Index: /trunk/src/VBox/VMM/PDMAsyncCompletionFile.cpp =================================================================== --- /trunk/src/VBox/VMM/PDMAsyncCompletionFile.cpp (revision 23955) +++ /trunk/src/VBox/VMM/PDMAsyncCompletionFile.cpp (revision 23956) @@ -25,4 +25,5 @@ *******************************************************************************/ #define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION +#define RT_STRICT #include "PDMInternal.h" #include @@ -381,6 +382,7 @@ * @param pEpClass Pointer to the endpoint class data. * @param ppAioMgr Where to store the pointer to the new async I/O manager on success. + * @param fFailsafe Flag to force a failsafe manager even if the global flag is not set. */ -int pdmacFileAioMgrCreate(PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass, PPPDMACEPFILEMGR ppAioMgr) +int pdmacFileAioMgrCreate(PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass, PPPDMACEPFILEMGR ppAioMgr, bool fFailsafe) { int rc = VINF_SUCCESS; @@ -392,5 +394,5 @@ if (RT_SUCCESS(rc)) { - pAioMgrNew->fFailsafe = pEpClass->fFailsafe; + pAioMgrNew->fFailsafe = fFailsafe ? true : pEpClass->fFailsafe; rc = RTSemEventCreate(&pAioMgrNew->EventSem); @@ -637,5 +639,5 @@ { /* Safe mode. Every file has its own async I/O manager. */ - rc = pdmacFileAioMgrCreate(pEpClassFile, &pAioMgr); + rc = pdmacFileAioMgrCreate(pEpClassFile, &pAioMgr, false); AssertRC(rc); } @@ -656,5 +658,5 @@ if (!pEpClassFile->pAioMgrHead) { - rc = pdmacFileAioMgrCreate(pEpClassFile, &pAioMgr); + rc = pdmacFileAioMgrCreate(pEpClassFile, &pAioMgr, false); AssertRC(rc); } Index: /trunk/src/VBox/VMM/PDMAsyncCompletionFileFailsafe.cpp =================================================================== --- /trunk/src/VBox/VMM/PDMAsyncCompletionFileFailsafe.cpp (revision 23955) +++ /trunk/src/VBox/VMM/PDMAsyncCompletionFileFailsafe.cpp (revision 23956) @@ -21,13 +21,15 @@ */ #define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION +#define RT_STRICT #include +#include #include #include "PDMAsyncCompletionFileInternal.h" -static int pdmacFileAioMgrFailsafeProcessEndpoint(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint) +static int pdmacFileAioMgrFailsafeProcessEndpointTaskList(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, + PPDMACTASKFILE pTasks) { int rc = VINF_SUCCESS; - PPDMACTASKFILE pTasks = pdmacFileEpGetNewTasks(pEndpoint); while (pTasks) @@ -83,4 +85,27 @@ } +static int pdmacFileAioMgrFailsafeProcessEndpoint(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint) +{ + int rc = VINF_SUCCESS; + PPDMACTASKFILE pTasks = pEndpoint->AioMgr.pReqsPendingHead; + + pEndpoint->AioMgr.pReqsPendingHead = NULL; + pEndpoint->AioMgr.pReqsPendingTail = NULL; + + /* Process the request pending list first in case the endpoint was migrated due to an error. */ + if (pTasks) + rc = pdmacFileAioMgrFailsafeProcessEndpointTaskList(pEndpoint, pTasks); + + if (RT_SUCCESS(rc)) + { + pTasks = pdmacFileEpGetNewTasks(pEndpoint); + + if (pTasks); + rc = pdmacFileAioMgrFailsafeProcessEndpointTaskList(pEndpoint, pTasks); + } + + return rc; +} + /** * A fallback method in case something goes wrong with the normal Index: /trunk/src/VBox/VMM/PDMAsyncCompletionFileInternal.h =================================================================== --- /trunk/src/VBox/VMM/PDMAsyncCompletionFileInternal.h (revision 23955) +++ /trunk/src/VBox/VMM/PDMAsyncCompletionFileInternal.h (revision 23956) @@ -528,5 +528,5 @@ void pdmacFileAioMgrNormalDestroy(PPDMACEPFILEMGR pAioMgr); -int pdmacFileAioMgrCreate(PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass, PPPDMACEPFILEMGR ppAioMgr); +int pdmacFileAioMgrCreate(PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass, PPPDMACEPFILEMGR ppAioMgr, bool fFailsafe); int pdmacFileAioMgrAddEndpoint(PPDMACEPFILEMGR pAioMgr, PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint); Index: /trunk/src/VBox/VMM/PDMAsyncCompletionFileNormal.cpp =================================================================== --- /trunk/src/VBox/VMM/PDMAsyncCompletionFileNormal.cpp (revision 23955) +++ /trunk/src/VBox/VMM/PDMAsyncCompletionFileNormal.cpp (revision 23956) @@ -21,4 +21,5 @@ */ #define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION +#define RT_STRICT #include #include @@ -26,4 +27,5 @@ #include #include +#include #include @@ -206,5 +208,5 @@ { rc = pdmacFileAioMgrCreate((PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass, - &pAioMgrNew); + &pAioMgrNew, false); if (RT_SUCCESS(rc)) { @@ -850,78 +852,136 @@ pEndpoint = pTask->pEndpoint; - AssertMsg( RT_SUCCESS(rcReq) - && ( (cbTransfered == pTask->DataSeg.cbSeg) - || (pTask->fBounceBuffer)), - ("Task didn't completed successfully (rc=%Rrc) or was incomplete (cbTransfered=%u)\n", rcReq, cbTransfered)); - - if (pTask->fPrefetch) + /* + * It is possible that the request failed on Linux with kernels < 2.6.23 + * if the passed buffer was allocated with remap_pfn_range or if the file + * is on an NFS endpoint which does not support async and direct I/O at the same time. + * The endpoint will be migrated to a failsafe manager in case a request fails. + */ + if (RT_FAILURE(rcReq)) { - Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE); - Assert(pTask->fBounceBuffer); - - memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset, - pTask->DataSeg.pvSeg, - pTask->DataSeg.cbSeg); - - /* Write it now. */ - pTask->fPrefetch = false; - size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg, 512); - RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1); - - /* Grow the file if needed. */ - if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile)) + /* Free bounce buffers and the IPRT request. */ + pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i]; + pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries; + + pAioMgr->cRequestsActive--; + pEndpoint->AioMgr.cRequestsActive--; + pEndpoint->AioMgr.cReqsProcessed++; + + if (pTask->fBounceBuffer) + RTMemFree(pTask->pvBounceBuffer); + + /* Queue the request on the pending list. */ + pTask->pNext = pEndpoint->AioMgr.pReqsPendingHead; + pEndpoint->AioMgr.pReqsPendingHead = pTask; + + /* Create a new failsafe manager if neccessary. */ + if (!pEndpoint->AioMgr.fMoving) { - ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg); - RTFileSetSize(pEndpoint->File, pTask->Off + pTask->DataSeg.cbSeg); + PPDMACEPFILEMGR pAioMgrFailsafe; + + pEndpoint->AioMgr.fMoving = true; + + rc = pdmacFileAioMgrCreate((PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass, + &pAioMgrFailsafe, true); + AssertRC(rc); + + pEndpoint->AioMgr.pAioMgrDst = pAioMgrFailsafe; } - rc = RTFileAioReqPrepareWrite(apReqs[i], pEndpoint->File, - offStart, pTask->pvBounceBuffer, cbToTransfer, pTask); - AssertRC(rc); - rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, &apReqs[i], 1); - AssertRC(rc); + /* If this was the last request for the endpoint migrate it to the new manager. */ + if (!pEndpoint->AioMgr.cRequestsActive) + { + bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pEndpoint); + Assert(!fReqsPending); + + rc = pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint); + AssertRC(rc); + } } else { - if (pTask->fBounceBuffer) + AssertMsg(( (cbTransfered == pTask->DataSeg.cbSeg) + || (pTask->fBounceBuffer && (cbTransfered >= pTask->DataSeg.cbSeg))), + ("Task didn't completed successfully (rc=%Rrc) or was incomplete (cbTransfered=%u)\n", rcReq, cbTransfered)); + + if (pTask->fPrefetch) { - if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ) - memcpy(pTask->DataSeg.pvSeg, - ((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset, - pTask->DataSeg.cbSeg); - - RTMemPageFree(pTask->pvBounceBuffer); + Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE); + Assert(pTask->fBounceBuffer); + + memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset, + pTask->DataSeg.pvSeg, + pTask->DataSeg.cbSeg); + + /* Write it now. */ + pTask->fPrefetch = false; + size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg, 512); + RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1); + + /* Grow the file if needed. */ + if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile)) + { + ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg); + RTFileSetSize(pEndpoint->File, pTask->Off + pTask->DataSeg.cbSeg); + } + + rc = RTFileAioReqPrepareWrite(apReqs[i], pEndpoint->File, + offStart, pTask->pvBounceBuffer, cbToTransfer, pTask); + AssertRC(rc); + rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, &apReqs[i], 1); + AssertRC(rc); } - - /* Put the entry on the free array */ - pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i]; - pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries; - - pAioMgr->cRequestsActive--; - pEndpoint->AioMgr.cRequestsActive--; - pEndpoint->AioMgr.cReqsProcessed++; - - /* Call completion callback */ - pTask->pfnCompleted(pTask, pTask->pvUser); - pdmacFileTaskFree(pEndpoint, pTask); - - /* - * If there is no request left on the endpoint but a flush request is set - * it completed now and we notify the owner. - * Furthermore we look for new requests and continue. - */ - if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq) + else { + if (pTask->fBounceBuffer) + { + if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ) + memcpy(pTask->DataSeg.pvSeg, + ((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset, + pTask->DataSeg.cbSeg); + + RTMemPageFree(pTask->pvBounceBuffer); + } + + /* Put the entry on the free array */ + pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i]; + pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries; + + pAioMgr->cRequestsActive--; + pEndpoint->AioMgr.cRequestsActive--; + pEndpoint->AioMgr.cReqsProcessed++; + /* Call completion callback */ - pTask = pEndpoint->pFlushReq; - pEndpoint->pFlushReq = NULL; - - AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n")); - pTask->pfnCompleted(pTask, pTask->pvUser); pdmacFileTaskFree(pEndpoint, pTask); + + /* + * If there is no request left on the endpoint but a flush request is set + * it completed now and we notify the owner. + * Furthermore we look for new requests and continue. + */ + if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq) + { + /* Call completion callback */ + pTask = pEndpoint->pFlushReq; + pEndpoint->pFlushReq = NULL; + + AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n")); + + pTask->pfnCompleted(pTask, pTask->pvUser); + pdmacFileTaskFree(pEndpoint, pTask); + } + else if (RT_UNLIKELY(!pEndpoint->AioMgr.cRequestsActive && pEndpoint->AioMgr.fMoving)) + { + /* If we the endpoint is about to be migrated do it now. */ + bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pEndpoint); + Assert(!fReqsPending); + + rc = pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint); + AssertRC(rc); + } } - } - } + } /* request completed successfully */ + } /* for every completed request */ /* Check for an external blocking event before we go to sleep again. */