PDMAsyncCompletionFileNormal.cpp@ 23911

Last change on this file since 23911 was 23812, checked in by vboxsync, 15 years ago
PDMAsyncCompletion: Bugfixes
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 36.8 KB

Line
1	/* $Id: PDMAsyncCompletionFileNormal.cpp 23812 2009-10-15 17:52:15Z vboxsync $ */
2	/** @file
3	* PDM Async I/O - Transport data asynchronous in R3 using EMT.
4	* Async File I/O manager.
5	*/
6
7	/*
8	* Copyright (C) 2006-2008 Sun Microsystems, Inc.
9	*
10	* This file is part of VirtualBox Open Source Edition (OSE), as
11	* available from http://www.virtualbox.org. This file is free software;
12	* you can redistribute it and/or modify it under the terms of the GNU
13	* General Public License (GPL) as published by the Free Software
14	* Foundation, in version 2 as it comes in the "COPYING" file of the
15	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
16	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
17	*
18	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
19	* Clara, CA 95054 USA or visit http://www.sun.com if you need
20	* additional information or have any questions.
21	*/
22	#define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION
23	#include <iprt/types.h>
24	#include <iprt/asm.h>
25	#include <iprt/file.h>
26	#include <iprt/mem.h>
27	#include <iprt/string.h>
28	#include <VBox/log.h>
29
30	#include "PDMAsyncCompletionFileInternal.h"
31
32	/** The update period for the I/O load statistics in ms. */
33	#define PDMACEPFILEMGR_LOAD_UPDATE_PERIOD 1000
34	/** Maximum number of requests a manager will handle. */
35	#define PDMACEPFILEMGR_REQS_MAX 512 /* @todo: Find better solution wrt. the request number*/
36
37	int pdmacFileAioMgrNormalInit(PPDMACEPFILEMGR pAioMgr)
38	{
39	int rc = VINF_SUCCESS;
40
41	rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, RTFILEAIO_UNLIMITED_REQS);
42	if (rc == VERR_OUT_OF_RANGE)
43	rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, PDMACEPFILEMGR_REQS_MAX);
44
45	if (RT_SUCCESS(rc))
46	{
47	/* Initialize request handle array. */
48	pAioMgr->iFreeEntryNext = 0;
49	pAioMgr->iFreeReqNext = 0;
50	pAioMgr->cReqEntries = PDMACEPFILEMGR_REQS_MAX + 1;
51	pAioMgr->pahReqsFree = (RTFILEAIOREQ )RTMemAllocZ(pAioMgr->cReqEntries sizeof(RTFILEAIOREQ));
52
53	if (pAioMgr->pahReqsFree)
54	{
55	return VINF_SUCCESS;
56	}
57	else
58	{
59	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
60	rc = VERR_NO_MEMORY;
61	}
62	}
63
64	return rc;
65	}
66
67	void pdmacFileAioMgrNormalDestroy(PPDMACEPFILEMGR pAioMgr)
68	{
69	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
70
71	while (pAioMgr->iFreeReqNext != pAioMgr->iFreeEntryNext)
72	{
73	RTFileAioReqDestroy(pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext]);
74	pAioMgr->iFreeReqNext = (pAioMgr->iFreeReqNext + 1) % pAioMgr->cReqEntries;
75	}
76
77	RTMemFree(pAioMgr->pahReqsFree);
78	}
79
80	/**
81	* Sorts the endpoint list with insertion sort.
82	*/
83	static void pdmacFileAioMgrNormalEndpointsSortByLoad(PPDMACEPFILEMGR pAioMgr)
84	{
85	PPDMASYNCCOMPLETIONENDPOINTFILE pEpPrev, pEpCurr, pEpNextToSort;
86
87	pEpPrev = pAioMgr->pEndpointsHead;
88	pEpCurr = pEpPrev->AioMgr.pEndpointNext;
89
90	while (pEpCurr)
91	{
92	/* Remember the next element to sort because the list might change. */
93	pEpNextToSort = pEpCurr->AioMgr.pEndpointNext;
94
95	/* Unlink the current element from the list. */
96	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
97	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
98
99	if (pPrev)
100	pPrev->AioMgr.pEndpointNext = pNext;
101	else
102	pAioMgr->pEndpointsHead = pNext;
103
104	if (pNext)
105	pNext->AioMgr.pEndpointPrev = pPrev;
106
107	/* Go back until we reached the place to insert the current endpoint into. */
108	while (pEpPrev && (pEpPrev->AioMgr.cReqsPerSec < pEpCurr->AioMgr.cReqsPerSec))
109	pEpPrev = pEpPrev->AioMgr.pEndpointPrev;
110
111	/* Link the endpoint into the list. */
112	if (pEpPrev)
113	pNext = pEpPrev->AioMgr.pEndpointNext;
114	else
115	pNext = pAioMgr->pEndpointsHead;
116
117	pEpCurr->AioMgr.pEndpointNext = pNext;
118	pEpCurr->AioMgr.pEndpointPrev = pEpPrev;
119	pNext->AioMgr.pEndpointPrev = pEpCurr;
120	if (pEpPrev)
121	pEpPrev->AioMgr.pEndpointNext = pEpCurr;
122	else
123	pAioMgr->pEndpointsHead = pEpCurr;
124
125	pEpCurr = pEpNextToSort;
126	}
127
128	#ifdef DEBUG
129	/* Validate sorting alogrithm */
130	unsigned cEndpoints = 0;
131	pEpCurr = pAioMgr->pEndpointsHead;
132
133	AssertMsg(pEpCurr, ("No endpoint in the list?\n"));
134	AssertMsg(!pEpCurr->AioMgr.pEndpointPrev, ("First element in the list points to previous element\n"));
135
136	while (pEpCurr)
137	{
138	cEndpoints++;
139
140	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
141	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
142
143	Assert(!pNext \|\| pNext->AioMgr.cReqsPerSec <= pEpCurr->AioMgr.cReqsPerSec);
144	Assert(!pPrev \|\| pPrev->AioMgr.cReqsPerSec >= pEpCurr->AioMgr.cReqsPerSec);
145
146	pEpCurr = pNext;
147	}
148
149	AssertMsg(cEndpoints == pAioMgr->cEndpoints, ("Endpoints lost during sort!\n"));
150
151	#endif
152	}
153
154	/**
155	* Removes an endpoint from the currently assigned manager.
156	*
157	* @returns TRUE if there are still requests pending on the current manager for this endpoint.
158	* FALSE otherwise.
159	* @param pEndpointRemove The endpoint to remove.
160	*/
161	static bool pdmacFileAioMgrNormalRemoveEndpoint(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove)
162	{
163	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEndpointRemove->AioMgr.pEndpointPrev;
164	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEndpointRemove->AioMgr.pEndpointNext;
165	PPDMACEPFILEMGR pAioMgr = pEndpointRemove->pAioMgr;
166
167	pAioMgr->cEndpoints--;
168
169	if (pPrev)
170	pPrev->AioMgr.pEndpointNext = pNext;
171	else
172	pAioMgr->pEndpointsHead = pNext;
173
174	if (pNext)
175	pNext->AioMgr.pEndpointPrev = pPrev;
176
177	/* Make sure that there is no request pending on this manager for the endpoint. */
178	if (!pEndpointRemove->AioMgr.cRequestsActive)
179	{
180	Assert(!pEndpointRemove->pFlushReq);
181
182	/* Reopen the file so that the new endpoint can reassociate with the file */
183	RTFileClose(pEndpointRemove->File);
184	int rc = RTFileOpen(&pEndpointRemove->File, pEndpointRemove->Core.pszUri, pEndpointRemove->fFlags);
185	AssertRC(rc);
186	return false;
187	}
188
189	return true;
190	}
191
192	/**
193	* Creates a new I/O manager and spreads the I/O load of the endpoints
194	* between the given I/O manager and the new one.
195	*
196	* @returns nothing.
197	* @param pAioMgr The I/O manager with high I/O load.
198	*/
199	static void pdmacFileAioMgrNormalBalanceLoad(PPDMACEPFILEMGR pAioMgr)
200	{
201	PPDMACEPFILEMGR pAioMgrNew = NULL;
202	int rc = VINF_SUCCESS;
203
204	/* Splitting can't be done with only one open endpoint. */
205	if (pAioMgr->cEndpoints > 1)
206	{
207	rc = pdmacFileAioMgrCreate((PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass,
208	&pAioMgrNew);
209	if (RT_SUCCESS(rc))
210	{
211	/* We will sort the list by request count per second. */
212	pdmacFileAioMgrNormalEndpointsSortByLoad(pAioMgr);
213
214	/* Now move some endpoints to the new manager. */
215	unsigned cReqsHere = pAioMgr->pEndpointsHead->AioMgr.cReqsPerSec;
216	unsigned cReqsOther = 0;
217	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead->AioMgr.pEndpointNext;
218
219	while (pCurr)
220	{
221	if (cReqsHere <= cReqsOther)
222	{
223	/*
224	* The other manager has more requests to handle now.
225	* We will keep the current endpoint.
226	*/
227	Log(("Keeping endpoint %#p{%s} with %u reqs/s\n", pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
228	cReqsHere += pCurr->AioMgr.cReqsPerSec;
229	pCurr = pCurr->AioMgr.pEndpointNext;
230	}
231	else
232	{
233	/* Move to other endpoint. */
234	Log(("Moving endpoint %#p{%s} with %u reqs/s to other manager\n", pCurr, pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
235	cReqsOther += pCurr->AioMgr.cReqsPerSec;
236
237	PPDMASYNCCOMPLETIONENDPOINTFILE pMove = pCurr;
238
239	pCurr = pCurr->AioMgr.pEndpointNext;
240
241	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pMove);
242
243	if (fReqsPending)
244	{
245	pMove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
246	pMove->AioMgr.fMoving = true;
247	pMove->AioMgr.pAioMgrDst = pAioMgrNew;
248	}
249	else
250	{
251	pMove->AioMgr.fMoving = false;
252	pMove->AioMgr.pAioMgrDst = NULL;
253	pdmacFileAioMgrAddEndpoint(pAioMgrNew, pMove);
254	}
255	}
256	}
257	}
258	else
259	{
260	/* Don't process further but leave a log entry about reduced performance. */
261	LogRel(("AIOMgr: Could not create new I/O manager (rc=%Rrc). Expect reduced performance\n", rc));
262	}
263	}
264	}
265
266	/**
267	* Error handler which will create the failsafe managers and destroy the failed I/O manager.
268	*
269	* @returns VBox status code
270	* @param pAioMgr The I/O manager the error ocurred on.
271	* @param rc The error code.
272	*/
273	static int pdmacFileAioMgrNormalErrorHandler(PPDMACEPFILEMGR pAioMgr, int rc, RT_SRC_POS_DECL)
274	{
275	LogRel(("AIOMgr: I/O manager %#p encountered a critical error (rc=%Rrc) during operation. Falling back to failsafe mode. Expect reduced performance\n",
276	pAioMgr, rc));
277	LogRel(("AIOMgr: Error happened in %s:(%u){%s}\n", RT_SRC_POS_ARGS));
278	LogRel(("AIOMgr: Please contact the product vendor\n"));
279
280	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass;
281
282	pAioMgr->enmState = PDMACEPFILEMGRSTATE_FAULT;
283	ASMAtomicWriteBool(&pEpClassFile->fFailsafe, true);
284
285	AssertMsgFailed(("Implement\n"));
286	return VINF_SUCCESS;
287	}
288
289	/**
290	* Put a list of tasks in the pending request list of an endpoint.
291	*/
292	DECLINLINE(void) pdmacFileAioMgrEpAddTaskList(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTaskHead)
293	{
294	/* Add the rest of the tasks to the pending list */
295	if (!pEndpoint->AioMgr.pReqsPendingHead)
296	{
297	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
298	pEndpoint->AioMgr.pReqsPendingHead = pTaskHead;
299	}
300	else
301	{
302	Assert(pEndpoint->AioMgr.pReqsPendingTail);
303	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTaskHead;
304	}
305
306	/* Update the tail. */
307	while (pTaskHead->pNext)
308	pTaskHead = pTaskHead->pNext;
309
310	pEndpoint->AioMgr.pReqsPendingTail = pTaskHead;
311	}
312
313	/**
314	* Put one task in the pending request list of an endpoint.
315	*/
316	DECLINLINE(void) pdmacFileAioMgrEpAddTask(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTask)
317	{
318	/* Add the rest of the tasks to the pending list */
319	if (!pEndpoint->AioMgr.pReqsPendingHead)
320	{
321	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
322	pEndpoint->AioMgr.pReqsPendingHead = pTask;
323	}
324	else
325	{
326	Assert(pEndpoint->AioMgr.pReqsPendingTail);
327	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTask;
328	}
329
330	pEndpoint->AioMgr.pReqsPendingTail = pTask;
331	}
332
333	/**
334	* Wrapper around RTFIleAioCtxSubmit() which is also doing error handling.
335	*/
336	static int pdmacFileAioMgrNormalReqsEnqueue(PPDMACEPFILEMGR pAioMgr,
337	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
338	PRTFILEAIOREQ pahReqs, size_t cReqs)
339	{
340	int rc;
341
342	pAioMgr->cRequestsActive += cReqs;
343	pEndpoint->AioMgr.cRequestsActive += cReqs;
344
345	LogFlow(("Enqueuing %d requests. I/O manager has a total of %d active requests now\n", cReqs, pAioMgr->cRequestsActive));
346	LogFlow(("Endpoint has a total of %d active requests now\n", pEndpoint->AioMgr.cRequestsActive));
347
348	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, pahReqs, cReqs);
349	if (RT_FAILURE(rc))
350	{
351	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
352	{
353	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
354
355	/*
356	* We run out of resources.
357	* Need to check which requests got queued
358	* and put the rest on the pending list again.
359	*/
360	if (RT_UNLIKELY(!pEpClass->fOutOfResourcesWarningPrinted))
361	{
362	pEpClass->fOutOfResourcesWarningPrinted = true;
363	LogRel(("AIOMgr: The operating system doesn't have enough resources "
364	"to handle the I/O load of the VM. Expect reduced I/O performance\n"));
365	}
366
367	for (size_t i = 0; i < cReqs; i++)
368	{
369	int rcReq = RTFileAioReqGetRC(pahReqs[i], NULL);
370
371	if (rcReq != VERR_FILE_AIO_IN_PROGRESS)
372	{
373	AssertMsg(rcReq == VERR_FILE_AIO_NOT_SUBMITTED,
374	("Request returned unexpected return code: rc=%Rrc\n", rcReq));
375
376	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(pahReqs[i]);
377
378	/* Put the entry on the free array */
379	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = pahReqs[i];
380	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries;
381
382	pdmacFileAioMgrEpAddTask(pEndpoint, pTask);
383	pAioMgr->cRequestsActive--;
384	pEndpoint->AioMgr.cRequestsActive--;
385	}
386	}
387	LogFlow(("Removed requests. I/O manager has a total of %d active requests now\n", pAioMgr->cRequestsActive));
388	LogFlow(("Endpoint has a total of %d active requests now\n", pEndpoint->AioMgr.cRequestsActive));
389	}
390	else
391	AssertMsgFailed(("Unexpected return code rc=%Rrc\n", rc));
392	}
393
394	return rc;
395	}
396
397	static int pdmacFileAioMgrNormalProcessTaskList(PPDMACTASKFILE pTaskHead,
398	PPDMACEPFILEMGR pAioMgr,
399	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
400	{
401	RTFILEAIOREQ apReqs[20];
402	unsigned cRequests = 0;
403	unsigned cMaxRequests = PDMACEPFILEMGR_REQS_MAX - pAioMgr->cRequestsActive;
404	int rc = VINF_SUCCESS;
405	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
406
407	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
408	("Trying to process request lists of a non active endpoint!\n"));
409
410	/* Go through the list and queue the requests until we get a flush request */
411	while ( pTaskHead
412	&& !pEndpoint->pFlushReq
413	&& (cMaxRequests > 0)
414	&& RT_SUCCESS(rc))
415	{
416	PPDMACTASKFILE pCurr = pTaskHead;
417
418	pTaskHead = pTaskHead->pNext;
419
420	pCurr->pNext = NULL;
421
422	AssertMsg(VALID_PTR(pCurr->pEndpoint) && (pCurr->pEndpoint == pEndpoint),
423	("Endpoints do not match\n"));
424
425	switch (pCurr->enmTransferType)
426	{
427	case PDMACTASKFILETRANSFER_FLUSH:
428	{
429	/* If there is no data transfer request this flush request finished immediately. */
430	if (!pEndpoint->AioMgr.cRequestsActive)
431	{
432	pCurr->pfnCompleted(pCurr, pCurr->pvUser);
433	pdmacFileTaskFree(pEndpoint, pCurr);
434	}
435	else
436	{
437	pEndpoint->pFlushReq = pCurr;
438	}
439	break;
440	}
441	case PDMACTASKFILETRANSFER_READ:
442	case PDMACTASKFILETRANSFER_WRITE:
443	{
444	RTFILEAIOREQ hReq = NIL_RTFILEAIOREQ;
445	void *pvBuf = pCurr->DataSeg.pvSeg;
446
447	/* Get a request handle. */
448	if (pAioMgr->iFreeReqNext != pAioMgr->iFreeEntryNext)
449	{
450	hReq = pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext];
451	pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext] = NIL_RTFILEAIOREQ;
452	pAioMgr->iFreeReqNext = (pAioMgr->iFreeReqNext + 1) % pAioMgr->cReqEntries;
453	}
454	else
455	{
456	rc = RTFileAioReqCreate(&hReq);
457	AssertRC(rc);
458	}
459
460	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
461
462	/* Check if the alignment requirements are met.
463	* Offset, transfer size and buffer address
464	* need to be on a 512 boundary. */
465	RTFOFF offStart = pCurr->Off & ~(RTFOFF)(512-1);
466	size_t cbToTransfer = RT_ALIGN_Z(pCurr->DataSeg.cbSeg + (pCurr->Off - offStart), 512);
467	PDMACTASKFILETRANSFER enmTransferType = pCurr->enmTransferType;
468
469	AssertMsg( pCurr->enmTransferType == PDMACTASKFILETRANSFER_WRITE
470	\|\| (uint64_t)(offStart + cbToTransfer) <= pEndpoint->cbFile,
471	("Read exceeds file size offStart=%RTfoff cbToTransfer=%d cbFile=%llu\n",
472	offStart, cbToTransfer, pEndpoint->cbFile));
473
474	pCurr->fPrefetch = false;
475
476	if ( RT_UNLIKELY(cbToTransfer != pCurr->DataSeg.cbSeg)
477	\|\| RT_UNLIKELY(offStart != pCurr->Off)
478	\|\| ((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) != (RTR3UINTPTR)pvBuf))
479	{
480	LogFlow(("Using bounce buffer for task %#p cbToTransfer=%zd cbSeg=%zd offStart=%RTfoff off=%RTfoff\n",
481	pCurr, cbToTransfer, pCurr->DataSeg.cbSeg, offStart, pCurr->Off));
482
483	/* Create bounce buffer. */
484	pCurr->fBounceBuffer = true;
485
486	AssertMsg(pCurr->Off >= offStart, ("Overflow in calculation Off=%llu offStart=%llu\n",
487	pCurr->Off, offStart));
488	pCurr->uBounceBufOffset = pCurr->Off - offStart;
489
490	/** @todo: I think we need something like a RTMemAllocAligned method here.
491	* Current assumption is that the maximum alignment is 4096byte
492	* (GPT disk on Windows)
493	* so we can use RTMemPageAlloc here.
494	*/
495	pCurr->pvBounceBuffer = RTMemPageAlloc(cbToTransfer);
496	AssertPtr(pCurr->pvBounceBuffer);
497	pvBuf = pCurr->pvBounceBuffer;
498
499	if (pCurr->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
500	{
501	if ( RT_UNLIKELY(cbToTransfer != pCurr->DataSeg.cbSeg)
502	\|\| RT_UNLIKELY(offStart != pCurr->Off))
503	{
504	/* We have to fill the buffer first before we can update the data. */
505	LogFlow(("Prefetching data for task %#p\n", pCurr));
506	pCurr->fPrefetch = true;
507	enmTransferType = PDMACTASKFILETRANSFER_READ;
508	}
509	else
510	memcpy(pvBuf, pCurr->DataSeg.pvSeg, pCurr->DataSeg.cbSeg);
511	}
512	}
513	else
514	pCurr->fBounceBuffer = false;
515
516	AssertMsg((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) == (RTR3UINTPTR)pvBuf,
517	("AIO: Alignment restrictions not met! pvBuf=%p uBitmaskAlignment=%p\n", pvBuf, pEpClassFile->uBitmaskAlignment));
518
519	if (enmTransferType == PDMACTASKFILETRANSFER_WRITE)
520	{
521	/* Grow the file if needed. */
522	if (RT_UNLIKELY((uint64_t)(pCurr->Off + pCurr->DataSeg.cbSeg) > pEndpoint->cbFile))
523	{
524	ASMAtomicWriteU64(&pEndpoint->cbFile, pCurr->Off + pCurr->DataSeg.cbSeg);
525	RTFileSetSize(pEndpoint->File, pCurr->Off + pCurr->DataSeg.cbSeg);
526	}
527
528	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->File,
529	offStart, pvBuf, cbToTransfer, pCurr);
530	}
531	else
532	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->File,
533	offStart, pvBuf, cbToTransfer, pCurr);
534	AssertRC(rc);
535
536	apReqs[cRequests] = hReq;
537	pEndpoint->AioMgr.cReqsProcessed++;
538	cMaxRequests--;
539	cRequests++;
540	if (cRequests == RT_ELEMENTS(apReqs))
541	{
542	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
543	cRequests = 0;
544	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
545	("Unexpected return code\n"));
546	}
547	break;
548	}
549	default:
550	AssertMsgFailed(("Invalid transfer type %d\n", pCurr->enmTransferType));
551	}
552	}
553
554	if (cRequests)
555	{
556	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
557	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
558	("Unexpected return code rc=%Rrc\n", rc));
559	}
560
561	if (pTaskHead)
562	{
563	/* Add the rest of the tasks to the pending list */
564	pdmacFileAioMgrEpAddTaskList(pEndpoint, pTaskHead);
565
566	if (RT_UNLIKELY(!cMaxRequests && !pEndpoint->pFlushReq))
567	{
568	/*
569	* The I/O manager has no room left for more requests
570	* but there are still requests to process.
571	* Create a new I/O manager and let it handle some endpoints.
572	*/
573	pdmacFileAioMgrNormalBalanceLoad(pAioMgr);
574	}
575	}
576
577	/* Insufficient resources are not fatal. */
578	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
579	rc = VINF_SUCCESS;
580
581	return rc;
582	}
583
584	/**
585	* Adds all pending requests for the given endpoint
586	* until a flush request is encountered or there is no
587	* request anymore.
588	*
589	* @returns VBox status code.
590	* @param pAioMgr The async I/O manager for the endpoint
591	* @param pEndpoint The endpoint to get the requests from.
592	*/
593	static int pdmacFileAioMgrNormalQueueReqs(PPDMACEPFILEMGR pAioMgr,
594	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
595	{
596	int rc = VINF_SUCCESS;
597	PPDMACTASKFILE pTasksHead = NULL;
598
599	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
600	("Trying to process request lists of a non active endpoint!\n"));
601
602	Assert(!pEndpoint->pFlushReq);
603
604	/* Check the pending list first */
605	if (pEndpoint->AioMgr.pReqsPendingHead)
606	{
607	LogFlow(("Queuing pending requests first\n"));
608
609	pTasksHead = pEndpoint->AioMgr.pReqsPendingHead;
610	/*
611	* Clear the list as the processing routine will insert them into the list
612	* again if it gets a flush request.
613	*/
614	pEndpoint->AioMgr.pReqsPendingHead = NULL;
615	pEndpoint->AioMgr.pReqsPendingTail = NULL;
616	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
617	AssertRC(rc);
618	}
619
620	if (!pEndpoint->pFlushReq && !pEndpoint->AioMgr.pReqsPendingHead)
621	{
622	/* Now the request queue. */
623	pTasksHead = pdmacFileEpGetNewTasks(pEndpoint);
624	if (pTasksHead)
625	{
626	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
627	AssertRC(rc);
628	}
629	}
630
631	return rc;
632	}
633
634	static int pdmacFileAioMgrNormalProcessBlockingEvent(PPDMACEPFILEMGR pAioMgr)
635	{
636	int rc = VINF_SUCCESS;
637	bool fNotifyWaiter = false;
638
639	LogFlowFunc((": Enter\n"));
640
641	Assert(pAioMgr->fBlockingEventPending);
642
643	switch (pAioMgr->enmBlockingEvent)
644	{
645	case PDMACEPFILEAIOMGRBLOCKINGEVENT_ADD_ENDPOINT:
646	{
647	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointNew = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint);
648	AssertMsg(VALID_PTR(pEndpointNew), ("Adding endpoint event without a endpoint to add\n"));
649
650	pEndpointNew->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE;
651
652	pEndpointNew->AioMgr.pEndpointNext = pAioMgr->pEndpointsHead;
653	pEndpointNew->AioMgr.pEndpointPrev = NULL;
654	if (pAioMgr->pEndpointsHead)
655	pAioMgr->pEndpointsHead->AioMgr.pEndpointPrev = pEndpointNew;
656	pAioMgr->pEndpointsHead = pEndpointNew;
657
658	/* Assign the completion point to this file. */
659	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pEndpointNew->File);
660	fNotifyWaiter = true;
661	pAioMgr->cEndpoints++;
662	break;
663	}
664	case PDMACEPFILEAIOMGRBLOCKINGEVENT_REMOVE_ENDPOINT:
665	{
666	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint);
667	AssertMsg(VALID_PTR(pEndpointRemove), ("Removing endpoint event without a endpoint to remove\n"));
668
669	pEndpointRemove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
670	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointRemove);
671	break;
672	}
673	case PDMACEPFILEAIOMGRBLOCKINGEVENT_CLOSE_ENDPOINT:
674	{
675	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointClose = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint);
676	AssertMsg(VALID_PTR(pEndpointClose), ("Close endpoint event without a endpoint to close\n"));
677
678	LogFlowFunc((": Closing endpoint %#p{%s}\n", pEndpointClose, pEndpointClose->Core.pszUri));
679
680	/* Make sure all tasks finished. Process the queues a last time first. */
681	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpointClose);
682	AssertRC(rc);
683
684	pEndpointClose->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING;
685	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointClose);
686	break;
687	}
688	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SHUTDOWN:
689	{
690	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SHUTDOWN;
691	if (!pAioMgr->cRequestsActive)
692	fNotifyWaiter = true;
693	break;
694	}
695	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SUSPEND:
696	{
697	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SUSPENDING;
698	break;
699	}
700	case PDMACEPFILEAIOMGRBLOCKINGEVENT_RESUME:
701	{
702	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
703	fNotifyWaiter = true;
704	break;
705	}
706	default:
707	AssertReleaseMsgFailed(("Invalid event type %d\n", pAioMgr->enmBlockingEvent));
708	}
709
710	if (fNotifyWaiter)
711	{
712	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
713	pAioMgr->enmBlockingEvent = PDMACEPFILEAIOMGRBLOCKINGEVENT_INVALID;
714
715	/* Release the waiting thread. */
716	LogFlow(("Signalling waiter\n"));
717	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
718	AssertRC(rc);
719	}
720
721	LogFlowFunc((": Leave\n"));
722	return rc;
723	}
724
725	/**
726	* Checks all endpoints for pending events or new requests.
727	*
728	* @returns VBox status code.
729	* @param pAioMgr The I/O manager handle.
730	*/
731	static int pdmacFileAioMgrNormalCheckEndpoints(PPDMACEPFILEMGR pAioMgr)
732	{
733	/* Check the assigned endpoints for new tasks if there isn't a flush request active at the moment. */
734	int rc = VINF_SUCCESS;
735	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint = pAioMgr->pEndpointsHead;
736
737	while (pEndpoint)
738	{
739	if (!pEndpoint->pFlushReq && (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE))
740	{
741	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
742	if (RT_FAILURE(rc))
743	return rc;
744	}
745	else if (!pEndpoint->AioMgr.cRequestsActive)
746	{
747	/* Reopen the file so that the new endpoint can reassociate with the file */
748	RTFileClose(pEndpoint->File);
749	rc = RTFileOpen(&pEndpoint->File, pEndpoint->Core.pszUri, pEndpoint->fFlags);
750	AssertRC(rc);
751
752	if (pEndpoint->AioMgr.fMoving)
753	{
754	pEndpoint->AioMgr.fMoving = false;
755	pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
756	}
757	else
758	{
759	Assert(pAioMgr->fBlockingEventPending);
760	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
761
762	/* Release the waiting thread. */
763	LogFlow(("Signalling waiter\n"));
764	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
765	AssertRC(rc);
766	}
767	}
768
769	pEndpoint = pEndpoint->AioMgr.pEndpointNext;
770	}
771
772	return rc;
773	}
774
775	/** Helper macro for checking for error codes. */
776	#define CHECK_RC(pAioMgr, rc) \
777	if (RT_FAILURE(rc)) \
778	{\
779	int rc2 = pdmacFileAioMgrNormalErrorHandler(pAioMgr, rc, RT_SRC_POS);\
780	return rc2;\
781	}
782
783	/**
784	* The normal I/O manager using the RTFileAio* API
785	*
786	* @returns VBox status code.
787	* @param ThreadSelf Handle of the thread.
788	* @param pvUser Opaque user data.
789	*/
790	int pdmacFileAioMgrNormal(RTTHREAD ThreadSelf, void *pvUser)
791	{
792	int rc = VINF_SUCCESS;
793	PPDMACEPFILEMGR pAioMgr = (PPDMACEPFILEMGR)pvUser;
794	uint64_t uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
795
796	while ( (pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING)
797	\|\| (pAioMgr->enmState == PDMACEPFILEMGRSTATE_SUSPENDING))
798	{
799	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, true);
800	if (!ASMAtomicReadBool(&pAioMgr->fWokenUp))
801	rc = RTSemEventWait(pAioMgr->EventSem, RT_INDEFINITE_WAIT);
802	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, false);
803	AssertRC(rc);
804
805	LogFlow(("Got woken up\n"));
806	ASMAtomicWriteBool(&pAioMgr->fWokenUp, false);
807
808	/* Check for an external blocking event first. */
809	if (pAioMgr->fBlockingEventPending)
810	{
811	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
812	CHECK_RC(pAioMgr, rc);
813	}
814
815	if (RT_LIKELY(pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING))
816	{
817	/* We got woken up because an endpoint issued new requests. Queue them. */
818	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
819	CHECK_RC(pAioMgr, rc);
820
821	while (pAioMgr->cRequestsActive)
822	{
823	RTFILEAIOREQ apReqs[20];
824	uint32_t cReqsCompleted = 0;
825	size_t cReqsWait;
826
827	if (pAioMgr->cRequestsActive > RT_ELEMENTS(apReqs))
828	cReqsWait = RT_ELEMENTS(apReqs);
829	else
830	cReqsWait = pAioMgr->cRequestsActive;
831
832	LogFlow(("Waiting for %d of %d tasks to complete\n", pAioMgr->cRequestsActive, cReqsWait));
833
834	rc = RTFileAioCtxWait(pAioMgr->hAioCtx,
835	cReqsWait,
836	RT_INDEFINITE_WAIT, apReqs,
837	RT_ELEMENTS(apReqs), &cReqsCompleted);
838	if (RT_FAILURE(rc) && (rc != VERR_INTERRUPTED))
839	CHECK_RC(pAioMgr, rc);
840
841	LogFlow(("%d tasks completed\n", cReqsCompleted));
842
843	for (uint32_t i = 0; i < cReqsCompleted; i++)
844	{
845	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint;
846	size_t cbTransfered = 0;
847	int rcReq = RTFileAioReqGetRC(apReqs[i], &cbTransfered);
848	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(apReqs[i]);
849
850	pEndpoint = pTask->pEndpoint;
851
852	AssertMsg( RT_SUCCESS(rcReq)
853	&& ( (cbTransfered == pTask->DataSeg.cbSeg)
854	\|\| (pTask->fBounceBuffer)),
855	("Task didn't completed successfully (rc=%Rrc) or was incomplete (cbTransfered=%u)\n", rcReq, cbTransfered));
856
857	if (pTask->fPrefetch)
858	{
859	Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE);
860	Assert(pTask->fBounceBuffer);
861
862	memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
863	pTask->DataSeg.pvSeg,
864	pTask->DataSeg.cbSeg);
865
866	/* Write it now. */
867	pTask->fPrefetch = false;
868	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg, 512);
869	RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1);
870
871	/* Grow the file if needed. */
872	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
873	{
874	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
875	RTFileSetSize(pEndpoint->File, pTask->Off + pTask->DataSeg.cbSeg);
876	}
877
878	rc = RTFileAioReqPrepareWrite(apReqs[i], pEndpoint->File,
879	offStart, pTask->pvBounceBuffer, cbToTransfer, pTask);
880	AssertRC(rc);
881	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, &apReqs[i], 1);
882	AssertRC(rc);
883	}
884	else
885	{
886	if (pTask->fBounceBuffer)
887	{
888	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
889	memcpy(pTask->DataSeg.pvSeg,
890	((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
891	pTask->DataSeg.cbSeg);
892
893	RTMemPageFree(pTask->pvBounceBuffer);
894	}
895
896	/* Put the entry on the free array */
897	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i];
898	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries;
899
900	pAioMgr->cRequestsActive--;
901	pEndpoint->AioMgr.cRequestsActive--;
902	pEndpoint->AioMgr.cReqsProcessed++;
903
904	/* Call completion callback */
905	pTask->pfnCompleted(pTask, pTask->pvUser);
906	pdmacFileTaskFree(pEndpoint, pTask);
907
908	/*
909	* If there is no request left on the endpoint but a flush request is set
910	* it completed now and we notify the owner.
911	* Furthermore we look for new requests and continue.
912	*/
913	if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq)
914	{
915	/* Call completion callback */
916	pTask = pEndpoint->pFlushReq;
917	pEndpoint->pFlushReq = NULL;
918
919	AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n"));
920
921	pTask->pfnCompleted(pTask, pTask->pvUser);
922	pdmacFileTaskFree(pEndpoint, pTask);
923	}
924	}
925	}
926
927	/* Check for an external blocking event before we go to sleep again. */
928	if (pAioMgr->fBlockingEventPending)
929	{
930	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
931	CHECK_RC(pAioMgr, rc);
932	}
933
934	/* Update load statistics. */
935	uint64_t uMillisCurr = RTTimeMilliTS();
936	if (uMillisCurr > uMillisEnd)
937	{
938	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointCurr = pAioMgr->pEndpointsHead;
939
940	/* Calculate timespan. */
941	uMillisCurr -= uMillisEnd;
942
943	while (pEndpointCurr)
944	{
945	pEndpointCurr->AioMgr.cReqsPerSec = pEndpointCurr->AioMgr.cReqsProcessed / (uMillisCurr + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD);
946	pEndpointCurr->AioMgr.cReqsProcessed = 0;
947	pEndpointCurr = pEndpointCurr->AioMgr.pEndpointNext;
948	}
949
950	/* Set new update interval */
951	uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
952	}
953
954	/* Check endpoints for new requests. */
955	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
956	CHECK_RC(pAioMgr, rc);
957	} /* while requests are active. */
958	} /* if still running */
959	} /* while running */
960
961	return rc;
962	}
963
964	#undef CHECK_RC
965

Note: See TracBrowser for help on using the repository browser.

source: vbox/trunk/src/VBox/VMM/PDMAsyncCompletionFileNormal.cpp@ 23911

Download in other formats: