VirtualBox

source: vbox/trunk/src/VBox/VMM/VMMR3/PDMBlkCache.cpp@ 39042

Last change on this file since 39042 was 39042, checked in by vboxsync, 13 years ago

build fix
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 98.3 KB
Line 
1/* $Id: PDMBlkCache.cpp 39042 2011-10-19 14:55:57Z 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 VBOX_STRICT
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 VBOX_STRICT
119 pdmBlkCacheValidate(pCache);
120#endif
121}
122
123DECLINLINE(void) pdmBlkCacheLockLeave(PPDMBLKCACHEGLOBAL pCache)
124{
125#ifdef VBOX_STRICT
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, (uint32_t)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); NOREF(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 PFNPDMBLKCACHEXFERENQUEUEDISCARDDRV pfnXferEnqueueDiscard,
1274 const char *pcszId)
1275{
1276 int rc = VINF_SUCCESS;
1277 PPDMBLKCACHE pBlkCache;
1278
1279 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1280 if (RT_SUCCESS(rc))
1281 {
1282 pBlkCache->enmType = PDMBLKCACHETYPE_DRV;
1283 pBlkCache->u.Drv.pfnXferComplete = pfnXferComplete;
1284 pBlkCache->u.Drv.pfnXferEnqueue = pfnXferEnqueue;
1285 pBlkCache->u.Drv.pfnXferEnqueueDiscard = pfnXferEnqueueDiscard;
1286 pBlkCache->u.Drv.pDrvIns = pDrvIns;
1287 *ppBlkCache = pBlkCache;
1288 }
1289
1290 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1291 return rc;
1292}
1293
1294VMMR3DECL(int) PDMR3BlkCacheRetainDevice(PVM pVM, PPDMDEVINS pDevIns, PPPDMBLKCACHE ppBlkCache,
1295 PFNPDMBLKCACHEXFERCOMPLETEDEV pfnXferComplete,
1296 PFNPDMBLKCACHEXFERENQUEUEDEV pfnXferEnqueue,
1297 PFNPDMBLKCACHEXFERENQUEUEDISCARDDEV pfnXferEnqueueDiscard,
1298 const char *pcszId)
1299{
1300 int rc = VINF_SUCCESS;
1301 PPDMBLKCACHE pBlkCache;
1302
1303 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1304 if (RT_SUCCESS(rc))
1305 {
1306 pBlkCache->enmType = PDMBLKCACHETYPE_DEV;
1307 pBlkCache->u.Dev.pfnXferComplete = pfnXferComplete;
1308 pBlkCache->u.Dev.pfnXferEnqueue = pfnXferEnqueue;
1309 pBlkCache->u.Dev.pfnXferEnqueueDiscard = pfnXferEnqueueDiscard;
1310 pBlkCache->u.Dev.pDevIns = pDevIns;
1311 *ppBlkCache = pBlkCache;
1312 }
1313
1314 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1315 return rc;
1316
1317}
1318
1319VMMR3DECL(int) PDMR3BlkCacheRetainUsb(PVM pVM, PPDMUSBINS pUsbIns, PPPDMBLKCACHE ppBlkCache,
1320 PFNPDMBLKCACHEXFERCOMPLETEUSB pfnXferComplete,
1321 PFNPDMBLKCACHEXFERENQUEUEUSB pfnXferEnqueue,
1322 PFNPDMBLKCACHEXFERENQUEUEDISCARDUSB pfnXferEnqueueDiscard,
1323 const char *pcszId)
1324{
1325 int rc = VINF_SUCCESS;
1326 PPDMBLKCACHE pBlkCache;
1327
1328 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1329 if (RT_SUCCESS(rc))
1330 {
1331 pBlkCache->enmType = PDMBLKCACHETYPE_USB;
1332 pBlkCache->u.Usb.pfnXferComplete = pfnXferComplete;
1333 pBlkCache->u.Usb.pfnXferEnqueue = pfnXferEnqueue;
1334 pBlkCache->u.Usb.pfnXferEnqueueDiscard = pfnXferEnqueueDiscard;
1335 pBlkCache->u.Usb.pUsbIns = pUsbIns;
1336 *ppBlkCache = pBlkCache;
1337 }
1338
1339 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1340 return rc;
1341
1342}
1343
1344VMMR3DECL(int) PDMR3BlkCacheRetainInt(PVM pVM, void *pvUser, PPPDMBLKCACHE ppBlkCache,
1345 PFNPDMBLKCACHEXFERCOMPLETEINT pfnXferComplete,
1346 PFNPDMBLKCACHEXFERENQUEUEINT pfnXferEnqueue,
1347 PFNPDMBLKCACHEXFERENQUEUEDISCARDINT pfnXferEnqueueDiscard,
1348 const char *pcszId)
1349{
1350 int rc = VINF_SUCCESS;
1351 PPDMBLKCACHE pBlkCache;
1352
1353 rc = pdmR3BlkCacheRetain(pVM, &pBlkCache, pcszId);
1354 if (RT_SUCCESS(rc))
1355 {
1356 pBlkCache->enmType = PDMBLKCACHETYPE_INTERNAL;
1357 pBlkCache->u.Int.pfnXferComplete = pfnXferComplete;
1358 pBlkCache->u.Int.pfnXferEnqueue = pfnXferEnqueue;
1359 pBlkCache->u.Int.pfnXferEnqueueDiscard = pfnXferEnqueueDiscard;
1360 pBlkCache->u.Int.pvUser = pvUser;
1361 *ppBlkCache = pBlkCache;
1362 }
1363
1364 LogFlowFunc(("Leave rc=%Rrc\n", rc));
1365 return rc;
1366
1367}
1368
1369/**
1370 * Callback for the AVL destroy routine. Frees a cache entry for this endpoint.
1371 *
1372 * @returns IPRT status code.
1373 * @param pNode The node to destroy.
1374 * @param pvUser Opaque user data.
1375 */
1376static int pdmBlkCacheEntryDestroy(PAVLRU64NODECORE pNode, void *pvUser)
1377{
1378 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)pNode;
1379 PPDMBLKCACHEGLOBAL pCache = (PPDMBLKCACHEGLOBAL)pvUser;
1380 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
1381
1382 while (ASMAtomicReadU32(&pEntry->fFlags) & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS)
1383 {
1384 /* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
1385 pdmBlkCacheEntryRef(pEntry);
1386 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1387 pdmBlkCacheLockLeave(pCache);
1388
1389 RTThreadSleep(250);
1390
1391 /* Re-enter all locks */
1392 pdmBlkCacheLockEnter(pCache);
1393 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1394 pdmBlkCacheEntryRelease(pEntry);
1395 }
1396
1397 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
1398 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
1399
1400 bool fUpdateCache = pEntry->pList == &pCache->LruFrequentlyUsed
1401 || pEntry->pList == &pCache->LruRecentlyUsedIn;
1402
1403 pdmBlkCacheEntryRemoveFromList(pEntry);
1404
1405 if (fUpdateCache)
1406 pdmBlkCacheSub(pCache, pEntry->cbData);
1407
1408 RTMemPageFree(pEntry->pbData, pEntry->cbData);
1409 RTMemFree(pEntry);
1410
1411 return VINF_SUCCESS;
1412}
1413
1414/**
1415 * Destroys all cache resources used by the given endpoint.
1416 *
1417 * @returns nothing.
1418 * @param pEndpoint The endpoint to the destroy.
1419 */
1420VMMR3DECL(void) PDMR3BlkCacheRelease(PPDMBLKCACHE pBlkCache)
1421{
1422 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1423
1424 /*
1425 * Commit all dirty entries now (they are waited on for completion during the
1426 * destruction of the AVL tree below).
1427 * The exception is if the VM was paused because of an I/O error before.
1428 */
1429 if (!ASMAtomicReadBool(&pCache->fIoErrorVmSuspended))
1430 pdmBlkCacheCommit(pBlkCache);
1431
1432 /* Make sure nobody is accessing the cache while we delete the tree. */
1433 pdmBlkCacheLockEnter(pCache);
1434 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1435 RTAvlrU64Destroy(pBlkCache->pTree, pdmBlkCacheEntryDestroy, pCache);
1436 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1437
1438 RTSpinlockDestroy(pBlkCache->LockList);
1439
1440 pCache->cRefs--;
1441 RTListNodeRemove(&pBlkCache->NodeCacheUser);
1442
1443 pdmBlkCacheLockLeave(pCache);
1444
1445 RTSemRWDestroy(pBlkCache->SemRWEntries);
1446
1447#ifdef VBOX_WITH_STATISTICS
1448 STAMR3Deregister(pCache->pVM, &pBlkCache->StatWriteDeferred);
1449#endif
1450
1451 RTStrFree(pBlkCache->pszId);
1452 RTMemFree(pBlkCache);
1453}
1454
1455VMMR3DECL(void) PDMR3BlkCacheReleaseDevice(PVM pVM, PPDMDEVINS pDevIns)
1456{
1457 LogFlow(("%s: pDevIns=%p\n", __FUNCTION__, pDevIns));
1458
1459 /*
1460 * Validate input.
1461 */
1462 if (!pDevIns)
1463 return;
1464 VM_ASSERT_EMT(pVM);
1465
1466 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1467 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1468
1469 /* Return silently if not supported. */
1470 if (!pBlkCacheGlobal)
1471 return;
1472
1473 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1474
1475 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1476 {
1477 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DEV
1478 && pBlkCache->u.Dev.pDevIns == pDevIns)
1479 PDMR3BlkCacheRelease(pBlkCache);
1480 }
1481
1482 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1483}
1484
1485VMMR3DECL(void) PDMR3BlkCacheReleaseDriver(PVM pVM, PPDMDRVINS pDrvIns)
1486{
1487 LogFlow(("%s: pDrvIns=%p\n", __FUNCTION__, pDrvIns));
1488
1489 /*
1490 * Validate input.
1491 */
1492 if (!pDrvIns)
1493 return;
1494 VM_ASSERT_EMT(pVM);
1495
1496 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1497 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1498
1499 /* Return silently if not supported. */
1500 if (!pBlkCacheGlobal)
1501 return;
1502
1503 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1504
1505 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1506 {
1507 if ( pBlkCache->enmType == PDMBLKCACHETYPE_DRV
1508 && pBlkCache->u.Drv.pDrvIns == pDrvIns)
1509 PDMR3BlkCacheRelease(pBlkCache);
1510 }
1511
1512 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1513}
1514
1515VMMR3DECL(void) PDMR3BlkCacheReleaseUsb(PVM pVM, PPDMUSBINS pUsbIns)
1516{
1517 LogFlow(("%s: pUsbIns=%p\n", __FUNCTION__, pUsbIns));
1518
1519 /*
1520 * Validate input.
1521 */
1522 if (!pUsbIns)
1523 return;
1524 VM_ASSERT_EMT(pVM);
1525
1526 PPDMBLKCACHEGLOBAL pBlkCacheGlobal = pVM->pUVM->pdm.s.pBlkCacheGlobal;
1527 PPDMBLKCACHE pBlkCache, pBlkCacheNext;
1528
1529 /* Return silently if not supported. */
1530 if (!pBlkCacheGlobal)
1531 return;
1532
1533 pdmBlkCacheLockEnter(pBlkCacheGlobal);
1534
1535 RTListForEachSafe(&pBlkCacheGlobal->ListUsers, pBlkCache, pBlkCacheNext, PDMBLKCACHE, NodeCacheUser)
1536 {
1537 if ( pBlkCache->enmType == PDMBLKCACHETYPE_USB
1538 && pBlkCache->u.Usb.pUsbIns == pUsbIns)
1539 PDMR3BlkCacheRelease(pBlkCache);
1540 }
1541
1542 pdmBlkCacheLockLeave(pBlkCacheGlobal);
1543}
1544
1545static PPDMBLKCACHEENTRY pdmBlkCacheGetCacheEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off)
1546{
1547 STAM_PROFILE_ADV_START(&pBlkCache->pCache->StatTreeGet, Cache);
1548
1549 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1550 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)RTAvlrU64RangeGet(pBlkCache->pTree, off);
1551 if (pEntry)
1552 pdmBlkCacheEntryRef(pEntry);
1553 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1554
1555 STAM_PROFILE_ADV_STOP(&pBlkCache->pCache->StatTreeGet, Cache);
1556
1557 return pEntry;
1558}
1559
1560/**
1561 * Return the best fit cache entries for the given offset.
1562 *
1563 * @returns nothing.
1564 * @param pBlkCache The endpoint cache.
1565 * @param off The offset.
1566 * @param pEntryAbove Where to store the pointer to the best fit entry above the
1567 * the given offset. NULL if not required.
1568 */
1569static void pdmBlkCacheGetCacheBestFitEntryByOffset(PPDMBLKCACHE pBlkCache, uint64_t off,
1570 PPDMBLKCACHEENTRY *ppEntryAbove)
1571{
1572 STAM_PROFILE_ADV_START(&pBlkCache->pCache->StatTreeGet, Cache);
1573
1574 RTSemRWRequestRead(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1575 if (ppEntryAbove)
1576 {
1577 *ppEntryAbove = (PPDMBLKCACHEENTRY)RTAvlrU64GetBestFit(pBlkCache->pTree, off, true /*fAbove*/);
1578 if (*ppEntryAbove)
1579 pdmBlkCacheEntryRef(*ppEntryAbove);
1580 }
1581
1582 RTSemRWReleaseRead(pBlkCache->SemRWEntries);
1583
1584 STAM_PROFILE_ADV_STOP(&pBlkCache->pCache->StatTreeGet, Cache);
1585}
1586
1587static void pdmBlkCacheInsertEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEENTRY pEntry)
1588{
1589 STAM_PROFILE_ADV_START(&pBlkCache->pCache->StatTreeInsert, Cache);
1590 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1591 bool fInserted = RTAvlrU64Insert(pBlkCache->pTree, &pEntry->Core);
1592 AssertMsg(fInserted, ("Node was not inserted into tree\n")); NOREF(fInserted);
1593 STAM_PROFILE_ADV_STOP(&pBlkCache->pCache->StatTreeInsert, Cache);
1594 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1595}
1596
1597/**
1598 * Allocates and initializes a new entry for the cache.
1599 * The entry has a reference count of 1.
1600 *
1601 * @returns Pointer to the new cache entry or NULL if out of memory.
1602 * @param pBlkCache The cache the entry belongs to.
1603 * @param off Start offset.
1604 * @param cbData Size of the cache entry.
1605 * @param pbBuffer Pointer to the buffer to use.
1606 * NULL if a new buffer should be allocated.
1607 * The buffer needs to have the same size of the entry.
1608 */
1609static PPDMBLKCACHEENTRY pdmBlkCacheEntryAlloc(PPDMBLKCACHE pBlkCache,
1610 uint64_t off, size_t cbData, uint8_t *pbBuffer)
1611{
1612 AssertReturn(cbData <= UINT32_MAX, NULL);
1613 PPDMBLKCACHEENTRY pEntryNew = (PPDMBLKCACHEENTRY)RTMemAllocZ(sizeof(PDMBLKCACHEENTRY));
1614
1615 if (RT_UNLIKELY(!pEntryNew))
1616 return NULL;
1617
1618 pEntryNew->Core.Key = off;
1619 pEntryNew->Core.KeyLast = off + cbData - 1;
1620 pEntryNew->pBlkCache = pBlkCache;
1621 pEntryNew->fFlags = 0;
1622 pEntryNew->cRefs = 1; /* We are using it now. */
1623 pEntryNew->pList = NULL;
1624 pEntryNew->cbData = (uint32_t)cbData;
1625 pEntryNew->pWaitingHead = NULL;
1626 pEntryNew->pWaitingTail = NULL;
1627 if (pbBuffer)
1628 pEntryNew->pbData = pbBuffer;
1629 else
1630 pEntryNew->pbData = (uint8_t *)RTMemPageAlloc(cbData);
1631
1632 if (RT_UNLIKELY(!pEntryNew->pbData))
1633 {
1634 RTMemFree(pEntryNew);
1635 return NULL;
1636 }
1637
1638 return pEntryNew;
1639}
1640
1641/**
1642 * Checks that a set of flags is set/clear acquiring the R/W semaphore
1643 * in exclusive mode.
1644 *
1645 * @returns true if the flag in fSet is set and the one in fClear is clear.
1646 * false otherwise.
1647 * The R/W semaphore is only held if true is returned.
1648 *
1649 * @param pBlkCache The endpoint cache instance data.
1650 * @param pEntry The entry to check the flags for.
1651 * @param fSet The flag which is tested to be set.
1652 * @param fClear The flag which is tested to be clear.
1653 */
1654DECLINLINE(bool) pdmBlkCacheEntryFlagIsSetClearAcquireLock(PPDMBLKCACHE pBlkCache,
1655 PPDMBLKCACHEENTRY pEntry,
1656 uint32_t fSet, uint32_t fClear)
1657{
1658 uint32_t fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1659 bool fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1660
1661 if (fPassed)
1662 {
1663 /* Acquire the lock and check again because the completion callback might have raced us. */
1664 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
1665
1666 fFlags = ASMAtomicReadU32(&pEntry->fFlags);
1667 fPassed = ((fFlags & fSet) && !(fFlags & fClear));
1668
1669 /* Drop the lock if we didn't passed the test. */
1670 if (!fPassed)
1671 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
1672 }
1673
1674 return fPassed;
1675}
1676
1677/**
1678 * Adds a segment to the waiting list for a cache entry
1679 * which is currently in progress.
1680 *
1681 * @returns nothing.
1682 * @param pEntry The cache entry to add the segment to.
1683 * @param pSeg The segment to add.
1684 */
1685DECLINLINE(void) pdmBlkCacheEntryAddWaiter(PPDMBLKCACHEENTRY pEntry,
1686 PPDMBLKCACHEWAITER pWaiter)
1687{
1688 pWaiter->pNext = NULL;
1689
1690 if (pEntry->pWaitingHead)
1691 {
1692 AssertPtr(pEntry->pWaitingTail);
1693
1694 pEntry->pWaitingTail->pNext = pWaiter;
1695 pEntry->pWaitingTail = pWaiter;
1696 }
1697 else
1698 {
1699 Assert(!pEntry->pWaitingTail);
1700
1701 pEntry->pWaitingHead = pWaiter;
1702 pEntry->pWaitingTail = pWaiter;
1703 }
1704}
1705
1706/**
1707 * Add a buffer described by the I/O memory context
1708 * to the entry waiting for completion.
1709 *
1710 * @returns VBox status code.
1711 * @param pEntry The entry to add the buffer to.
1712 * @param pTask Task associated with the buffer.
1713 * @param pIoMemCtx The memory context to use.
1714 * @param offDiff Offset from the start of the buffer
1715 * in the entry.
1716 * @param cbData Amount of data to wait for onthis entry.
1717 * @param fWrite Flag whether the task waits because it wants to write
1718 * to the cache entry.
1719 */
1720static int pdmBlkCacheEntryWaitersAdd(PPDMBLKCACHEENTRY pEntry,
1721 PPDMBLKCACHEREQ pReq,
1722 PRTSGBUF pSgBuf, uint64_t offDiff,
1723 size_t cbData, bool fWrite)
1724{
1725 PPDMBLKCACHEWAITER pWaiter = (PPDMBLKCACHEWAITER)RTMemAllocZ(sizeof(PDMBLKCACHEWAITER));
1726 if (!pWaiter)
1727 return VERR_NO_MEMORY;
1728
1729 ASMAtomicIncU32(&pReq->cXfersPending);
1730 pWaiter->pReq = pReq;
1731 pWaiter->offCacheEntry = offDiff;
1732 pWaiter->cbTransfer = cbData;
1733 pWaiter->fWrite = fWrite;
1734 RTSgBufClone(&pWaiter->SgBuf, pSgBuf);
1735 RTSgBufAdvance(pSgBuf, cbData);
1736
1737 pdmBlkCacheEntryAddWaiter(pEntry, pWaiter);
1738
1739 return VINF_SUCCESS;
1740}
1741
1742/**
1743 * Calculate aligned offset and size for a new cache entry which do not
1744 * intersect with an already existing entry and the file end.
1745 *
1746 * @returns The number of bytes the entry can hold of the requested amount
1747 * of byte.
1748 * @param pEndpoint The endpoint.
1749 * @param pBlkCache The endpoint cache.
1750 * @param off The start offset.
1751 * @param cb The number of bytes the entry needs to hold at
1752 * least.
1753 * @param uAlignment Alignment of the boundary sizes.
1754 * @param poffAligned Where to store the aligned offset.
1755 * @param pcbAligned Where to store the aligned size of the entry.
1756 */
1757static uint32_t pdmBlkCacheEntryBoundariesCalc(PPDMBLKCACHE pBlkCache,
1758 uint64_t off, uint32_t cb,
1759 unsigned uAlignment,
1760 uint64_t *poffAligned, uint32_t *pcbAligned)
1761{
1762 /* Get the best fit entries around the offset */
1763 PPDMBLKCACHEENTRY pEntryAbove = NULL;
1764 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off, &pEntryAbove);
1765
1766 /* Log the info */
1767 LogFlow(("%sest fit entry above off=%llu (BestFit=%llu BestFitEnd=%llu BestFitSize=%u)\n",
1768 pEntryAbove ? "B" : "No b",
1769 off,
1770 pEntryAbove ? pEntryAbove->Core.Key : 0,
1771 pEntryAbove ? pEntryAbove->Core.KeyLast : 0,
1772 pEntryAbove ? pEntryAbove->cbData : 0));
1773
1774 /** @todo r=bird: Why is uAlignment disregarded here? */
1775 uint64_t offAligned = off;
1776
1777 uint32_t cbAligned;
1778 uint32_t cbInEntry;
1779 if ( pEntryAbove
1780 && off + cb > pEntryAbove->Core.Key)
1781 {
1782 cbInEntry = (uint32_t)(pEntryAbove->Core.Key - off);
1783 cbAligned = (uint32_t)(pEntryAbove->Core.Key - offAligned);
1784 }
1785 else
1786 {
1787 cbInEntry = cb;
1788 cbAligned = cb;
1789 }
1790
1791 /* A few sanity checks */
1792 AssertMsg(!pEntryAbove || offAligned + cbAligned <= pEntryAbove->Core.Key,
1793 ("Aligned size intersects with another cache entry\n"));
1794 Assert(cbInEntry <= cbAligned);
1795
1796 if (pEntryAbove)
1797 pdmBlkCacheEntryRelease(pEntryAbove);
1798
1799 LogFlow(("offAligned=%llu cbAligned=%u\n", offAligned, cbAligned));
1800
1801 *poffAligned = offAligned;
1802 *pcbAligned = cbAligned;
1803
1804 return cbInEntry;
1805}
1806
1807/**
1808 * Create a new cache entry evicting data from the cache if required.
1809 *
1810 * @returns Pointer to the new cache entry or NULL
1811 * if not enough bytes could be evicted from the cache.
1812 * @param pEndpoint The endpoint.
1813 * @param pBlkCache The endpoint cache.
1814 * @param off The offset.
1815 * @param cb Number of bytes the cache entry should have.
1816 * @param uAlignment Alignment the size of the entry should have.
1817 * @param pcbData Where to store the number of bytes the new
1818 * entry can hold. May be lower than actually requested
1819 * due to another entry intersecting the access range.
1820 */
1821static PPDMBLKCACHEENTRY pdmBlkCacheEntryCreate(PPDMBLKCACHE pBlkCache,
1822 uint64_t off, size_t cb,
1823 unsigned uAlignment,
1824 size_t *pcbData)
1825{
1826 uint64_t offStart = 0;
1827 uint32_t cbEntry = 0;
1828 *pcbData = pdmBlkCacheEntryBoundariesCalc(pBlkCache, off, (uint32_t)cb, uAlignment,
1829 &offStart, &cbEntry);
1830 AssertReturn(cb <= UINT32_MAX, NULL);
1831
1832 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1833 pdmBlkCacheLockEnter(pCache);
1834
1835 PPDMBLKCACHEENTRY pEntryNew = NULL;
1836 uint8_t *pbBuffer = NULL;
1837 bool fEnough = pdmBlkCacheReclaim(pCache, cbEntry, true, &pbBuffer);
1838 if (fEnough)
1839 {
1840 LogFlow(("Evicted enough bytes (%u requested). Creating new cache entry\n", cbEntry));
1841
1842 pEntryNew = pdmBlkCacheEntryAlloc(pBlkCache, offStart, cbEntry, pbBuffer);
1843 if (RT_LIKELY(pEntryNew))
1844 {
1845 pdmBlkCacheEntryAddToList(&pCache->LruRecentlyUsedIn, pEntryNew);
1846 pdmBlkCacheAdd(pCache, cbEntry);
1847 pdmBlkCacheLockLeave(pCache);
1848
1849 pdmBlkCacheInsertEntry(pBlkCache, pEntryNew);
1850
1851 AssertMsg( (off >= pEntryNew->Core.Key)
1852 && (off + *pcbData <= pEntryNew->Core.KeyLast + 1),
1853 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
1854 off, pEntryNew->Core.Key));
1855 }
1856 else
1857 pdmBlkCacheLockLeave(pCache);
1858 }
1859 else
1860 pdmBlkCacheLockLeave(pCache);
1861
1862 return pEntryNew;
1863}
1864
1865static PPDMBLKCACHEREQ pdmBlkCacheReqAlloc(void *pvUser)
1866{
1867 PPDMBLKCACHEREQ pReq = (PPDMBLKCACHEREQ)RTMemAlloc(sizeof(PDMBLKCACHEREQ));
1868
1869 if (RT_LIKELY(pReq))
1870 {
1871 pReq->pvUser = pvUser;
1872 pReq->rcReq = VINF_SUCCESS;
1873 pReq->cXfersPending = 0;
1874 }
1875
1876 return pReq;
1877}
1878
1879static void pdmBlkCacheReqComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq)
1880{
1881 switch (pBlkCache->enmType)
1882 {
1883 case PDMBLKCACHETYPE_DEV:
1884 {
1885 pBlkCache->u.Dev.pfnXferComplete(pBlkCache->u.Dev.pDevIns,
1886 pReq->pvUser, pReq->rcReq);
1887 break;
1888 }
1889 case PDMBLKCACHETYPE_DRV:
1890 {
1891 pBlkCache->u.Drv.pfnXferComplete(pBlkCache->u.Drv.pDrvIns,
1892 pReq->pvUser, pReq->rcReq);
1893 break;
1894 }
1895 case PDMBLKCACHETYPE_USB:
1896 {
1897 pBlkCache->u.Usb.pfnXferComplete(pBlkCache->u.Usb.pUsbIns,
1898 pReq->pvUser, pReq->rcReq);
1899 break;
1900 }
1901 case PDMBLKCACHETYPE_INTERNAL:
1902 {
1903 pBlkCache->u.Int.pfnXferComplete(pBlkCache->u.Int.pvUser,
1904 pReq->pvUser, pReq->rcReq);
1905 break;
1906 }
1907 default:
1908 AssertMsgFailed(("Unknown block cache type!\n"));
1909 }
1910
1911 RTMemFree(pReq);
1912}
1913
1914static bool pdmBlkCacheReqUpdate(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEREQ pReq,
1915 int rcReq, bool fCallHandler)
1916{
1917 if (RT_FAILURE(rcReq))
1918 ASMAtomicCmpXchgS32(&pReq->rcReq, rcReq, VINF_SUCCESS);
1919
1920 AssertMsg(pReq->cXfersPending > 0, ("No transfers are pending for this request\n"));
1921 uint32_t cXfersPending = ASMAtomicDecU32(&pReq->cXfersPending);
1922
1923 if (!cXfersPending)
1924 {
1925 if (fCallHandler)
1926 pdmBlkCacheReqComplete(pBlkCache, pReq);
1927 else
1928 RTMemFree(pReq);
1929 return true;
1930 }
1931
1932 LogFlowFunc(("pReq=%#p cXfersPending=%u\n", pReq, cXfersPending));
1933 return false;
1934}
1935
1936VMMR3DECL(int) PDMR3BlkCacheRead(PPDMBLKCACHE pBlkCache, uint64_t off,
1937 PCRTSGBUF pcSgBuf, size_t cbRead, void *pvUser)
1938{
1939 int rc = VINF_SUCCESS;
1940 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
1941 PPDMBLKCACHEENTRY pEntry;
1942 PPDMBLKCACHEREQ pReq;
1943
1944 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pcSgBuf=%#p cbRead=%u pvUser=%#p\n",
1945 pBlkCache, pBlkCache->pszId, off, pcSgBuf, cbRead, pvUser));
1946
1947 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
1948 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
1949
1950 RTSGBUF SgBuf;
1951 RTSgBufClone(&SgBuf, pcSgBuf);
1952
1953 /* Allocate new request structure. */
1954 pReq = pdmBlkCacheReqAlloc(pvUser);
1955 if (RT_UNLIKELY(!pReq))
1956 return VERR_NO_MEMORY;
1957
1958 /* Increment data transfer counter to keep the request valid while we access it. */
1959 ASMAtomicIncU32(&pReq->cXfersPending);
1960
1961 while (cbRead)
1962 {
1963 size_t cbToRead;
1964
1965 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
1966
1967 /*
1968 * If there is no entry we try to create a new one eviciting unused pages
1969 * if the cache is full. If this is not possible we will pass the request through
1970 * and skip the caching (all entries may be still in progress so they can't
1971 * be evicted)
1972 * If we have an entry it can be in one of the LRU lists where the entry
1973 * contains data (recently used or frequently used LRU) so we can just read
1974 * the data we need and put the entry at the head of the frequently used LRU list.
1975 * In case the entry is in one of the ghost lists it doesn't contain any data.
1976 * We have to fetch it again evicting pages from either T1 or T2 to make room.
1977 */
1978 if (pEntry)
1979 {
1980 uint64_t offDiff = off - pEntry->Core.Key;
1981
1982 AssertMsg(off >= pEntry->Core.Key,
1983 ("Overflow in calculation off=%llu OffsetAligned=%llu\n",
1984 off, pEntry->Core.Key));
1985
1986 AssertPtr(pEntry->pList);
1987
1988 cbToRead = RT_MIN(pEntry->cbData - offDiff, cbRead);
1989
1990 AssertMsg(off + cbToRead <= pEntry->Core.Key + pEntry->Core.KeyLast + 1,
1991 ("Buffer of cache entry exceeded off=%llu cbToRead=%d\n",
1992 off, cbToRead));
1993
1994 cbRead -= cbToRead;
1995
1996 if (!cbRead)
1997 STAM_COUNTER_INC(&pCache->cHits);
1998 else
1999 STAM_COUNTER_INC(&pCache->cPartialHits);
2000
2001 STAM_COUNTER_ADD(&pCache->StatRead, cbToRead);
2002
2003 /* Ghost lists contain no data. */
2004 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2005 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2006 {
2007 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2008 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2009 PDMBLKCACHE_ENTRY_IS_DIRTY))
2010 {
2011 /* Entry didn't completed yet. Append to the list */
2012 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2013 &SgBuf, offDiff, cbToRead,
2014 false /* fWrite */);
2015 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2016 }
2017 else
2018 {
2019 /* Read as much as we can from the entry. */
2020 RTSgBufCopyFromBuf(&SgBuf, pEntry->pbData + offDiff, cbToRead);
2021 }
2022
2023 /* Move this entry to the top position */
2024 if (pEntry->pList == &pCache->LruFrequentlyUsed)
2025 {
2026 pdmBlkCacheLockEnter(pCache);
2027 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2028 pdmBlkCacheLockLeave(pCache);
2029 }
2030 /* Release the entry */
2031 pdmBlkCacheEntryRelease(pEntry);
2032 }
2033 else
2034 {
2035 uint8_t *pbBuffer = NULL;
2036
2037 LogFlow(("Fetching data for ghost entry %#p from file\n", pEntry));
2038
2039 pdmBlkCacheLockEnter(pCache);
2040 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2041 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2042
2043 /* Move the entry to Am and fetch it to the cache. */
2044 if (fEnough)
2045 {
2046 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2047 pdmBlkCacheAdd(pCache, pEntry->cbData);
2048 pdmBlkCacheLockLeave(pCache);
2049
2050 if (pbBuffer)
2051 pEntry->pbData = pbBuffer;
2052 else
2053 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2054 AssertPtr(pEntry->pbData);
2055
2056 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2057 &SgBuf, offDiff, cbToRead,
2058 false /* fWrite */);
2059 pdmBlkCacheEntryReadFromMedium(pEntry);
2060 /* Release the entry */
2061 pdmBlkCacheEntryRelease(pEntry);
2062 }
2063 else
2064 {
2065 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2066 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2067 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2068 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2069 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2070
2071 pdmBlkCacheLockLeave(pCache);
2072
2073 RTMemFree(pEntry);
2074
2075 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2076 &SgBuf, off, cbToRead,
2077 PDMBLKCACHEXFERDIR_READ);
2078 }
2079 }
2080 }
2081 else
2082 {
2083#ifdef VBOX_WITH_IO_READ_CACHE
2084 /* No entry found for this offset. Create a new entry and fetch the data to the cache. */
2085 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
2086 off, cbRead,
2087 PAGE_SIZE,
2088 &cbToRead);
2089
2090 cbRead -= cbToRead;
2091
2092 if (pEntryNew)
2093 {
2094 if (!cbRead)
2095 STAM_COUNTER_INC(&pCache->cMisses);
2096 else
2097 STAM_COUNTER_INC(&pCache->cPartialHits);
2098
2099 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
2100 &SgBuf,
2101 off - pEntryNew->Core.Key,
2102 cbToRead,
2103 false /* fWrite */);
2104 pdmBlkCacheEntryReadFromMedium(pEntryNew);
2105 pdmBlkCacheEntryRelease(pEntryNew); /* it is protected by the I/O in progress flag now. */
2106 }
2107 else
2108 {
2109 /*
2110 * There is not enough free space in the cache.
2111 * Pass the request directly to the I/O manager.
2112 */
2113 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToRead));
2114
2115 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2116 &SgBuf, off, cbToRead,
2117 PDMBLKCACHEXFERDIR_READ);
2118 }
2119#else
2120 /* Clip read size if necessary. */
2121 PPDMBLKCACHEENTRY pEntryAbove;
2122 pdmBlkCacheGetCacheBestFitEntryByOffset(pBlkCache, off, &pEntryAbove);
2123
2124 if (pEntryAbove)
2125 {
2126 if (off + cbRead > pEntryAbove->Core.Key)
2127 cbToRead = pEntryAbove->Core.Key - off;
2128 else
2129 cbToRead = cbRead;
2130
2131 pdmBlkCacheEntryRelease(pEntryAbove);
2132 }
2133 else
2134 cbToRead = cbRead;
2135
2136 cbRead -= cbToRead;
2137 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2138 &SgBuf, off, cbToRead,
2139 PDMBLKCACHEXFERDIR_READ);
2140#endif
2141 }
2142 off += cbToRead;
2143 }
2144
2145 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, false))
2146 rc = VINF_AIO_TASK_PENDING;
2147
2148 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2149
2150 return rc;
2151}
2152
2153VMMR3DECL(int) PDMR3BlkCacheWrite(PPDMBLKCACHE pBlkCache, uint64_t off,
2154 PCRTSGBUF pcSgBuf, size_t cbWrite, void *pvUser)
2155{
2156 int rc = VINF_SUCCESS;
2157 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2158 PPDMBLKCACHEENTRY pEntry;
2159 PPDMBLKCACHEREQ pReq;
2160
2161 LogFlowFunc((": pBlkCache=%#p{%s} off=%llu pcSgBuf=%#p cbWrite=%u pvUser=%#p\n",
2162 pBlkCache, pBlkCache->pszId, off, pcSgBuf, cbWrite, pvUser));
2163
2164 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2165 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
2166
2167 RTSGBUF SgBuf;
2168 RTSgBufClone(&SgBuf, pcSgBuf);
2169
2170 /* Allocate new request structure. */
2171 pReq = pdmBlkCacheReqAlloc(pvUser);
2172 if (RT_UNLIKELY(!pReq))
2173 return VERR_NO_MEMORY;
2174
2175 /* Increment data transfer counter to keep the request valid while we access it. */
2176 ASMAtomicIncU32(&pReq->cXfersPending);
2177
2178 while (cbWrite)
2179 {
2180 size_t cbToWrite;
2181
2182 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, off);
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 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff, cbToWrite);
2216 else
2217 {
2218 /* The data isn't written to the file yet */
2219 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2220 &SgBuf, offDiff, cbToWrite,
2221 true /* fWrite */);
2222 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2223 }
2224
2225 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2226 }
2227 else /* Dirty bit not set */
2228 {
2229 /*
2230 * Check if a read is in progress for this entry.
2231 * We have to defer processing in that case.
2232 */
2233 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2234 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2235 0))
2236 {
2237 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2238 &SgBuf, offDiff, cbToWrite,
2239 true /* fWrite */);
2240 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2241 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2242 }
2243 else /* I/O in progress flag not set */
2244 {
2245 /* Write as much as we can into the entry and update the file. */
2246 RTSgBufCopyToBuf(&SgBuf, pEntry->pbData + offDiff, cbToWrite);
2247
2248 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2249 if (fCommit)
2250 pdmBlkCacheCommitDirtyEntries(pCache);
2251 }
2252 } /* Dirty bit not set */
2253
2254 /* Move this entry to the top position */
2255 if (pEntry->pList == &pCache->LruFrequentlyUsed)
2256 {
2257 pdmBlkCacheLockEnter(pCache);
2258 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2259 pdmBlkCacheLockLeave(pCache);
2260 }
2261
2262 pdmBlkCacheEntryRelease(pEntry);
2263 }
2264 else /* Entry is on the ghost list */
2265 {
2266 uint8_t *pbBuffer = NULL;
2267
2268 pdmBlkCacheLockEnter(pCache);
2269 pdmBlkCacheEntryRemoveFromList(pEntry); /* Remove it before we remove data, otherwise it may get freed when evicting data. */
2270 bool fEnough = pdmBlkCacheReclaim(pCache, pEntry->cbData, true, &pbBuffer);
2271
2272 if (fEnough)
2273 {
2274 /* Move the entry to Am and fetch it to the cache. */
2275 pdmBlkCacheEntryAddToList(&pCache->LruFrequentlyUsed, pEntry);
2276 pdmBlkCacheAdd(pCache, pEntry->cbData);
2277 pdmBlkCacheLockLeave(pCache);
2278
2279 if (pbBuffer)
2280 pEntry->pbData = pbBuffer;
2281 else
2282 pEntry->pbData = (uint8_t *)RTMemPageAlloc(pEntry->cbData);
2283 AssertPtr(pEntry->pbData);
2284
2285 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2286 &SgBuf, offDiff, cbToWrite,
2287 true /* fWrite */);
2288 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2289 pdmBlkCacheEntryReadFromMedium(pEntry);
2290
2291 /* Release the reference. If it is still needed the I/O in progress flag should protect it now. */
2292 pdmBlkCacheEntryRelease(pEntry);
2293 }
2294 else
2295 {
2296 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2297 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2298 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2299 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2300 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2301
2302 pdmBlkCacheLockLeave(pCache);
2303
2304 RTMemFree(pEntry);
2305 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2306 &SgBuf, off, cbToWrite,
2307 PDMBLKCACHEXFERDIR_WRITE);
2308 }
2309 }
2310 }
2311 else /* No entry found */
2312 {
2313 /*
2314 * No entry found. Try to create a new cache entry to store the data in and if that fails
2315 * write directly to the file.
2316 */
2317 PPDMBLKCACHEENTRY pEntryNew = pdmBlkCacheEntryCreate(pBlkCache,
2318 off, cbWrite,
2319 512, &cbToWrite);
2320
2321 cbWrite -= cbToWrite;
2322
2323 if (pEntryNew)
2324 {
2325 uint64_t offDiff = off - pEntryNew->Core.Key;
2326
2327 STAM_COUNTER_INC(&pCache->cHits);
2328
2329 /*
2330 * Check if it is possible to just write the data without waiting
2331 * for it to get fetched first.
2332 */
2333 if (!offDiff && pEntryNew->cbData == cbToWrite)
2334 {
2335 RTSgBufCopyToBuf(&SgBuf, pEntryNew->pbData, cbToWrite);
2336
2337 bool fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntryNew);
2338 if (fCommit)
2339 pdmBlkCacheCommitDirtyEntries(pCache);
2340 STAM_COUNTER_ADD(&pCache->StatWritten, cbToWrite);
2341 }
2342 else
2343 {
2344 /* Defer the write and fetch the data from the endpoint. */
2345 pdmBlkCacheEntryWaitersAdd(pEntryNew, pReq,
2346 &SgBuf, offDiff, cbToWrite,
2347 true /* fWrite */);
2348 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2349 pdmBlkCacheEntryReadFromMedium(pEntryNew);
2350 }
2351
2352 pdmBlkCacheEntryRelease(pEntryNew);
2353 }
2354 else
2355 {
2356 /*
2357 * There is not enough free space in the cache.
2358 * Pass the request directly to the I/O manager.
2359 */
2360 LogFlow(("Couldn't evict %u bytes from the cache. Remaining request will be passed through\n", cbToWrite));
2361
2362 STAM_COUNTER_INC(&pCache->cMisses);
2363
2364 pdmBlkCacheRequestPassthrough(pBlkCache, pReq,
2365 &SgBuf, off, cbToWrite,
2366 PDMBLKCACHEXFERDIR_WRITE);
2367 }
2368 }
2369
2370 off += cbToWrite;
2371 }
2372
2373 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, false))
2374 rc = VINF_AIO_TASK_PENDING;
2375
2376 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2377
2378 return rc;
2379}
2380
2381VMMR3DECL(int) PDMR3BlkCacheFlush(PPDMBLKCACHE pBlkCache, void *pvUser)
2382{
2383 int rc = VINF_SUCCESS;
2384 PPDMBLKCACHEREQ pReq;
2385
2386 LogFlowFunc((": pBlkCache=%#p{%s}\n", pBlkCache, pBlkCache->pszId));
2387
2388 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2389 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
2390
2391 /* Commit dirty entries in the cache. */
2392 pdmBlkCacheCommit(pBlkCache);
2393
2394 /* Allocate new request structure. */
2395 pReq = pdmBlkCacheReqAlloc(pvUser);
2396 if (RT_UNLIKELY(!pReq))
2397 return VERR_NO_MEMORY;
2398
2399 rc = pdmBlkCacheRequestPassthrough(pBlkCache, pReq, NULL, 0, 0,
2400 PDMBLKCACHEXFERDIR_FLUSH);
2401 AssertRC(rc);
2402
2403 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2404 return VINF_AIO_TASK_PENDING;
2405}
2406
2407VMMR3DECL(int) PDMR3BlkCacheDiscard(PPDMBLKCACHE pBlkCache, PCRTRANGE paRanges,
2408 unsigned cRanges, void *pvUser)
2409{
2410 int rc = VINF_SUCCESS;
2411 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2412 PPDMBLKCACHEENTRY pEntry;
2413 PPDMBLKCACHEREQ pReq;
2414
2415 LogFlowFunc((": pBlkCache=%#p{%s} paRanges=%#p cRanges=%u pvUser=%#p\n",
2416 pBlkCache, pBlkCache->pszId, paRanges, cRanges, pvUser));
2417
2418 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2419 AssertReturn(!pBlkCache->fSuspended, VERR_INVALID_STATE);
2420
2421 /* Allocate new request structure. */
2422 pReq = pdmBlkCacheReqAlloc(pvUser);
2423 if (RT_UNLIKELY(!pReq))
2424 return VERR_NO_MEMORY;
2425
2426 /* Increment data transfer counter to keep the request valid while we access it. */
2427 ASMAtomicIncU32(&pReq->cXfersPending);
2428
2429 for (unsigned i = 0; i < cRanges; i++)
2430 {
2431 uint64_t offCur = paRanges[i].offStart;
2432 size_t cbLeft = paRanges[i].cbRange;
2433
2434 while (cbLeft)
2435 {
2436 size_t cbThisDiscard = 0;
2437
2438 pEntry = pdmBlkCacheGetCacheEntryByOffset(pBlkCache, offCur);
2439
2440 if (pEntry)
2441 {
2442 /* Write the data into the entry and mark it as dirty */
2443 AssertPtr(pEntry->pList);
2444
2445 uint64_t offDiff = offCur - pEntry->Core.Key;
2446
2447 AssertMsg(offCur >= pEntry->Core.Key,
2448 ("Overflow in calculation offCur=%llu OffsetAligned=%llu\n",
2449 offCur, pEntry->Core.Key));
2450
2451 cbThisDiscard = RT_MIN(pEntry->cbData - offDiff, cbLeft);
2452
2453 /* Ghost lists contain no data. */
2454 if ( (pEntry->pList == &pCache->LruRecentlyUsedIn)
2455 || (pEntry->pList == &pCache->LruFrequentlyUsed))
2456 {
2457 /* Check if the entry is dirty. */
2458 if (pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2459 PDMBLKCACHE_ENTRY_IS_DIRTY,
2460 0))
2461 {
2462 /* If it is dirty but not yet in progress remove it. */
2463 if (!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS))
2464 {
2465 pdmBlkCacheLockEnter(pCache);
2466 pdmBlkCacheEntryRemoveFromList(pEntry);
2467
2468 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2469 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2470 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2471
2472 pdmBlkCacheLockLeave(pCache);
2473
2474 RTMemFree(pEntry);
2475 }
2476 else
2477 {
2478#if 0
2479 /* The data isn't written to the file yet */
2480 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2481 &SgBuf, offDiff, cbToWrite,
2482 true /* fWrite */);
2483 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2484#endif
2485 }
2486
2487 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2488 pdmBlkCacheEntryRelease(pEntry);
2489 }
2490 else /* Dirty bit not set */
2491 {
2492 /*
2493 * Check if a read is in progress for this entry.
2494 * We have to defer processing in that case.
2495 */
2496 if(pdmBlkCacheEntryFlagIsSetClearAcquireLock(pBlkCache, pEntry,
2497 PDMBLKCACHE_ENTRY_IO_IN_PROGRESS,
2498 0))
2499 {
2500#if 0
2501 pdmBlkCacheEntryWaitersAdd(pEntry, pReq,
2502 &SgBuf, offDiff, cbToWrite,
2503 true /* fWrite */);
2504#endif
2505 STAM_COUNTER_INC(&pBlkCache->StatWriteDeferred);
2506 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2507 pdmBlkCacheEntryRelease(pEntry);
2508 }
2509 else /* I/O in progress flag not set */
2510 {
2511 pdmBlkCacheLockEnter(pCache);
2512 pdmBlkCacheEntryRemoveFromList(pEntry);
2513
2514 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2515 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2516 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2517 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2518 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2519
2520 pdmBlkCacheLockLeave(pCache);
2521
2522 RTMemFree(pEntry);
2523 }
2524 } /* Dirty bit not set */
2525 }
2526 else /* Entry is on the ghost list just remove cache entry. */
2527 {
2528 pdmBlkCacheLockEnter(pCache);
2529 pdmBlkCacheEntryRemoveFromList(pEntry);
2530
2531 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2532 STAM_PROFILE_ADV_START(&pCache->StatTreeRemove, Cache);
2533 RTAvlrU64Remove(pBlkCache->pTree, pEntry->Core.Key);
2534 STAM_PROFILE_ADV_STOP(&pCache->StatTreeRemove, Cache);
2535 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2536
2537 pdmBlkCacheLockLeave(pCache);
2538
2539 RTMemFree(pEntry);
2540 }
2541 }
2542 /* else: no entry found. */
2543
2544 offCur += cbThisDiscard;
2545 cbLeft -= cbThisDiscard;
2546 }
2547 }
2548
2549 if (!pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, false))
2550 rc = VINF_AIO_TASK_PENDING;
2551
2552 LogFlowFunc((": Leave rc=%Rrc\n", rc));
2553
2554 return rc;
2555}
2556
2557/**
2558 * Completes a task segment freeing all resources and completes the task handle
2559 * if everything was transferred.
2560 *
2561 * @returns Next task segment handle.
2562 * @param pTaskSeg Task segment to complete.
2563 * @param rc Status code to set.
2564 */
2565static PPDMBLKCACHEWAITER pdmBlkCacheWaiterComplete(PPDMBLKCACHE pBlkCache,
2566 PPDMBLKCACHEWAITER pWaiter,
2567 int rc)
2568{
2569 PPDMBLKCACHEWAITER pNext = pWaiter->pNext;
2570 PPDMBLKCACHEREQ pReq = pWaiter->pReq;
2571
2572 pdmBlkCacheReqUpdate(pBlkCache, pReq, rc, true);
2573
2574 RTMemFree(pWaiter);
2575
2576 return pNext;
2577}
2578
2579static void pdmBlkCacheIoXferCompleteEntry(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2580{
2581 PPDMBLKCACHEENTRY pEntry = hIoXfer->pEntry;
2582 PPDMBLKCACHEGLOBAL pCache = pBlkCache->pCache;
2583
2584 /* Reference the entry now as we are clearing the I/O in progress flag
2585 * which protected the entry till now. */
2586 pdmBlkCacheEntryRef(pEntry);
2587
2588 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2589 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IO_IN_PROGRESS;
2590
2591 /* Process waiting segment list. The data in entry might have changed in-between. */
2592 bool fDirty = false;
2593 PPDMBLKCACHEWAITER pComplete = pEntry->pWaitingHead;
2594 PPDMBLKCACHEWAITER pCurr = pComplete;
2595
2596 AssertMsg((pCurr && pEntry->pWaitingTail) || (!pCurr && !pEntry->pWaitingTail),
2597 ("The list tail was not updated correctly\n"));
2598 pEntry->pWaitingTail = NULL;
2599 pEntry->pWaitingHead = NULL;
2600
2601 if (hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_WRITE)
2602 {
2603 /*
2604 * An error here is difficult to handle as the original request completed already.
2605 * The error is logged for now and the VM is paused.
2606 * If the user continues the entry is written again in the hope
2607 * the user fixed the problem and the next write succeeds.
2608 */
2609 if (RT_FAILURE(rcIoXfer))
2610 {
2611 LogRel(("I/O cache: Error while writing entry at offset %llu (%u bytes) to medium \"%s\" (rc=%Rrc)\n",
2612 pEntry->Core.Key, pEntry->cbData, pBlkCache->pszId, rcIoXfer));
2613
2614 if (!ASMAtomicXchgBool(&pCache->fIoErrorVmSuspended, true))
2615 {
2616 int rc = VMSetRuntimeError(pCache->pVM, VMSETRTERR_FLAGS_SUSPEND | VMSETRTERR_FLAGS_NO_WAIT, "BLKCACHE_IOERR",
2617 N_("The I/O cache encountered an error while updating data in medium \"%s\" (rc=%Rrc). "
2618 "Make sure there is enough free space on the disk and that the disk is working properly. "
2619 "Operation can be resumed afterwards"),
2620 pBlkCache->pszId, rcIoXfer);
2621 AssertRC(rc);
2622 }
2623
2624 /* Mark the entry as dirty again to get it added to the list later on. */
2625 fDirty = true;
2626 }
2627
2628 pEntry->fFlags &= ~PDMBLKCACHE_ENTRY_IS_DIRTY;
2629
2630 while (pCurr)
2631 {
2632 AssertMsg(pCurr->fWrite, ("Completed write entries should never have read tasks attached\n"));
2633
2634 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2635 fDirty = true;
2636 pCurr = pCurr->pNext;
2637 }
2638 }
2639 else
2640 {
2641 AssertMsg(hIoXfer->enmXferDir == PDMBLKCACHEXFERDIR_READ, ("Invalid transfer type\n"));
2642 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IS_DIRTY),
2643 ("Invalid flags set\n"));
2644
2645 while (pCurr)
2646 {
2647 if (pCurr->fWrite)
2648 {
2649 RTSgBufCopyToBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2650 fDirty = true;
2651 }
2652 else
2653 RTSgBufCopyFromBuf(&pCurr->SgBuf, pEntry->pbData + pCurr->offCacheEntry, pCurr->cbTransfer);
2654
2655 pCurr = pCurr->pNext;
2656 }
2657 }
2658
2659 bool fCommit = false;
2660 if (fDirty)
2661 fCommit = pdmBlkCacheAddDirtyEntry(pBlkCache, pEntry);
2662
2663 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2664
2665 /* Dereference so that it isn't protected anymore except we issued anyother write for it. */
2666 pdmBlkCacheEntryRelease(pEntry);
2667
2668 if (fCommit)
2669 pdmBlkCacheCommitDirtyEntries(pCache);
2670
2671 /* Complete waiters now. */
2672 while (pComplete)
2673 pComplete = pdmBlkCacheWaiterComplete(pBlkCache, pComplete, rcIoXfer);
2674}
2675
2676VMMR3DECL(void) PDMR3BlkCacheIoXferComplete(PPDMBLKCACHE pBlkCache, PPDMBLKCACHEIOXFER hIoXfer, int rcIoXfer)
2677{
2678 LogFlowFunc(("pBlkCache=%#p hIoXfer=%#p rcIoXfer=%Rrc\n", pBlkCache, hIoXfer, rcIoXfer));
2679
2680 if (hIoXfer->fIoCache)
2681 pdmBlkCacheIoXferCompleteEntry(pBlkCache, hIoXfer, rcIoXfer);
2682 else
2683 pdmBlkCacheReqUpdate(pBlkCache, hIoXfer->pReq, rcIoXfer, true);
2684 RTMemFree(hIoXfer);
2685}
2686
2687/**
2688 * Callback for the AVL do with all routine. Waits for a cachen entry to finish any pending I/O.
2689 *
2690 * @returns IPRT status code.
2691 * @param pNode The node to destroy.
2692 * @param pvUser Opaque user data.
2693 */
2694static int pdmBlkCacheEntryQuiesce(PAVLRU64NODECORE pNode, void *pvUser)
2695{
2696 PPDMBLKCACHEENTRY pEntry = (PPDMBLKCACHEENTRY)pNode;
2697 PPDMBLKCACHE pBlkCache = pEntry->pBlkCache;
2698
2699 while (ASMAtomicReadU32(&pEntry->fFlags) & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS)
2700 {
2701 /* Leave the locks to let the I/O thread make progress but reference the entry to prevent eviction. */
2702 pdmBlkCacheEntryRef(pEntry);
2703 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2704
2705 RTThreadSleep(1);
2706
2707 /* Re-enter all locks and drop the reference. */
2708 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2709 pdmBlkCacheEntryRelease(pEntry);
2710 }
2711
2712 AssertMsg(!(pEntry->fFlags & PDMBLKCACHE_ENTRY_IO_IN_PROGRESS),
2713 ("Entry is dirty and/or still in progress fFlags=%#x\n", pEntry->fFlags));
2714
2715 return VINF_SUCCESS;
2716}
2717
2718VMMR3DECL(int) PDMR3BlkCacheSuspend(PPDMBLKCACHE pBlkCache)
2719{
2720 int rc = VINF_SUCCESS;
2721 LogFlowFunc(("pBlkCache=%#p\n", pBlkCache));
2722
2723 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2724
2725 if (!ASMAtomicReadBool(&pBlkCache->pCache->fIoErrorVmSuspended))
2726 pdmBlkCacheCommit(pBlkCache); /* Can issue new I/O requests. */
2727 ASMAtomicXchgBool(&pBlkCache->fSuspended, true);
2728
2729 /* Wait for all I/O to complete. */
2730 RTSemRWRequestWrite(pBlkCache->SemRWEntries, RT_INDEFINITE_WAIT);
2731 rc = RTAvlrU64DoWithAll(pBlkCache->pTree, true, pdmBlkCacheEntryQuiesce, NULL);
2732 AssertRC(rc);
2733 RTSemRWReleaseWrite(pBlkCache->SemRWEntries);
2734
2735 return rc;
2736}
2737
2738VMMR3DECL(int) PDMR3BlkCacheResume(PPDMBLKCACHE pBlkCache)
2739{
2740 LogFlowFunc(("pBlkCache=%#p\n", pBlkCache));
2741
2742 AssertPtrReturn(pBlkCache, VERR_INVALID_POINTER);
2743
2744 ASMAtomicXchgBool(&pBlkCache->fSuspended, false);
2745
2746 return VINF_SUCCESS;
2747}
2748
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