PDMAsyncCompletionFileNormal.cpp@ 42062

Last change on this file since 42062 was 41862, checked in by vboxsync, 13 years ago
PDMAsyncCompletionFileNormal.cpp: Removed RT_STRICT as it breaks the strict build (added 'temporarily' about a year back, r71670).
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 64.4 KB

Line
1	/* $Id: PDMAsyncCompletionFileNormal.cpp 41862 2012-06-21 15:08:02Z vboxsync $ */
2	/** @file
3	* PDM Async I/O - Async File I/O manager.
4	*/
5
6	/*
7	* Copyright (C) 2006-2011 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	/*******************************************************************************
19	* Header Files *
20	*******************************************************************************/
21	#define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION
22	#include <iprt/types.h>
23	#include <iprt/asm.h>
24	#include <iprt/file.h>
25	#include <iprt/mem.h>
26	#include <iprt/string.h>
27	#include <iprt/assert.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_STEP 512
36
37
38	/*******************************************************************************
39	* Internal functions *
40	*******************************************************************************/
41	static int pdmacFileAioMgrNormalProcessTaskList(PPDMACTASKFILE pTaskHead,
42	PPDMACEPFILEMGR pAioMgr,
43	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint);
44
45	static PPDMACTASKFILE pdmacFileAioMgrNormalRangeLockFree(PPDMACEPFILEMGR pAioMgr,
46	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
47	PPDMACFILERANGELOCK pRangeLock);
48
49	static void pdmacFileAioMgrNormalReqCompleteRc(PPDMACEPFILEMGR pAioMgr, RTFILEAIOREQ hReq,
50	int rc, size_t cbTransfered);
51
52
53	int pdmacFileAioMgrNormalInit(PPDMACEPFILEMGR pAioMgr)
54	{
55	pAioMgr->cRequestsActiveMax = PDMACEPFILEMGR_REQS_STEP;
56
57	int rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, RTFILEAIO_UNLIMITED_REQS);
58	if (rc == VERR_OUT_OF_RANGE)
59	rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, pAioMgr->cRequestsActiveMax);
60
61	if (RT_SUCCESS(rc))
62	{
63	/* Initialize request handle array. */
64	pAioMgr->iFreeEntry = 0;
65	pAioMgr->cReqEntries = pAioMgr->cRequestsActiveMax;
66	pAioMgr->pahReqsFree = (RTFILEAIOREQ )RTMemAllocZ(pAioMgr->cReqEntries sizeof(RTFILEAIOREQ));
67
68	if (pAioMgr->pahReqsFree)
69	{
70	/* Create the range lock memcache. */
71	rc = RTMemCacheCreate(&pAioMgr->hMemCacheRangeLocks, sizeof(PDMACFILERANGELOCK),
72	0, UINT32_MAX, NULL, NULL, NULL, 0);
73	if (RT_SUCCESS(rc))
74	return VINF_SUCCESS;
75
76	RTMemFree(pAioMgr->pahReqsFree);
77	}
78	else
79	{
80	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
81	rc = VERR_NO_MEMORY;
82	}
83	}
84
85	return rc;
86	}
87
88	void pdmacFileAioMgrNormalDestroy(PPDMACEPFILEMGR pAioMgr)
89	{
90	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
91
92	while (pAioMgr->iFreeEntry > 0)
93	{
94	pAioMgr->iFreeEntry--;
95	Assert(pAioMgr->pahReqsFree[pAioMgr->iFreeEntry] != NIL_RTFILEAIOREQ);
96	RTFileAioReqDestroy(pAioMgr->pahReqsFree[pAioMgr->iFreeEntry]);
97	}
98
99	RTMemFree(pAioMgr->pahReqsFree);
100	RTMemCacheDestroy(pAioMgr->hMemCacheRangeLocks);
101	}
102
103	#if 0 /* currently unused */
104	/**
105	* Sorts the endpoint list with insertion sort.
106	*/
107	static void pdmacFileAioMgrNormalEndpointsSortByLoad(PPDMACEPFILEMGR pAioMgr)
108	{
109	PPDMASYNCCOMPLETIONENDPOINTFILE pEpPrev, pEpCurr, pEpNextToSort;
110
111	pEpPrev = pAioMgr->pEndpointsHead;
112	pEpCurr = pEpPrev->AioMgr.pEndpointNext;
113
114	while (pEpCurr)
115	{
116	/* Remember the next element to sort because the list might change. */
117	pEpNextToSort = pEpCurr->AioMgr.pEndpointNext;
118
119	/* Unlink the current element from the list. */
120	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
121	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
122
123	if (pPrev)
124	pPrev->AioMgr.pEndpointNext = pNext;
125	else
126	pAioMgr->pEndpointsHead = pNext;
127
128	if (pNext)
129	pNext->AioMgr.pEndpointPrev = pPrev;
130
131	/* Go back until we reached the place to insert the current endpoint into. */
132	while (pEpPrev && (pEpPrev->AioMgr.cReqsPerSec < pEpCurr->AioMgr.cReqsPerSec))
133	pEpPrev = pEpPrev->AioMgr.pEndpointPrev;
134
135	/* Link the endpoint into the list. */
136	if (pEpPrev)
137	pNext = pEpPrev->AioMgr.pEndpointNext;
138	else
139	pNext = pAioMgr->pEndpointsHead;
140
141	pEpCurr->AioMgr.pEndpointNext = pNext;
142	pEpCurr->AioMgr.pEndpointPrev = pEpPrev;
143
144	if (pNext)
145	pNext->AioMgr.pEndpointPrev = pEpCurr;
146
147	if (pEpPrev)
148	pEpPrev->AioMgr.pEndpointNext = pEpCurr;
149	else
150	pAioMgr->pEndpointsHead = pEpCurr;
151
152	pEpCurr = pEpNextToSort;
153	}
154
155	#ifdef DEBUG
156	/* Validate sorting algorithm */
157	unsigned cEndpoints = 0;
158	pEpCurr = pAioMgr->pEndpointsHead;
159
160	AssertMsg(pEpCurr, ("No endpoint in the list?\n"));
161	AssertMsg(!pEpCurr->AioMgr.pEndpointPrev, ("First element in the list points to previous element\n"));
162
163	while (pEpCurr)
164	{
165	cEndpoints++;
166
167	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
168	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
169
170	Assert(!pNext \|\| pNext->AioMgr.cReqsPerSec <= pEpCurr->AioMgr.cReqsPerSec);
171	Assert(!pPrev \|\| pPrev->AioMgr.cReqsPerSec >= pEpCurr->AioMgr.cReqsPerSec);
172
173	pEpCurr = pNext;
174	}
175
176	AssertMsg(cEndpoints == pAioMgr->cEndpoints, ("Endpoints lost during sort!\n"));
177
178	#endif
179	}
180	#endif /* currently unused */
181
182	/**
183	* Removes an endpoint from the currently assigned manager.
184	*
185	* @returns TRUE if there are still requests pending on the current manager for this endpoint.
186	* FALSE otherwise.
187	* @param pEndpointRemove The endpoint to remove.
188	*/
189	static bool pdmacFileAioMgrNormalRemoveEndpoint(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove)
190	{
191	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEndpointRemove->AioMgr.pEndpointPrev;
192	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEndpointRemove->AioMgr.pEndpointNext;
193	PPDMACEPFILEMGR pAioMgr = pEndpointRemove->pAioMgr;
194
195	pAioMgr->cEndpoints--;
196
197	if (pPrev)
198	pPrev->AioMgr.pEndpointNext = pNext;
199	else
200	pAioMgr->pEndpointsHead = pNext;
201
202	if (pNext)
203	pNext->AioMgr.pEndpointPrev = pPrev;
204
205	/* Make sure that there is no request pending on this manager for the endpoint. */
206	if (!pEndpointRemove->AioMgr.cRequestsActive)
207	{
208	Assert(!pEndpointRemove->pFlushReq);
209
210	/* Reopen the file so that the new endpoint can re-associate with the file */
211	RTFileClose(pEndpointRemove->hFile);
212	int rc = RTFileOpen(&pEndpointRemove->hFile, pEndpointRemove->Core.pszUri, pEndpointRemove->fFlags);
213	AssertRC(rc);
214	return false;
215	}
216
217	return true;
218	}
219
220	#if 0 /* currently unused */
221
222	static bool pdmacFileAioMgrNormalIsBalancePossible(PPDMACEPFILEMGR pAioMgr)
223	{
224	/* Balancing doesn't make sense with only one endpoint. */
225	if (pAioMgr->cEndpoints == 1)
226	return false;
227
228	/* Doesn't make sens to move endpoints if only one produces the whole load */
229	unsigned cEndpointsWithLoad = 0;
230
231	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead;
232
233	while (pCurr)
234	{
235	if (pCurr->AioMgr.cReqsPerSec)
236	cEndpointsWithLoad++;
237
238	pCurr = pCurr->AioMgr.pEndpointNext;
239	}
240
241	return (cEndpointsWithLoad > 1);
242	}
243
244	/**
245	* Creates a new I/O manager and spreads the I/O load of the endpoints
246	* between the given I/O manager and the new one.
247	*
248	* @returns nothing.
249	* @param pAioMgr The I/O manager with high I/O load.
250	*/
251	static void pdmacFileAioMgrNormalBalanceLoad(PPDMACEPFILEMGR pAioMgr)
252	{
253	/*
254	* Check if balancing would improve the situation.
255	*/
256	if (pdmacFileAioMgrNormalIsBalancePossible(pAioMgr))
257	{
258	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass;
259	PPDMACEPFILEMGR pAioMgrNew = NULL;
260
261	int rc = pdmacFileAioMgrCreate(pEpClassFile, &pAioMgrNew, PDMACEPFILEMGRTYPE_ASYNC);
262	if (RT_SUCCESS(rc))
263	{
264	/* We will sort the list by request count per second. */
265	pdmacFileAioMgrNormalEndpointsSortByLoad(pAioMgr);
266
267	/* Now move some endpoints to the new manager. */
268	unsigned cReqsHere = pAioMgr->pEndpointsHead->AioMgr.cReqsPerSec;
269	unsigned cReqsOther = 0;
270	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead->AioMgr.pEndpointNext;
271
272	while (pCurr)
273	{
274	if (cReqsHere <= cReqsOther)
275	{
276	/*
277	* The other manager has more requests to handle now.
278	* We will keep the current endpoint.
279	*/
280	Log(("Keeping endpoint %#p{%s} with %u reqs/s\n", pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
281	cReqsHere += pCurr->AioMgr.cReqsPerSec;
282	pCurr = pCurr->AioMgr.pEndpointNext;
283	}
284	else
285	{
286	/* Move to other endpoint. */
287	Log(("Moving endpoint %#p{%s} with %u reqs/s to other manager\n", pCurr, pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
288	cReqsOther += pCurr->AioMgr.cReqsPerSec;
289
290	PPDMASYNCCOMPLETIONENDPOINTFILE pMove = pCurr;
291
292	pCurr = pCurr->AioMgr.pEndpointNext;
293
294	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pMove);
295
296	if (fReqsPending)
297	{
298	pMove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
299	pMove->AioMgr.fMoving = true;
300	pMove->AioMgr.pAioMgrDst = pAioMgrNew;
301	}
302	else
303	{
304	pMove->AioMgr.fMoving = false;
305	pMove->AioMgr.pAioMgrDst = NULL;
306	pdmacFileAioMgrAddEndpoint(pAioMgrNew, pMove);
307	}
308	}
309	}
310	}
311	else
312	{
313	/* Don't process further but leave a log entry about reduced performance. */
314	LogRel(("AIOMgr: Could not create new I/O manager (rc=%Rrc). Expect reduced performance\n", rc));
315	}
316	}
317	else
318	Log(("AIOMgr: Load balancing would not improve anything\n"));
319	}
320
321	#endif /* unused */
322
323	/**
324	* Increase the maximum number of active requests for the given I/O manager.
325	*
326	* @returns VBox status code.
327	* @param pAioMgr The I/O manager to grow.
328	*/
329	static int pdmacFileAioMgrNormalGrow(PPDMACEPFILEMGR pAioMgr)
330	{
331	LogFlowFunc(("pAioMgr=%#p\n", pAioMgr));
332
333	AssertMsg( pAioMgr->enmState == PDMACEPFILEMGRSTATE_GROWING
334	&& !pAioMgr->cRequestsActive,
335	("Invalid state of the I/O manager\n"));
336
337	#ifdef RT_OS_WINDOWS
338	/*
339	* Reopen the files of all assigned endpoints first so we can assign them to the new
340	* I/O context.
341	*/
342	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead;
343
344	while (pCurr)
345	{
346	RTFileClose(pCurr->hFile);
347	int rc2 = RTFileOpen(&pCurr->hFile, pCurr->Core.pszUri, pCurr->fFlags); AssertRC(rc2);
348
349	pCurr = pCurr->AioMgr.pEndpointNext;
350	}
351	#endif
352
353	/* Create the new bigger context. */
354	pAioMgr->cRequestsActiveMax += PDMACEPFILEMGR_REQS_STEP;
355
356	RTFILEAIOCTX hAioCtxNew = NIL_RTFILEAIOCTX;
357	int rc = RTFileAioCtxCreate(&hAioCtxNew, RTFILEAIO_UNLIMITED_REQS);
358	if (rc == VERR_OUT_OF_RANGE)
359	rc = RTFileAioCtxCreate(&hAioCtxNew, pAioMgr->cRequestsActiveMax);
360
361	if (RT_SUCCESS(rc))
362	{
363	/* Close the old context. */
364	rc = RTFileAioCtxDestroy(pAioMgr->hAioCtx);
365	AssertRC(rc); /** @todo r=bird: Ignoring error code, will propagate. */
366
367	pAioMgr->hAioCtx = hAioCtxNew;
368
369	/* Create a new I/O task handle array */
370	uint32_t cReqEntriesNew = pAioMgr->cRequestsActiveMax + 1;
371	RTFILEAIOREQ pahReqNew = (RTFILEAIOREQ )RTMemAllocZ(cReqEntriesNew * sizeof(RTFILEAIOREQ));
372
373	if (pahReqNew)
374	{
375	/* Copy the cached request handles. */
376	for (uint32_t iReq = 0; iReq < pAioMgr->cReqEntries; iReq++)
377	pahReqNew[iReq] = pAioMgr->pahReqsFree[iReq];
378
379	RTMemFree(pAioMgr->pahReqsFree);
380	pAioMgr->pahReqsFree = pahReqNew;
381	pAioMgr->cReqEntries = cReqEntriesNew;
382	LogFlowFunc(("I/O manager increased to handle a maximum of %u requests\n",
383	pAioMgr->cRequestsActiveMax));
384	}
385	else
386	rc = VERR_NO_MEMORY;
387	}
388
389	#ifdef RT_OS_WINDOWS
390	/* Assign the file to the new context. */
391	pCurr = pAioMgr->pEndpointsHead;
392	while (pCurr)
393	{
394	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pCurr->hFile);
395	AssertRC(rc); /** @todo r=bird: Ignoring error code, will propagate. */
396
397	pCurr = pCurr->AioMgr.pEndpointNext;
398	}
399	#endif
400
401	if (RT_FAILURE(rc))
402	{
403	LogFlow(("Increasing size of the I/O manager failed with rc=%Rrc\n", rc));
404	pAioMgr->cRequestsActiveMax -= PDMACEPFILEMGR_REQS_STEP;
405	}
406
407	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
408	LogFlowFunc(("returns rc=%Rrc\n", rc));
409
410	return rc;
411	}
412
413	/**
414	* Checks if a given status code is fatal.
415	* Non fatal errors can be fixed by migrating the endpoint to a
416	* failsafe manager.
417	*
418	* @returns true If the error is fatal and migrating to a failsafe manager doesn't help
419	* false If the error can be fixed by a migration. (image on NFS disk for example)
420	* @param rcReq The status code to check.
421	*/
422	DECLINLINE(bool) pdmacFileAioMgrNormalRcIsFatal(int rcReq)
423	{
424	return rcReq == VERR_DEV_IO_ERROR
425	\|\| rcReq == VERR_FILE_IO_ERROR
426	\|\| rcReq == VERR_DISK_IO_ERROR
427	\|\| rcReq == VERR_DISK_FULL
428	\|\| rcReq == VERR_FILE_TOO_BIG;
429	}
430
431	/**
432	* Error handler which will create the failsafe managers and destroy the failed I/O manager.
433	*
434	* @returns VBox status code
435	* @param pAioMgr The I/O manager the error occurred on.
436	* @param rc The error code.
437	*/
438	static int pdmacFileAioMgrNormalErrorHandler(PPDMACEPFILEMGR pAioMgr, int rc, RT_SRC_POS_DECL)
439	{
440	LogRel(("AIOMgr: I/O manager %#p encountered a critical error (rc=%Rrc) during operation. Falling back to failsafe mode. Expect reduced performance\n",
441	pAioMgr, rc));
442	LogRel(("AIOMgr: Error happened in %s:(%u){%s}\n", RT_SRC_POS_ARGS));
443	LogRel(("AIOMgr: Please contact the product vendor\n"));
444
445	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass;
446
447	pAioMgr->enmState = PDMACEPFILEMGRSTATE_FAULT;
448	ASMAtomicWriteU32((volatile uint32_t *)&pEpClassFile->enmMgrTypeOverride, PDMACEPFILEMGRTYPE_SIMPLE);
449
450	AssertMsgFailed(("Implement\n"));
451	return VINF_SUCCESS;
452	}
453
454	/**
455	* Put a list of tasks in the pending request list of an endpoint.
456	*/
457	DECLINLINE(void) pdmacFileAioMgrEpAddTaskList(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTaskHead)
458	{
459	/* Add the rest of the tasks to the pending list */
460	if (!pEndpoint->AioMgr.pReqsPendingHead)
461	{
462	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
463	pEndpoint->AioMgr.pReqsPendingHead = pTaskHead;
464	}
465	else
466	{
467	Assert(pEndpoint->AioMgr.pReqsPendingTail);
468	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTaskHead;
469	}
470
471	/* Update the tail. */
472	while (pTaskHead->pNext)
473	pTaskHead = pTaskHead->pNext;
474
475	pEndpoint->AioMgr.pReqsPendingTail = pTaskHead;
476	pTaskHead->pNext = NULL;
477	}
478
479	/**
480	* Put one task in the pending request list of an endpoint.
481	*/
482	DECLINLINE(void) pdmacFileAioMgrEpAddTask(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTask)
483	{
484	/* Add the rest of the tasks to the pending list */
485	if (!pEndpoint->AioMgr.pReqsPendingHead)
486	{
487	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
488	pEndpoint->AioMgr.pReqsPendingHead = pTask;
489	}
490	else
491	{
492	Assert(pEndpoint->AioMgr.pReqsPendingTail);
493	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTask;
494	}
495
496	pEndpoint->AioMgr.pReqsPendingTail = pTask;
497	pTask->pNext = NULL;
498	}
499
500	/**
501	* Allocates a async I/O request.
502	*
503	* @returns Handle to the request.
504	* @param pAioMgr The I/O manager.
505	*/
506	static RTFILEAIOREQ pdmacFileAioMgrNormalRequestAlloc(PPDMACEPFILEMGR pAioMgr)
507	{
508	/* Get a request handle. */
509	RTFILEAIOREQ hReq;
510	if (pAioMgr->iFreeEntry > 0)
511	{
512	pAioMgr->iFreeEntry--;
513	hReq = pAioMgr->pahReqsFree[pAioMgr->iFreeEntry];
514	pAioMgr->pahReqsFree[pAioMgr->iFreeEntry] = NIL_RTFILEAIOREQ;
515	Assert(hReq != NIL_RTFILEAIOREQ);
516	}
517	else
518	{
519	int rc = RTFileAioReqCreate(&hReq);
520	AssertRCReturn(rc, NIL_RTFILEAIOREQ);
521	}
522
523	return hReq;
524	}
525
526	/**
527	* Frees a async I/O request handle.
528	*
529	* @returns nothing.
530	* @param pAioMgr The I/O manager.
531	* @param hReq The I/O request handle to free.
532	*/
533	static void pdmacFileAioMgrNormalRequestFree(PPDMACEPFILEMGR pAioMgr, RTFILEAIOREQ hReq)
534	{
535	Assert(pAioMgr->iFreeEntry < pAioMgr->cReqEntries);
536	Assert(pAioMgr->pahReqsFree[pAioMgr->iFreeEntry] == NIL_RTFILEAIOREQ);
537
538	pAioMgr->pahReqsFree[pAioMgr->iFreeEntry] = hReq;
539	pAioMgr->iFreeEntry++;
540	}
541
542	/**
543	* Wrapper around RTFIleAioCtxSubmit() which is also doing error handling.
544	*/
545	static int pdmacFileAioMgrNormalReqsEnqueue(PPDMACEPFILEMGR pAioMgr,
546	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
547	PRTFILEAIOREQ pahReqs, unsigned cReqs)
548	{
549	pAioMgr->cRequestsActive += cReqs;
550	pEndpoint->AioMgr.cRequestsActive += cReqs;
551
552	LogFlow(("Enqueuing %d requests. I/O manager has a total of %d active requests now\n", cReqs, pAioMgr->cRequestsActive));
553	LogFlow(("Endpoint has a total of %d active requests now\n", pEndpoint->AioMgr.cRequestsActive));
554
555	int rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, pahReqs, cReqs);
556	if (RT_FAILURE(rc))
557	{
558	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
559	{
560	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
561
562	/* Append any not submitted task to the waiting list. */
563	for (size_t i = 0; i < cReqs; i++)
564	{
565	int rcReq = RTFileAioReqGetRC(pahReqs[i], NULL);
566
567	if (rcReq != VERR_FILE_AIO_IN_PROGRESS)
568	{
569	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(pahReqs[i]);
570
571	Assert(pTask->hReq == pahReqs[i]);
572	pdmacFileAioMgrEpAddTask(pEndpoint, pTask);
573	pAioMgr->cRequestsActive--;
574	pEndpoint->AioMgr.cRequestsActive--;
575
576	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_FLUSH)
577	{
578	/* Clear the pending flush */
579	Assert(pEndpoint->pFlushReq == pTask);
580	pEndpoint->pFlushReq = NULL;
581	}
582	}
583	}
584
585	pAioMgr->cRequestsActiveMax = pAioMgr->cRequestsActive;
586
587	/* Print an entry in the release log */
588	if (RT_UNLIKELY(!pEpClass->fOutOfResourcesWarningPrinted))
589	{
590	pEpClass->fOutOfResourcesWarningPrinted = true;
591	LogRel(("AIOMgr: Host limits number of active IO requests to %u. Expect a performance impact.\n",
592	pAioMgr->cRequestsActive));
593	}
594
595	LogFlow(("Removed requests. I/O manager has a total of %u active requests now\n", pAioMgr->cRequestsActive));
596	LogFlow(("Endpoint has a total of %u active requests now\n", pEndpoint->AioMgr.cRequestsActive));
597	rc = VINF_SUCCESS;
598	}
599	else /* Another kind of error happened (full disk, ...) */
600	{
601	/* An error happened. Find out which one caused the error and resubmit all other tasks. */
602	for (size_t i = 0; i < cReqs; i++)
603	{
604	int rcReq = RTFileAioReqGetRC(pahReqs[i], NULL);
605
606	if (rcReq == VERR_FILE_AIO_NOT_SUBMITTED)
607	{
608	/* We call ourself again to do any error handling which might come up now. */
609	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, &pahReqs[i], 1);
610	AssertRC(rc);
611	}
612	else if (rcReq != VERR_FILE_AIO_IN_PROGRESS)
613	pdmacFileAioMgrNormalReqCompleteRc(pAioMgr, pahReqs[i], rcReq, 0);
614	}
615
616
617	if ( pEndpoint->pFlushReq
618	&& !pAioMgr->cRequestsActive
619	&& !pEndpoint->fAsyncFlushSupported)
620	{
621	/*
622	* Complete a pending flush if we don't have requests enqueued and the host doesn't support
623	* the async flush API.
624	* Happens only if this we just noticed that this is not supported
625	* and the only active request was a flush.
626	*/
627	PPDMACTASKFILE pFlush = pEndpoint->pFlushReq;
628	pEndpoint->pFlushReq = NULL;
629	pFlush->pfnCompleted(pFlush, pFlush->pvUser, VINF_SUCCESS);
630	pdmacFileTaskFree(pEndpoint, pFlush);
631	}
632	}
633	}
634
635	return VINF_SUCCESS;
636	}
637
638	static bool pdmacFileAioMgrNormalIsRangeLocked(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
639	RTFOFF offStart, size_t cbRange,
640	PPDMACTASKFILE pTask)
641	{
642	AssertMsg( pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE
643	\|\| pTask->enmTransferType == PDMACTASKFILETRANSFER_READ,
644	("Invalid task type %d\n", pTask->enmTransferType));
645
646	PPDMACFILERANGELOCK pRangeLock;
647	pRangeLock = (PPDMACFILERANGELOCK)RTAvlrFileOffsetRangeGet(pEndpoint->AioMgr.pTreeRangesLocked, offStart);
648	if (!pRangeLock)
649	{
650	pRangeLock = (PPDMACFILERANGELOCK)RTAvlrFileOffsetGetBestFit(pEndpoint->AioMgr.pTreeRangesLocked, offStart, true);
651	/* Check if we intersect with the range. */
652	if ( !pRangeLock
653	\|\| !( (pRangeLock->Core.Key) <= (offStart + (RTFOFF)cbRange - 1)
654	&& (pRangeLock->Core.KeyLast) >= offStart))
655	{
656	pRangeLock = NULL; /* False alarm */
657	}
658	}
659
660	/* Check whether we have one of the situations explained below */
661	if ( pRangeLock
662	#if 0 /** @todo later. For now we will just block all requests if they interfere */
663	&& ( (pRangeLock->fReadLock && pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
664	\|\| (!pRangeLock->fReadLock)
665	#endif
666	)
667	{
668	/* Add to the list. */
669	pTask->pNext = NULL;
670
671	if (!pRangeLock->pWaitingTasksHead)
672	{
673	Assert(!pRangeLock->pWaitingTasksTail);
674	pRangeLock->pWaitingTasksHead = pTask;
675	pRangeLock->pWaitingTasksTail = pTask;
676	}
677	else
678	{
679	AssertPtr(pRangeLock->pWaitingTasksTail);
680	pRangeLock->pWaitingTasksTail->pNext = pTask;
681	pRangeLock->pWaitingTasksTail = pTask;
682	}
683	return true;
684	}
685
686	return false;
687	}
688
689	static int pdmacFileAioMgrNormalRangeLock(PPDMACEPFILEMGR pAioMgr,
690	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
691	RTFOFF offStart, size_t cbRange,
692	PPDMACTASKFILE pTask)
693	{
694	AssertMsg(!pdmacFileAioMgrNormalIsRangeLocked(pEndpoint, offStart, cbRange, pTask),
695	("Range is already locked offStart=%RTfoff cbRange=%u\n",
696	offStart, cbRange));
697
698	PPDMACFILERANGELOCK pRangeLock = (PPDMACFILERANGELOCK)RTMemCacheAlloc(pAioMgr->hMemCacheRangeLocks);
699	if (!pRangeLock)
700	return VERR_NO_MEMORY;
701
702	/* Init the lock. */
703	pRangeLock->Core.Key = offStart;
704	pRangeLock->Core.KeyLast = offStart + cbRange - 1;
705	pRangeLock->cRefs = 1;
706	pRangeLock->fReadLock = pTask->enmTransferType == PDMACTASKFILETRANSFER_READ;
707	pRangeLock->pWaitingTasksHead = NULL;
708	pRangeLock->pWaitingTasksTail = NULL;
709
710	bool fInserted = RTAvlrFileOffsetInsert(pEndpoint->AioMgr.pTreeRangesLocked, &pRangeLock->Core);
711	AssertMsg(fInserted, ("Range lock was not inserted!\n"));
712
713	/* Let the task point to its lock. */
714	pTask->pRangeLock = pRangeLock;
715
716	return VINF_SUCCESS;
717	}
718
719	static PPDMACTASKFILE pdmacFileAioMgrNormalRangeLockFree(PPDMACEPFILEMGR pAioMgr,
720	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
721	PPDMACFILERANGELOCK pRangeLock)
722	{
723	PPDMACTASKFILE pTasksWaitingHead;
724
725	AssertPtr(pRangeLock);
726	Assert(pRangeLock->cRefs == 1);
727
728	RTAvlrFileOffsetRemove(pEndpoint->AioMgr.pTreeRangesLocked, pRangeLock->Core.Key);
729	pTasksWaitingHead = pRangeLock->pWaitingTasksHead;
730	pRangeLock->pWaitingTasksHead = NULL;
731	pRangeLock->pWaitingTasksTail = NULL;
732	RTMemCacheFree(pAioMgr->hMemCacheRangeLocks, pRangeLock);
733
734	return pTasksWaitingHead;
735	}
736
737	static int pdmacFileAioMgrNormalTaskPrepareBuffered(PPDMACEPFILEMGR pAioMgr,
738	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
739	PPDMACTASKFILE pTask, PRTFILEAIOREQ phReq)
740	{
741	AssertMsg( pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE
742	\|\| (uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) <= pEndpoint->cbFile,
743	("Read exceeds file size offStart=%RTfoff cbToTransfer=%d cbFile=%llu\n",
744	pTask->Off, pTask->DataSeg.cbSeg, pEndpoint->cbFile));
745
746	pTask->fPrefetch = false;
747	pTask->cbBounceBuffer = 0;
748
749	/*
750	* Before we start to setup the request we have to check whether there is a task
751	* already active which range intersects with ours. We have to defer execution
752	* of this task in two cases:
753	* - The pending task is a write and the current is either read or write
754	* - The pending task is a read and the current task is a write task.
755	*
756	* To check whether a range is currently "locked" we use the AVL tree where every pending task
757	* is stored by its file offset range. The current task will be added to the active task
758	* and will be executed when the active one completes. (The method below
759	* which checks whether a range is already used will add the task)
760	*
761	* This is necessary because of the requirement to align all requests to a 512 boundary
762	* which is enforced by the host OS (Linux and Windows atm). It is possible that
763	* we have to process unaligned tasks and need to align them using bounce buffers.
764	* While the data is fetched from the file another request might arrive writing to
765	* the same range. This will result in data corruption if both are executed concurrently.
766	*/
767	int rc = VINF_SUCCESS;
768	bool fLocked = pdmacFileAioMgrNormalIsRangeLocked(pEndpoint, pTask->Off, pTask->DataSeg.cbSeg, pTask);
769	if (!fLocked)
770	{
771	/* Get a request handle. */
772	RTFILEAIOREQ hReq = pdmacFileAioMgrNormalRequestAlloc(pAioMgr);
773	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
774
775	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
776	{
777	/* Grow the file if needed. */
778	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
779	{
780	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
781	RTFileSetSize(pEndpoint->hFile, pTask->Off + pTask->DataSeg.cbSeg);
782	}
783
784	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->hFile,
785	pTask->Off, pTask->DataSeg.pvSeg,
786	pTask->DataSeg.cbSeg, pTask);
787	}
788	else
789	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->hFile,
790	pTask->Off, pTask->DataSeg.pvSeg,
791	pTask->DataSeg.cbSeg, pTask);
792	AssertRC(rc);
793
794	rc = pdmacFileAioMgrNormalRangeLock(pAioMgr, pEndpoint, pTask->Off,
795	pTask->DataSeg.cbSeg,
796	pTask);
797
798	if (RT_SUCCESS(rc))
799	{
800	pTask->hReq = hReq;
801	*phReq = hReq;
802	}
803	}
804	else
805	LogFlow(("Task %#p was deferred because the access range is locked\n", pTask));
806
807	return rc;
808	}
809
810	static int pdmacFileAioMgrNormalTaskPrepareNonBuffered(PPDMACEPFILEMGR pAioMgr,
811	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
812	PPDMACTASKFILE pTask, PRTFILEAIOREQ phReq)
813	{
814	/*
815	* Check if the alignment requirements are met.
816	* Offset, transfer size and buffer address
817	* need to be on a 512 boundary.
818	*/
819	RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1);
820	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg + (pTask->Off - offStart), 512);
821	PDMACTASKFILETRANSFER enmTransferType = pTask->enmTransferType;
822
823	AssertMsg( pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE
824	\|\| (uint64_t)(offStart + cbToTransfer) <= pEndpoint->cbFile,
825	("Read exceeds file size offStart=%RTfoff cbToTransfer=%d cbFile=%llu\n",
826	offStart, cbToTransfer, pEndpoint->cbFile));
827
828	pTask->fPrefetch = false;
829
830	/*
831	* Before we start to setup the request we have to check whether there is a task
832	* already active which range intersects with ours. We have to defer execution
833	* of this task in two cases:
834	* - The pending task is a write and the current is either read or write
835	* - The pending task is a read and the current task is a write task.
836	*
837	* To check whether a range is currently "locked" we use the AVL tree where every pending task
838	* is stored by its file offset range. The current task will be added to the active task
839	* and will be executed when the active one completes. (The method below
840	* which checks whether a range is already used will add the task)
841	*
842	* This is necessary because of the requirement to align all requests to a 512 boundary
843	* which is enforced by the host OS (Linux and Windows atm). It is possible that
844	* we have to process unaligned tasks and need to align them using bounce buffers.
845	* While the data is fetched from the file another request might arrive writing to
846	* the same range. This will result in data corruption if both are executed concurrently.
847	*/
848	int rc = VINF_SUCCESS;
849	bool fLocked = pdmacFileAioMgrNormalIsRangeLocked(pEndpoint, offStart, cbToTransfer, pTask);
850	if (!fLocked)
851	{
852	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
853	void *pvBuf = pTask->DataSeg.pvSeg;
854
855	/* Get a request handle. */
856	RTFILEAIOREQ hReq = pdmacFileAioMgrNormalRequestAlloc(pAioMgr);
857	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
858
859	if ( RT_UNLIKELY(cbToTransfer != pTask->DataSeg.cbSeg)
860	\|\| RT_UNLIKELY(offStart != pTask->Off)
861	\|\| ((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) != (RTR3UINTPTR)pvBuf))
862	{
863	LogFlow(("Using bounce buffer for task %#p cbToTransfer=%zd cbSeg=%zd offStart=%RTfoff off=%RTfoff\n",
864	pTask, cbToTransfer, pTask->DataSeg.cbSeg, offStart, pTask->Off));
865
866	/* Create bounce buffer. */
867	pTask->cbBounceBuffer = cbToTransfer;
868
869	AssertMsg(pTask->Off >= offStart, ("Overflow in calculation Off=%llu offStart=%llu\n",
870	pTask->Off, offStart));
871	pTask->offBounceBuffer = pTask->Off - offStart;
872
873	/** @todo: I think we need something like a RTMemAllocAligned method here.
874	* Current assumption is that the maximum alignment is 4096byte
875	* (GPT disk on Windows)
876	* so we can use RTMemPageAlloc here.
877	*/
878	pTask->pvBounceBuffer = RTMemPageAlloc(cbToTransfer);
879	if (RT_LIKELY(pTask->pvBounceBuffer))
880	{
881	pvBuf = pTask->pvBounceBuffer;
882
883	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
884	{
885	if ( RT_UNLIKELY(cbToTransfer != pTask->DataSeg.cbSeg)
886	\|\| RT_UNLIKELY(offStart != pTask->Off))
887	{
888	/* We have to fill the buffer first before we can update the data. */
889	LogFlow(("Prefetching data for task %#p\n", pTask));
890	pTask->fPrefetch = true;
891	enmTransferType = PDMACTASKFILETRANSFER_READ;
892	}
893	else
894	memcpy(pvBuf, pTask->DataSeg.pvSeg, pTask->DataSeg.cbSeg);
895	}
896	}
897	else
898	rc = VERR_NO_MEMORY;
899	}
900	else
901	pTask->cbBounceBuffer = 0;
902
903	if (RT_SUCCESS(rc))
904	{
905	AssertMsg((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) == (RTR3UINTPTR)pvBuf,
906	("AIO: Alignment restrictions not met! pvBuf=%p uBitmaskAlignment=%p\n", pvBuf, pEpClassFile->uBitmaskAlignment));
907
908	if (enmTransferType == PDMACTASKFILETRANSFER_WRITE)
909	{
910	/* Grow the file if needed. */
911	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
912	{
913	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
914	RTFileSetSize(pEndpoint->hFile, pTask->Off + pTask->DataSeg.cbSeg);
915	}
916
917	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->hFile,
918	offStart, pvBuf, cbToTransfer, pTask);
919	}
920	else
921	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->hFile,
922	offStart, pvBuf, cbToTransfer, pTask);
923	AssertRC(rc);
924
925	rc = pdmacFileAioMgrNormalRangeLock(pAioMgr, pEndpoint, offStart, cbToTransfer, pTask);
926
927	if (RT_SUCCESS(rc))
928	{
929	pTask->hReq = hReq;
930	*phReq = hReq;
931	}
932	else
933	{
934	/* Cleanup */
935	if (pTask->cbBounceBuffer)
936	RTMemPageFree(pTask->pvBounceBuffer, pTask->cbBounceBuffer);
937	}
938	}
939	}
940	else
941	LogFlow(("Task %#p was deferred because the access range is locked\n", pTask));
942
943	return rc;
944	}
945
946	static int pdmacFileAioMgrNormalProcessTaskList(PPDMACTASKFILE pTaskHead,
947	PPDMACEPFILEMGR pAioMgr,
948	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
949	{
950	RTFILEAIOREQ apReqs[20];
951	unsigned cRequests = 0;
952	int rc = VINF_SUCCESS;
953
954	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
955	("Trying to process request lists of a non active endpoint!\n"));
956
957	/* Go through the list and queue the requests until we get a flush request */
958	while ( pTaskHead
959	&& !pEndpoint->pFlushReq
960	&& (pAioMgr->cRequestsActive + cRequests < pAioMgr->cRequestsActiveMax)
961	&& RT_SUCCESS(rc))
962	{
963	RTMSINTERVAL msWhenNext;
964	PPDMACTASKFILE pCurr = pTaskHead;
965
966	if (!pdmacEpIsTransferAllowed(&pEndpoint->Core, (uint32_t)pCurr->DataSeg.cbSeg, &msWhenNext))
967	{
968	pAioMgr->msBwLimitExpired = RT_MIN(pAioMgr->msBwLimitExpired, msWhenNext);
969	break;
970	}
971
972	pTaskHead = pTaskHead->pNext;
973
974	pCurr->pNext = NULL;
975
976	AssertMsg(VALID_PTR(pCurr->pEndpoint) && (pCurr->pEndpoint == pEndpoint),
977	("Endpoints do not match\n"));
978
979	switch (pCurr->enmTransferType)
980	{
981	case PDMACTASKFILETRANSFER_FLUSH:
982	{
983	/* If there is no data transfer request this flush request finished immediately. */
984	if (pEndpoint->fAsyncFlushSupported)
985	{
986	/* Issue a flush to the host. */
987	RTFILEAIOREQ hReq = pdmacFileAioMgrNormalRequestAlloc(pAioMgr);
988	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
989
990	LogFlow(("Flush request %#p\n", hReq));
991
992	rc = RTFileAioReqPrepareFlush(hReq, pEndpoint->hFile, pCurr);
993	if (RT_FAILURE(rc))
994	{
995	pEndpoint->fAsyncFlushSupported = false;
996	pdmacFileAioMgrNormalRequestFree(pAioMgr, hReq);
997	rc = VINF_SUCCESS; /* Fake success */
998	}
999	else
1000	{
1001	pCurr->hReq = hReq;
1002	apReqs[cRequests] = hReq;
1003	pEndpoint->AioMgr.cReqsProcessed++;
1004	cRequests++;
1005	}
1006	}
1007
1008	if ( !pEndpoint->AioMgr.cRequestsActive
1009	&& !pEndpoint->fAsyncFlushSupported)
1010	{
1011	pCurr->pfnCompleted(pCurr, pCurr->pvUser, VINF_SUCCESS);
1012	pdmacFileTaskFree(pEndpoint, pCurr);
1013	}
1014	else
1015	{
1016	Assert(!pEndpoint->pFlushReq);
1017	pEndpoint->pFlushReq = pCurr;
1018	}
1019	break;
1020	}
1021	case PDMACTASKFILETRANSFER_READ:
1022	case PDMACTASKFILETRANSFER_WRITE:
1023	{
1024	RTFILEAIOREQ hReq = NIL_RTFILEAIOREQ;
1025
1026	if (pCurr->hReq == NIL_RTFILEAIOREQ)
1027	{
1028	if (pEndpoint->enmBackendType == PDMACFILEEPBACKEND_BUFFERED)
1029	rc = pdmacFileAioMgrNormalTaskPrepareBuffered(pAioMgr, pEndpoint, pCurr, &hReq);
1030	else if (pEndpoint->enmBackendType == PDMACFILEEPBACKEND_NON_BUFFERED)
1031	rc = pdmacFileAioMgrNormalTaskPrepareNonBuffered(pAioMgr, pEndpoint, pCurr, &hReq);
1032	else
1033	AssertMsgFailed(("Invalid backend type %d\n", pEndpoint->enmBackendType));
1034
1035	AssertRC(rc);
1036	}
1037	else
1038	{
1039	LogFlow(("Task %#p has I/O request %#p already\n", pCurr, pCurr->hReq));
1040	hReq = pCurr->hReq;
1041	}
1042
1043	LogFlow(("Read/Write request %#p\n", hReq));
1044
1045	if (hReq != NIL_RTFILEAIOREQ)
1046	{
1047	apReqs[cRequests] = hReq;
1048	cRequests++;
1049	}
1050	break;
1051	}
1052	default:
1053	AssertMsgFailed(("Invalid transfer type %d\n", pCurr->enmTransferType));
1054	} /* switch transfer type */
1055
1056	/* Queue the requests if the array is full. */
1057	if (cRequests == RT_ELEMENTS(apReqs))
1058	{
1059	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
1060	cRequests = 0;
1061	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
1062	("Unexpected return code\n"));
1063	}
1064	}
1065
1066	if (cRequests)
1067	{
1068	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
1069	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
1070	("Unexpected return code rc=%Rrc\n", rc));
1071	}
1072
1073	if (pTaskHead)
1074	{
1075	/* Add the rest of the tasks to the pending list */
1076	pdmacFileAioMgrEpAddTaskList(pEndpoint, pTaskHead);
1077
1078	if (RT_UNLIKELY( pAioMgr->cRequestsActiveMax == pAioMgr->cRequestsActive
1079	&& !pEndpoint->pFlushReq))
1080	{
1081	#if 0
1082	/*
1083	* The I/O manager has no room left for more requests
1084	* but there are still requests to process.
1085	* Create a new I/O manager and let it handle some endpoints.
1086	*/
1087	pdmacFileAioMgrNormalBalanceLoad(pAioMgr);
1088	#else
1089	/* Grow the I/O manager */
1090	pAioMgr->enmState = PDMACEPFILEMGRSTATE_GROWING;
1091	#endif
1092	}
1093	}
1094
1095	/* Insufficient resources are not fatal. */
1096	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
1097	rc = VINF_SUCCESS;
1098
1099	return rc;
1100	}
1101
1102	/**
1103	* Adds all pending requests for the given endpoint
1104	* until a flush request is encountered or there is no
1105	* request anymore.
1106	*
1107	* @returns VBox status code.
1108	* @param pAioMgr The async I/O manager for the endpoint
1109	* @param pEndpoint The endpoint to get the requests from.
1110	*/
1111	static int pdmacFileAioMgrNormalQueueReqs(PPDMACEPFILEMGR pAioMgr,
1112	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
1113	{
1114	int rc = VINF_SUCCESS;
1115	PPDMACTASKFILE pTasksHead = NULL;
1116
1117	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
1118	("Trying to process request lists of a non active endpoint!\n"));
1119
1120	Assert(!pEndpoint->pFlushReq);
1121
1122	/* Check the pending list first */
1123	if (pEndpoint->AioMgr.pReqsPendingHead)
1124	{
1125	LogFlow(("Queuing pending requests first\n"));
1126
1127	pTasksHead = pEndpoint->AioMgr.pReqsPendingHead;
1128	/*
1129	* Clear the list as the processing routine will insert them into the list
1130	* again if it gets a flush request.
1131	*/
1132	pEndpoint->AioMgr.pReqsPendingHead = NULL;
1133	pEndpoint->AioMgr.pReqsPendingTail = NULL;
1134	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
1135	AssertRC(rc); /** @todo r=bird: status code potentially overwritten. */
1136	}
1137
1138	if (!pEndpoint->pFlushReq && !pEndpoint->AioMgr.pReqsPendingHead)
1139	{
1140	/* Now the request queue. */
1141	pTasksHead = pdmacFileEpGetNewTasks(pEndpoint);
1142	if (pTasksHead)
1143	{
1144	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
1145	AssertRC(rc);
1146	}
1147	}
1148
1149	return rc;
1150	}
1151
1152	static int pdmacFileAioMgrNormalProcessBlockingEvent(PPDMACEPFILEMGR pAioMgr)
1153	{
1154	int rc = VINF_SUCCESS;
1155	bool fNotifyWaiter = false;
1156
1157	LogFlowFunc((": Enter\n"));
1158
1159	Assert(pAioMgr->fBlockingEventPending);
1160
1161	switch (pAioMgr->enmBlockingEvent)
1162	{
1163	case PDMACEPFILEAIOMGRBLOCKINGEVENT_ADD_ENDPOINT:
1164	{
1165	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointNew = ASMAtomicReadPtrT(&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint, PPDMASYNCCOMPLETIONENDPOINTFILE);
1166	AssertMsg(VALID_PTR(pEndpointNew), ("Adding endpoint event without a endpoint to add\n"));
1167
1168	pEndpointNew->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE;
1169
1170	pEndpointNew->AioMgr.pEndpointNext = pAioMgr->pEndpointsHead;
1171	pEndpointNew->AioMgr.pEndpointPrev = NULL;
1172	if (pAioMgr->pEndpointsHead)
1173	pAioMgr->pEndpointsHead->AioMgr.pEndpointPrev = pEndpointNew;
1174	pAioMgr->pEndpointsHead = pEndpointNew;
1175
1176	/* Assign the completion point to this file. */
1177	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pEndpointNew->hFile);
1178	fNotifyWaiter = true;
1179	pAioMgr->cEndpoints++;
1180	break;
1181	}
1182	case PDMACEPFILEAIOMGRBLOCKINGEVENT_REMOVE_ENDPOINT:
1183	{
1184	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove = ASMAtomicReadPtrT(&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint, PPDMASYNCCOMPLETIONENDPOINTFILE);
1185	AssertMsg(VALID_PTR(pEndpointRemove), ("Removing endpoint event without a endpoint to remove\n"));
1186
1187	pEndpointRemove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
1188	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointRemove);
1189	break;
1190	}
1191	case PDMACEPFILEAIOMGRBLOCKINGEVENT_CLOSE_ENDPOINT:
1192	{
1193	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointClose = ASMAtomicReadPtrT(&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint, PPDMASYNCCOMPLETIONENDPOINTFILE);
1194	AssertMsg(VALID_PTR(pEndpointClose), ("Close endpoint event without a endpoint to close\n"));
1195
1196	if (pEndpointClose->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
1197	{
1198	LogFlowFunc((": Closing endpoint %#p{%s}\n", pEndpointClose, pEndpointClose->Core.pszUri));
1199
1200	/* Make sure all tasks finished. Process the queues a last time first. */
1201	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpointClose);
1202	AssertRC(rc);
1203
1204	pEndpointClose->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING;
1205	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointClose);
1206	}
1207	else if ( (pEndpointClose->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING)
1208	&& (!pEndpointClose->AioMgr.cRequestsActive))
1209	fNotifyWaiter = true;
1210	break;
1211	}
1212	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SHUTDOWN:
1213	{
1214	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SHUTDOWN;
1215	if (!pAioMgr->cRequestsActive)
1216	fNotifyWaiter = true;
1217	break;
1218	}
1219	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SUSPEND:
1220	{
1221	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SUSPENDING;
1222	break;
1223	}
1224	case PDMACEPFILEAIOMGRBLOCKINGEVENT_RESUME:
1225	{
1226	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
1227	fNotifyWaiter = true;
1228	break;
1229	}
1230	default:
1231	AssertReleaseMsgFailed(("Invalid event type %d\n", pAioMgr->enmBlockingEvent));
1232	}
1233
1234	if (fNotifyWaiter)
1235	{
1236	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
1237	pAioMgr->enmBlockingEvent = PDMACEPFILEAIOMGRBLOCKINGEVENT_INVALID;
1238
1239	/* Release the waiting thread. */
1240	LogFlow(("Signalling waiter\n"));
1241	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
1242	AssertRC(rc);
1243	}
1244
1245	LogFlowFunc((": Leave\n"));
1246	return rc;
1247	}
1248
1249	/**
1250	* Checks all endpoints for pending events or new requests.
1251	*
1252	* @returns VBox status code.
1253	* @param pAioMgr The I/O manager handle.
1254	*/
1255	static int pdmacFileAioMgrNormalCheckEndpoints(PPDMACEPFILEMGR pAioMgr)
1256	{
1257	/* Check the assigned endpoints for new tasks if there isn't a flush request active at the moment. */
1258	int rc = VINF_SUCCESS;
1259	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint = pAioMgr->pEndpointsHead;
1260
1261	pAioMgr->msBwLimitExpired = RT_INDEFINITE_WAIT;
1262
1263	while (pEndpoint)
1264	{
1265	if (!pEndpoint->pFlushReq
1266	&& (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
1267	&& !pEndpoint->AioMgr.fMoving)
1268	{
1269	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
1270	if (RT_FAILURE(rc))
1271	return rc;
1272	}
1273	else if ( !pEndpoint->AioMgr.cRequestsActive
1274	&& pEndpoint->enmState != PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
1275	{
1276	/* Reopen the file so that the new endpoint can re-associate with the file */
1277	RTFileClose(pEndpoint->hFile);
1278	rc = RTFileOpen(&pEndpoint->hFile, pEndpoint->Core.pszUri, pEndpoint->fFlags);
1279	AssertRC(rc);
1280
1281	if (pEndpoint->AioMgr.fMoving)
1282	{
1283	pEndpoint->AioMgr.fMoving = false;
1284	pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
1285	}
1286	else
1287	{
1288	Assert(pAioMgr->fBlockingEventPending);
1289	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
1290
1291	/* Release the waiting thread. */
1292	LogFlow(("Signalling waiter\n"));
1293	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
1294	AssertRC(rc);
1295	}
1296	}
1297
1298	pEndpoint = pEndpoint->AioMgr.pEndpointNext;
1299	}
1300
1301	return rc;
1302	}
1303
1304	/**
1305	* Wrapper around pdmacFileAioMgrNormalReqCompleteRc().
1306	*/
1307	static void pdmacFileAioMgrNormalReqComplete(PPDMACEPFILEMGR pAioMgr, RTFILEAIOREQ hReq)
1308	{
1309	size_t cbTransfered = 0;
1310	int rcReq = RTFileAioReqGetRC(hReq, &cbTransfered);
1311
1312	pdmacFileAioMgrNormalReqCompleteRc(pAioMgr, hReq, rcReq, cbTransfered);
1313	}
1314
1315	static void pdmacFileAioMgrNormalReqCompleteRc(PPDMACEPFILEMGR pAioMgr, RTFILEAIOREQ hReq,
1316	int rcReq, size_t cbTransfered)
1317	{
1318	int rc = VINF_SUCCESS;
1319	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint;
1320	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(hReq);
1321	PPDMACTASKFILE pTasksWaiting;
1322
1323	LogFlowFunc(("pAioMgr=%#p hReq=%#p\n", pAioMgr, hReq));
1324
1325	pEndpoint = pTask->pEndpoint;
1326
1327	pTask->hReq = NIL_RTFILEAIOREQ;
1328
1329	pAioMgr->cRequestsActive--;
1330	pEndpoint->AioMgr.cRequestsActive--;
1331	pEndpoint->AioMgr.cReqsProcessed++;
1332
1333	/*
1334	* It is possible that the request failed on Linux with kernels < 2.6.23
1335	* if the passed buffer was allocated with remap_pfn_range or if the file
1336	* is on an NFS endpoint which does not support async and direct I/O at the same time.
1337	* The endpoint will be migrated to a failsafe manager in case a request fails.
1338	*/
1339	if (RT_FAILURE(rcReq))
1340	{
1341	/* Free bounce buffers and the IPRT request. */
1342	pdmacFileAioMgrNormalRequestFree(pAioMgr, hReq);
1343
1344	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_FLUSH)
1345	{
1346	LogFlow(("Async flushes are not supported for this endpoint, disabling\n"));
1347	pEndpoint->fAsyncFlushSupported = false;
1348	AssertMsg(pEndpoint->pFlushReq == pTask, ("Failed flush request doesn't match active one\n"));
1349	/* The other method will take over now. */
1350	}
1351	else
1352	{
1353	/* Free the lock and process pending tasks if necessary */
1354	pTasksWaiting = pdmacFileAioMgrNormalRangeLockFree(pAioMgr, pEndpoint, pTask->pRangeLock);
1355	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksWaiting, pAioMgr, pEndpoint);
1356	AssertRC(rc);
1357
1358	if (pTask->cbBounceBuffer)
1359	RTMemPageFree(pTask->pvBounceBuffer, pTask->cbBounceBuffer);
1360
1361	/*
1362	* Fatal errors are reported to the guest and non-fatal errors
1363	* will cause a migration to the failsafe manager in the hope
1364	* that the error disappears.
1365	*/
1366	if (!pdmacFileAioMgrNormalRcIsFatal(rcReq))
1367	{
1368	/* Queue the request on the pending list. */
1369	pTask->pNext = pEndpoint->AioMgr.pReqsPendingHead;
1370	pEndpoint->AioMgr.pReqsPendingHead = pTask;
1371
1372	/* Create a new failsafe manager if necessary. */
1373	if (!pEndpoint->AioMgr.fMoving)
1374	{
1375	PPDMACEPFILEMGR pAioMgrFailsafe;
1376
1377	LogRel(("%s: Request %#p failed with rc=%Rrc, migrating endpoint %s to failsafe manager.\n",
1378	RTThreadGetName(pAioMgr->Thread), pTask, rcReq, pEndpoint->Core.pszUri));
1379
1380	pEndpoint->AioMgr.fMoving = true;
1381
1382	rc = pdmacFileAioMgrCreate((PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass,
1383	&pAioMgrFailsafe, PDMACEPFILEMGRTYPE_SIMPLE);
1384	AssertRC(rc);
1385
1386	pEndpoint->AioMgr.pAioMgrDst = pAioMgrFailsafe;
1387
1388	/* Update the flags to open the file with. Disable async I/O and enable the host cache. */
1389	pEndpoint->fFlags &= ~(RTFILE_O_ASYNC_IO \| RTFILE_O_NO_CACHE);
1390	}
1391
1392	/* If this was the last request for the endpoint migrate it to the new manager. */
1393	if (!pEndpoint->AioMgr.cRequestsActive)
1394	{
1395	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pEndpoint);
1396	Assert(!fReqsPending);
1397
1398	rc = pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
1399	AssertRC(rc);
1400	}
1401	}
1402	else
1403	{
1404	pTask->pfnCompleted(pTask, pTask->pvUser, rcReq);
1405	pdmacFileTaskFree(pEndpoint, pTask);
1406	}
1407	}
1408	}
1409	else
1410	{
1411	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_FLUSH)
1412	{
1413	/* Clear pending flush */
1414	AssertMsg(pEndpoint->pFlushReq == pTask, ("Completed flush request doesn't match active one\n"));
1415	pEndpoint->pFlushReq = NULL;
1416	pdmacFileAioMgrNormalRequestFree(pAioMgr, hReq);
1417
1418	/* Call completion callback */
1419	LogFlow(("Flush task=%#p completed with %Rrc\n", pTask, rcReq));
1420	pTask->pfnCompleted(pTask, pTask->pvUser, rcReq);
1421	pdmacFileTaskFree(pEndpoint, pTask);
1422	}
1423	else
1424	{
1425	/*
1426	* Restart an incomplete transfer.
1427	* This usually means that the request will return an error now
1428	* but to get the cause of the error (disk full, file too big, I/O error, ...)
1429	* the transfer needs to be continued.
1430	*/
1431	if (RT_UNLIKELY( cbTransfered < pTask->DataSeg.cbSeg
1432	\|\| ( pTask->cbBounceBuffer
1433	&& cbTransfered < pTask->cbBounceBuffer)))
1434	{
1435	RTFOFF offStart;
1436	size_t cbToTransfer;
1437	uint8_t *pbBuf = NULL;
1438
1439	LogFlow(("Restarting incomplete transfer %#p (%zu bytes transferred)\n",
1440	pTask, cbTransfered));
1441	Assert(cbTransfered % 512 == 0);
1442
1443	if (pTask->cbBounceBuffer)
1444	{
1445	AssertPtr(pTask->pvBounceBuffer);
1446	offStart = (pTask->Off & ~((RTFOFF)512-1)) + cbTransfered;
1447	cbToTransfer = pTask->cbBounceBuffer - cbTransfered;
1448	pbBuf = (uint8_t *)pTask->pvBounceBuffer + cbTransfered;
1449	}
1450	else
1451	{
1452	Assert(!pTask->pvBounceBuffer);
1453	offStart = pTask->Off + cbTransfered;
1454	cbToTransfer = pTask->DataSeg.cbSeg - cbTransfered;
1455	pbBuf = (uint8_t *)pTask->DataSeg.pvSeg + cbTransfered;
1456	}
1457
1458	if (pTask->fPrefetch \|\| pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
1459	{
1460	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->hFile, offStart,
1461	pbBuf, cbToTransfer, pTask);
1462	}
1463	else
1464	{
1465	AssertMsg(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE,
1466	("Invalid transfer type\n"));
1467	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->hFile, offStart,
1468	pbBuf, cbToTransfer, pTask);
1469	}
1470	AssertRC(rc);
1471
1472	pTask->hReq = hReq;
1473	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, &hReq, 1);
1474	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
1475	("Unexpected return code rc=%Rrc\n", rc));
1476	}
1477	else if (pTask->fPrefetch)
1478	{
1479	Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE);
1480	Assert(pTask->cbBounceBuffer);
1481
1482	memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->offBounceBuffer,
1483	pTask->DataSeg.pvSeg,
1484	pTask->DataSeg.cbSeg);
1485
1486	/* Write it now. */
1487	pTask->fPrefetch = false;
1488	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg, 512);
1489	RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1);
1490
1491	/* Grow the file if needed. */
1492	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
1493	{
1494	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
1495	RTFileSetSize(pEndpoint->hFile, pTask->Off + pTask->DataSeg.cbSeg);
1496	}
1497
1498	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->hFile,
1499	offStart, pTask->pvBounceBuffer, cbToTransfer, pTask);
1500	AssertRC(rc);
1501	pTask->hReq = hReq;
1502	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, &hReq, 1);
1503	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
1504	("Unexpected return code rc=%Rrc\n", rc));
1505	}
1506	else
1507	{
1508	if (RT_SUCCESS(rc) && pTask->cbBounceBuffer)
1509	{
1510	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
1511	memcpy(pTask->DataSeg.pvSeg,
1512	((uint8_t *)pTask->pvBounceBuffer) + pTask->offBounceBuffer,
1513	pTask->DataSeg.cbSeg);
1514
1515	RTMemPageFree(pTask->pvBounceBuffer, pTask->cbBounceBuffer);
1516	}
1517
1518	pdmacFileAioMgrNormalRequestFree(pAioMgr, hReq);
1519
1520	/* Free the lock and process pending tasks if necessary */
1521	pTasksWaiting = pdmacFileAioMgrNormalRangeLockFree(pAioMgr, pEndpoint, pTask->pRangeLock);
1522	if (pTasksWaiting)
1523	{
1524	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksWaiting, pAioMgr, pEndpoint);
1525	AssertRC(rc);
1526	}
1527
1528	/* Call completion callback */
1529	LogFlow(("Task=%#p completed with %Rrc\n", pTask, rcReq));
1530	pTask->pfnCompleted(pTask, pTask->pvUser, rcReq);
1531	pdmacFileTaskFree(pEndpoint, pTask);
1532
1533	/*
1534	* If there is no request left on the endpoint but a flush request is set
1535	* it completed now and we notify the owner.
1536	* Furthermore we look for new requests and continue.
1537	*/
1538	if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq)
1539	{
1540	/* Call completion callback */
1541	pTask = pEndpoint->pFlushReq;
1542	pEndpoint->pFlushReq = NULL;
1543
1544	AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n"));
1545
1546	pTask->pfnCompleted(pTask, pTask->pvUser, VINF_SUCCESS);
1547	pdmacFileTaskFree(pEndpoint, pTask);
1548	}
1549	else if (RT_UNLIKELY(!pEndpoint->AioMgr.cRequestsActive && pEndpoint->AioMgr.fMoving))
1550	{
1551	/* If the endpoint is about to be migrated do it now. */
1552	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pEndpoint);
1553	Assert(!fReqsPending);
1554
1555	rc = pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
1556	AssertRC(rc);
1557	}
1558	}
1559	} /* Not a flush request */
1560	} /* request completed successfully */
1561	}
1562
1563	/** Helper macro for checking for error codes. */
1564	#define CHECK_RC(pAioMgr, rc) \
1565	if (RT_FAILURE(rc)) \
1566	{\
1567	int rc2 = pdmacFileAioMgrNormalErrorHandler(pAioMgr, rc, RT_SRC_POS);\
1568	return rc2;\
1569	}
1570
1571	/**
1572	* The normal I/O manager using the RTFileAio* API
1573	*
1574	* @returns VBox status code.
1575	* @param hThreadSelf Handle of the thread.
1576	* @param pvUser Opaque user data.
1577	*/
1578	DECLCALLBACK(int) pdmacFileAioMgrNormal(RTTHREAD hThreadSelf, void *pvUser)
1579	{
1580	int rc = VINF_SUCCESS;
1581	PPDMACEPFILEMGR pAioMgr = (PPDMACEPFILEMGR)pvUser;
1582	uint64_t uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
1583	NOREF(hThreadSelf);
1584
1585	while ( pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING
1586	\|\| pAioMgr->enmState == PDMACEPFILEMGRSTATE_SUSPENDING
1587	\|\| pAioMgr->enmState == PDMACEPFILEMGRSTATE_GROWING)
1588	{
1589	if (!pAioMgr->cRequestsActive)
1590	{
1591	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, true);
1592	if (!ASMAtomicReadBool(&pAioMgr->fWokenUp))
1593	rc = RTSemEventWait(pAioMgr->EventSem, pAioMgr->msBwLimitExpired);
1594	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, false);
1595	Assert(RT_SUCCESS(rc) \|\| rc == VERR_TIMEOUT);
1596
1597	LogFlow(("Got woken up\n"));
1598	ASMAtomicWriteBool(&pAioMgr->fWokenUp, false);
1599	}
1600
1601	/* Check for an external blocking event first. */
1602	if (pAioMgr->fBlockingEventPending)
1603	{
1604	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
1605	CHECK_RC(pAioMgr, rc);
1606	}
1607
1608	if (RT_LIKELY( pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING
1609	\|\| pAioMgr->enmState == PDMACEPFILEMGRSTATE_GROWING))
1610	{
1611	/* We got woken up because an endpoint issued new requests. Queue them. */
1612	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
1613	CHECK_RC(pAioMgr, rc);
1614
1615	while (pAioMgr->cRequestsActive)
1616	{
1617	RTFILEAIOREQ apReqs[20];
1618	uint32_t cReqsCompleted = 0;
1619	size_t cReqsWait;
1620
1621	if (pAioMgr->cRequestsActive > RT_ELEMENTS(apReqs))
1622	cReqsWait = RT_ELEMENTS(apReqs);
1623	else
1624	cReqsWait = pAioMgr->cRequestsActive;
1625
1626	LogFlow(("Waiting for %d of %d tasks to complete\n", 1, cReqsWait));
1627
1628	rc = RTFileAioCtxWait(pAioMgr->hAioCtx,
1629	1,
1630	RT_INDEFINITE_WAIT, apReqs,
1631	cReqsWait, &cReqsCompleted);
1632	if (RT_FAILURE(rc) && (rc != VERR_INTERRUPTED))
1633	CHECK_RC(pAioMgr, rc);
1634
1635	LogFlow(("%d tasks completed\n", cReqsCompleted));
1636
1637	for (uint32_t i = 0; i < cReqsCompleted; i++)
1638	pdmacFileAioMgrNormalReqComplete(pAioMgr, apReqs[i]);
1639
1640	/* Check for an external blocking event before we go to sleep again. */
1641	if (pAioMgr->fBlockingEventPending)
1642	{
1643	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
1644	CHECK_RC(pAioMgr, rc);
1645	}
1646
1647	/* Update load statistics. */
1648	uint64_t uMillisCurr = RTTimeMilliTS();
1649	if (uMillisCurr > uMillisEnd)
1650	{
1651	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointCurr = pAioMgr->pEndpointsHead;
1652
1653	/* Calculate timespan. */
1654	uMillisCurr -= uMillisEnd;
1655
1656	while (pEndpointCurr)
1657	{
1658	pEndpointCurr->AioMgr.cReqsPerSec = pEndpointCurr->AioMgr.cReqsProcessed / (uMillisCurr + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD);
1659	pEndpointCurr->AioMgr.cReqsProcessed = 0;
1660	pEndpointCurr = pEndpointCurr->AioMgr.pEndpointNext;
1661	}
1662
1663	/* Set new update interval */
1664	uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
1665	}
1666
1667	/* Check endpoints for new requests. */
1668	if (pAioMgr->enmState != PDMACEPFILEMGRSTATE_GROWING)
1669	{
1670	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
1671	CHECK_RC(pAioMgr, rc);
1672	}
1673	} /* while requests are active. */
1674
1675	if (pAioMgr->enmState == PDMACEPFILEMGRSTATE_GROWING)
1676	{
1677	rc = pdmacFileAioMgrNormalGrow(pAioMgr);
1678	AssertRC(rc);
1679	Assert(pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING);
1680
1681	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
1682	CHECK_RC(pAioMgr, rc);
1683	}
1684	} /* if still running */
1685	} /* while running */
1686
1687	LogFlowFunc(("rc=%Rrc\n", rc));
1688	return rc;
1689	}
1690
1691	#undef CHECK_RC
1692

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PDMAsyncCompletionFileNormal.cpp@ 42062

Download in other formats: