=%Rrc\n", pBlkCache, hIoXfer, rcIoXfer));
2706
2707 if (hIoXfer->fIoCache)
2708 pdmBlkCacheIoXferCompleteEntry(pBlkCache, hIoXfer, rcIoXfer);
2709 else
2710 pdmBlkCacheReqUpdate(pBlkCache, hIoXfer->pReq, rcIoXfer, true);
2711
2712 ASMAtomicDecU32(&pBlkCache->cIoXfersActive);
2713 pdmBlkCacheR3TraceMsgF(pBlkCache, "BlkCache: I/O req %#p (%RTbool) completed (%u now active)",
2714 hIoXfer, hIoXfer->fIoCache, pBlkCache->cIoXfersActive);
2715 RTMemFree(hIoXfer);
2716}
2717
2718/**
2719 * Callback for the AVL do with all routine. Waits for a cachen entry to finish any pending I/O.
2720 *
2721 * @returns IPRT status code.
2722 * @param pNode The node to destroy.
2723 * @param pvUser Opaque user data.
2724 */
2725static DECLCALLBACK(int) pdmBlkCacheEntryQuiesce(PAVLRU64NODECORE pNode, void *pvUser)
2726{
2727 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)pNode;
2728 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
2729 NOREF(pvUser);
2730
2731 while (ASMAtomicReadU32(&pEntry->fFlags) & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS)
2732 {
2733 /* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
2734 pdmBlkCacheEntryRef(pEntry);
2735 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2736
2737 RTThreadSleep(1);
2738
2739 /* Re-enter all locks and drop the reference. */
2740 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2741 pdmBlkCacheEntryRelease(pEntry);
2742 }
2743
2744 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
2745 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
2746
2747 return VINF_SUCCESS;
2748}
2749
2750VMMR3DECL(int) PDMR3BlkCacheSuspend(PPDMBLKCACHE pBlkCache)
2751{
2752 int rc = VINF_SUCCESS;
2753 LogFlowFunc(("pBlkCache=%#p\n", pBlkCache));
2754
2755 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2756
2757 if (!ASMAtomicReadBool(&pBlkCache->pCache->fIoErrorVmSuspended))
2758 pdmBlkCacheCommit(pBlkCache); /* Can issue new I/O requests. */
2759 ASMAtomicXchgBool(&pBlkCache->fSuspended, true);
2760
2761 /* Wait for all I/O to complete. */
2762 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2763 rc = RTAvlrU64DoWithAll(pBlkCache->pTree, true, pdmBlkCacheEntryQuiesce, NULL);
2764 AssertRC(rc);
2765 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2766
2767 return rc;
2768}
2769
2770VMMR3DECL(int) PDMR3BlkCacheResume(PPDMBLKCACHE pBlkCache)
2771{
2772 LogFlowFunc(("pBlkCache=%#p\n", pBlkCache));
2773
2774 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2775
2776 ASMAtomicXchgBool(&pBlkCache->fSuspended, false);
2777
2778 return VINF_SUCCESS;
2779}
2780
2781VMMR3DECL(int) PDMR3BlkCacheClear(PPDMBLKCACHE pBlkCache)
2782{
2783 int rc = VINF_SUCCESS;
2784 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2785
2786 /*
2787 * Commit all dirty entries now (they are waited on for completion during the
2788 * destruction of the AVL tree below).
2789 * The exception is if the VM was paused because of an I/O error before.
2790 */
2791 if (!ASMAtomicReadBool(&pCache->fIoErrorVmSuspended))
2792 pdmBlkCacheCommit(pBlkCache);
2793
2794 /* Make sure nobody is accessing the cache while we delete the tree. */
2795 pdmBlkCacheLockEnter(pCache);
2796 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2797 RTAvlrU64Destroy(pBlkCache->pTree, pdmBlkCacheEntryDestroy, pCache);
2798 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2799
2800 pdmBlkCacheLockLeave(pCache);
2801 return rc;
2802}
2803
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