PDMAsyncCompletionFileNormal.cpp@ 96854

Last change on this file since 96854 was 96407, checked in by vboxsync, 2 years ago
scm copyright and license note update
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 67.1 KB

Line
1	/* $Id: PDMAsyncCompletionFileNormal.cpp 96407 2022-08-22 17:43:14Z vboxsync $ */
2	/** @file
3	* PDM Async I/O - Async File I/O manager.
4	*/
5
6	/*
7	* Copyright (C) 2006-2022 Oracle and/or its affiliates.
8	*
9	* This file is part of VirtualBox base platform packages, as
10	* available from https://www.virtualbox.org.
11	*
12	* This program is free software; you can redistribute it and/or
13	* modify it under the terms of the GNU General Public License
14	* as published by the Free Software Foundation, in version 3 of the
15	* License.
16	*
17	* This program is distributed in the hope that it will be useful, but
18	* WITHOUT ANY WARRANTY; without even the implied warranty of
19	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20	* General Public License for more details.
21	*
22	* You should have received a copy of the GNU General Public License
23	* along with this program; if not, see <https://www.gnu.org/licenses>.
24	*
25	* SPDX-License-Identifier: GPL-3.0-only
26	*/
27
28
29	/*********************************************************************************************************************************
30	* Header Files *
31	*********************************************************************************************************************************/
32	#define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION
33	#include <iprt/types.h>
34	#include <iprt/asm.h>
35	#include <iprt/file.h>
36	#include <iprt/mem.h>
37	#include <iprt/string.h>
38	#include <iprt/assert.h>
39	#include <VBox/log.h>
40
41	#include "PDMAsyncCompletionFileInternal.h"
42
43
44	/*********************************************************************************************************************************
45	* Defined Constants And Macros *
46	*********************************************************************************************************************************/
47	/** The update period for the I/O load statistics in ms. */
48	#define PDMACEPFILEMGR_LOAD_UPDATE_PERIOD 1000
49	/** Maximum number of requests a manager will handle. */
50	#define PDMACEPFILEMGR_REQS_STEP 64
51
52
53	/*********************************************************************************************************************************
54	* Internal functions *
55	*********************************************************************************************************************************/
56	static int pdmacFileAioMgrNormalProcessTaskList(PPDMACTASKFILE pTaskHead,
57	PPDMACEPFILEMGR pAioMgr,
58	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint);
59
60	static PPDMACTASKFILE pdmacFileAioMgrNormalRangeLockFree(PPDMACEPFILEMGR pAioMgr,
61	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
62	PPDMACFILERANGELOCK pRangeLock);
63
64	static void pdmacFileAioMgrNormalReqCompleteRc(PPDMACEPFILEMGR pAioMgr, RTFILEAIOREQ hReq,
65	int rc, size_t cbTransfered);
66
67
68	int pdmacFileAioMgrNormalInit(PPDMACEPFILEMGR pAioMgr)
69	{
70	pAioMgr->cRequestsActiveMax = PDMACEPFILEMGR_REQS_STEP;
71
72	int rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, RTFILEAIO_UNLIMITED_REQS, 0 /* fFlags */);
73	if (rc == VERR_OUT_OF_RANGE)
74	rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, pAioMgr->cRequestsActiveMax, 0 /* fFlags */);
75
76	if (RT_SUCCESS(rc))
77	{
78	/* Initialize request handle array. */
79	pAioMgr->iFreeEntry = 0;
80	pAioMgr->cReqEntries = pAioMgr->cRequestsActiveMax;
81	pAioMgr->pahReqsFree = (RTFILEAIOREQ )RTMemAllocZ(pAioMgr->cReqEntries sizeof(RTFILEAIOREQ));
82
83	if (pAioMgr->pahReqsFree)
84	{
85	/* Create the range lock memcache. */
86	rc = RTMemCacheCreate(&pAioMgr->hMemCacheRangeLocks, sizeof(PDMACFILERANGELOCK),
87	0, UINT32_MAX, NULL, NULL, NULL, 0);
88	if (RT_SUCCESS(rc))
89	return VINF_SUCCESS;
90
91	RTMemFree(pAioMgr->pahReqsFree);
92	}
93	else
94	{
95	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
96	rc = VERR_NO_MEMORY;
97	}
98	}
99
100	return rc;
101	}
102
103	void pdmacFileAioMgrNormalDestroy(PPDMACEPFILEMGR pAioMgr)
104	{
105	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
106
107	while (pAioMgr->iFreeEntry > 0)
108	{
109	pAioMgr->iFreeEntry--;
110	Assert(pAioMgr->pahReqsFree[pAioMgr->iFreeEntry] != NIL_RTFILEAIOREQ);
111	RTFileAioReqDestroy(pAioMgr->pahReqsFree[pAioMgr->iFreeEntry]);
112	}
113
114	RTMemFree(pAioMgr->pahReqsFree);
115	RTMemCacheDestroy(pAioMgr->hMemCacheRangeLocks);
116	}
117
118	#if 0 /* currently unused */
119	/**
120	* Sorts the endpoint list with insertion sort.
121	*/
122	static void pdmacFileAioMgrNormalEndpointsSortByLoad(PPDMACEPFILEMGR pAioMgr)
123	{
124	PPDMASYNCCOMPLETIONENDPOINTFILE pEpPrev, pEpCurr, pEpNextToSort;
125
126	pEpPrev = pAioMgr->pEndpointsHead;
127	pEpCurr = pEpPrev->AioMgr.pEndpointNext;
128
129	while (pEpCurr)
130	{
131	/* Remember the next element to sort because the list might change. */
132	pEpNextToSort = pEpCurr->AioMgr.pEndpointNext;
133
134	/* Unlink the current element from the list. */
135	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
136	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
137
138	if (pPrev)
139	pPrev->AioMgr.pEndpointNext = pNext;
140	else
141	pAioMgr->pEndpointsHead = pNext;
142
143	if (pNext)
144	pNext->AioMgr.pEndpointPrev = pPrev;
145
146	/* Go back until we reached the place to insert the current endpoint into. */
147	while (pEpPrev && (pEpPrev->AioMgr.cReqsPerSec < pEpCurr->AioMgr.cReqsPerSec))
148	pEpPrev = pEpPrev->AioMgr.pEndpointPrev;
149
150	/* Link the endpoint into the list. */
151	if (pEpPrev)
152	pNext = pEpPrev->AioMgr.pEndpointNext;
153	else
154	pNext = pAioMgr->pEndpointsHead;
155
156	pEpCurr->AioMgr.pEndpointNext = pNext;
157	pEpCurr->AioMgr.pEndpointPrev = pEpPrev;
158
159	if (pNext)
160	pNext->AioMgr.pEndpointPrev = pEpCurr;
161
162	if (pEpPrev)
163	pEpPrev->AioMgr.pEndpointNext = pEpCurr;
164	else
165	pAioMgr->pEndpointsHead = pEpCurr;
166
167	pEpCurr = pEpNextToSort;
168	}
169
170	#ifdef DEBUG
171	/* Validate sorting algorithm */
172	unsigned cEndpoints = 0;
173	pEpCurr = pAioMgr->pEndpointsHead;
174
175	AssertMsg(pEpCurr, ("No endpoint in the list?\n"));
176	AssertMsg(!pEpCurr->AioMgr.pEndpointPrev, ("First element in the list points to previous element\n"));
177
178	while (pEpCurr)
179	{
180	cEndpoints++;
181
182	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
183	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
184
185	Assert(!pNext \|\| pNext->AioMgr.cReqsPerSec <= pEpCurr->AioMgr.cReqsPerSec);
186	Assert(!pPrev \|\| pPrev->AioMgr.cReqsPerSec >= pEpCurr->AioMgr.cReqsPerSec);
187
188	pEpCurr = pNext;
189	}
190
191	AssertMsg(cEndpoints == pAioMgr->cEndpoints, ("Endpoints lost during sort!\n"));
192
193	#endif
194	}
195	#endif /* currently unused */
196
197	/**
198	* Removes an endpoint from the currently assigned manager.
199	*
200	* @returns TRUE if there are still requests pending on the current manager for this endpoint.
201	* FALSE otherwise.
202	* @param pEndpointRemove The endpoint to remove.
203	*/
204	static bool pdmacFileAioMgrNormalRemoveEndpoint(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove)
205	{
206	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEndpointRemove->AioMgr.pEndpointPrev;
207	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEndpointRemove->AioMgr.pEndpointNext;
208	PPDMACEPFILEMGR pAioMgr = pEndpointRemove->pAioMgr;
209
210	pAioMgr->cEndpoints--;
211
212	if (pPrev)
213	pPrev->AioMgr.pEndpointNext = pNext;
214	else
215	pAioMgr->pEndpointsHead = pNext;
216
217	if (pNext)
218	pNext->AioMgr.pEndpointPrev = pPrev;
219
220	/* Make sure that there is no request pending on this manager for the endpoint. */
221	if (!pEndpointRemove->AioMgr.cRequestsActive)
222	{
223	Assert(!pEndpointRemove->pFlushReq);
224
225	/* Reopen the file so that the new endpoint can re-associate with the file */
226	RTFileClose(pEndpointRemove->hFile);
227	int rc = RTFileOpen(&pEndpointRemove->hFile, pEndpointRemove->Core.pszUri, pEndpointRemove->fFlags);
228	AssertRC(rc);
229	return false;
230	}
231
232	return true;
233	}
234
235	#if 0 /* currently unused */
236
237	static bool pdmacFileAioMgrNormalIsBalancePossible(PPDMACEPFILEMGR pAioMgr)
238	{
239	/* Balancing doesn't make sense with only one endpoint. */
240	if (pAioMgr->cEndpoints == 1)
241	return false;
242
243	/* Doesn't make sens to move endpoints if only one produces the whole load */
244	unsigned cEndpointsWithLoad = 0;
245
246	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead;
247
248	while (pCurr)
249	{
250	if (pCurr->AioMgr.cReqsPerSec)
251	cEndpointsWithLoad++;
252
253	pCurr = pCurr->AioMgr.pEndpointNext;
254	}
255
256	return (cEndpointsWithLoad > 1);
257	}
258
259	/**
260	* Creates a new I/O manager and spreads the I/O load of the endpoints
261	* between the given I/O manager and the new one.
262	*
263	* @returns nothing.
264	* @param pAioMgr The I/O manager with high I/O load.
265	*/
266	static void pdmacFileAioMgrNormalBalanceLoad(PPDMACEPFILEMGR pAioMgr)
267	{
268	/*
269	* Check if balancing would improve the situation.
270	*/
271	if (pdmacFileAioMgrNormalIsBalancePossible(pAioMgr))
272	{
273	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass;
274	PPDMACEPFILEMGR pAioMgrNew = NULL;
275
276	int rc = pdmacFileAioMgrCreate(pEpClassFile, &pAioMgrNew, PDMACEPFILEMGRTYPE_ASYNC);
277	if (RT_SUCCESS(rc))
278	{
279	/* We will sort the list by request count per second. */
280	pdmacFileAioMgrNormalEndpointsSortByLoad(pAioMgr);
281
282	/* Now move some endpoints to the new manager. */
283	unsigned cReqsHere = pAioMgr->pEndpointsHead->AioMgr.cReqsPerSec;
284	unsigned cReqsOther = 0;
285	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead->AioMgr.pEndpointNext;
286
287	while (pCurr)
288	{
289	if (cReqsHere <= cReqsOther)
290	{
291	/*
292	* The other manager has more requests to handle now.
293	* We will keep the current endpoint.
294	*/
295	Log(("Keeping endpoint %#p{%s} with %u reqs/s\n", pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
296	cReqsHere += pCurr->AioMgr.cReqsPerSec;
297	pCurr = pCurr->AioMgr.pEndpointNext;
298	}
299	else
300	{
301	/* Move to other endpoint. */
302	Log(("Moving endpoint %#p{%s} with %u reqs/s to other manager\n", pCurr, pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
303	cReqsOther += pCurr->AioMgr.cReqsPerSec;
304
305	PPDMASYNCCOMPLETIONENDPOINTFILE pMove = pCurr;
306
307	pCurr = pCurr->AioMgr.pEndpointNext;
308
309	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pMove);
310
311	if (fReqsPending)
312	{
313	pMove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
314	pMove->AioMgr.fMoving = true;
315	pMove->AioMgr.pAioMgrDst = pAioMgrNew;
316	}
317	else
318	{
319	pMove->AioMgr.fMoving = false;
320	pMove->AioMgr.pAioMgrDst = NULL;
321	pdmacFileAioMgrAddEndpoint(pAioMgrNew, pMove);
322	}
323	}
324	}
325	}
326	else
327	{
328	/* Don't process further but leave a log entry about reduced performance. */
329	LogRel(("AIOMgr: Could not create new I/O manager (rc=%Rrc). Expect reduced performance\n", rc));
330	}
331	}
332	else
333	Log(("AIOMgr: Load balancing would not improve anything\n"));
334	}
335
336	#endif /* unused */
337
338	/**
339	* Increase the maximum number of active requests for the given I/O manager.
340	*
341	* @returns VBox status code.
342	* @param pAioMgr The I/O manager to grow.
343	*/
344	static int pdmacFileAioMgrNormalGrow(PPDMACEPFILEMGR pAioMgr)
345	{
346	LogFlowFunc(("pAioMgr=%#p\n", pAioMgr));
347
348	AssertMsg( pAioMgr->enmState == PDMACEPFILEMGRSTATE_GROWING
349	&& !pAioMgr->cRequestsActive,
350	("Invalid state of the I/O manager\n"));
351
352	#ifdef RT_OS_WINDOWS
353	/*
354	* Reopen the files of all assigned endpoints first so we can assign them to the new
355	* I/O context.
356	*/
357	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead;
358
359	while (pCurr)
360	{
361	RTFileClose(pCurr->hFile);
362	int rc2 = RTFileOpen(&pCurr->hFile, pCurr->Core.pszUri, pCurr->fFlags); AssertRC(rc2);
363
364	pCurr = pCurr->AioMgr.pEndpointNext;
365	}
366	#endif
367
368	/* Create the new bigger context. */
369	pAioMgr->cRequestsActiveMax += PDMACEPFILEMGR_REQS_STEP;
370
371	RTFILEAIOCTX hAioCtxNew = NIL_RTFILEAIOCTX;
372	int rc = RTFileAioCtxCreate(&hAioCtxNew, RTFILEAIO_UNLIMITED_REQS, 0 /* fFlags */);
373	if (rc == VERR_OUT_OF_RANGE)
374	rc = RTFileAioCtxCreate(&hAioCtxNew, pAioMgr->cRequestsActiveMax, 0 /* fFlags */);
375
376	if (RT_SUCCESS(rc))
377	{
378	/* Close the old context. */
379	rc = RTFileAioCtxDestroy(pAioMgr->hAioCtx);
380	AssertRC(rc); /** @todo r=bird: Ignoring error code, will propagate. */
381
382	pAioMgr->hAioCtx = hAioCtxNew;
383
384	/* Create a new I/O task handle array */
385	uint32_t cReqEntriesNew = pAioMgr->cRequestsActiveMax + 1;
386	RTFILEAIOREQ pahReqNew = (RTFILEAIOREQ )RTMemAllocZ(cReqEntriesNew * sizeof(RTFILEAIOREQ));
387
388	if (pahReqNew)
389	{
390	/* Copy the cached request handles. */
391	for (uint32_t iReq = 0; iReq < pAioMgr->cReqEntries; iReq++)
392	pahReqNew[iReq] = pAioMgr->pahReqsFree[iReq];
393
394	RTMemFree(pAioMgr->pahReqsFree);
395	pAioMgr->pahReqsFree = pahReqNew;
396	pAioMgr->cReqEntries = cReqEntriesNew;
397	LogFlowFunc(("I/O manager increased to handle a maximum of %u requests\n",
398	pAioMgr->cRequestsActiveMax));
399	}
400	else
401	rc = VERR_NO_MEMORY;
402	}
403
404	#ifdef RT_OS_WINDOWS
405	/* Assign the file to the new context. */
406	pCurr = pAioMgr->pEndpointsHead;
407	while (pCurr)
408	{
409	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pCurr->hFile);
410	AssertRC(rc); /** @todo r=bird: Ignoring error code, will propagate. */
411
412	pCurr = pCurr->AioMgr.pEndpointNext;
413	}
414	#endif
415
416	if (RT_FAILURE(rc))
417	{
418	LogFlow(("Increasing size of the I/O manager failed with rc=%Rrc\n", rc));
419	pAioMgr->cRequestsActiveMax -= PDMACEPFILEMGR_REQS_STEP;
420	}
421
422	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
423	LogFlowFunc(("returns rc=%Rrc\n", rc));
424
425	return rc;
426	}
427
428	/**
429	* Checks if a given status code is fatal.
430	* Non fatal errors can be fixed by migrating the endpoint to a
431	* failsafe manager.
432	*
433	* @returns true If the error is fatal and migrating to a failsafe manager doesn't help
434	* false If the error can be fixed by a migration. (image on NFS disk for example)
435	* @param rcReq The status code to check.
436	*/
437	DECLINLINE(bool) pdmacFileAioMgrNormalRcIsFatal(int rcReq)
438	{
439	return rcReq == VERR_DEV_IO_ERROR
440	\|\| rcReq == VERR_FILE_IO_ERROR
441	\|\| rcReq == VERR_DISK_IO_ERROR
442	\|\| rcReq == VERR_DISK_FULL
443	\|\| rcReq == VERR_FILE_TOO_BIG;
444	}
445
446	/**
447	* Error handler which will create the failsafe managers and destroy the failed I/O manager.
448	*
449	* @returns VBox status code
450	* @param pAioMgr The I/O manager the error occurred on.
451	* @param rc The error code.
452	* @param SRC_POS The source location of the error (use RT_SRC_POS).
453	*/
454	static int pdmacFileAioMgrNormalErrorHandler(PPDMACEPFILEMGR pAioMgr, int rc, RT_SRC_POS_DECL)
455	{
456	LogRel(("AIOMgr: I/O manager %#p encountered a critical error (rc=%Rrc) during operation. Falling back to failsafe mode. Expect reduced performance\n",
457	pAioMgr, rc));
458	LogRel(("AIOMgr: Error happened in %s:(%u){%s}\n", RT_SRC_POS_ARGS));
459	LogRel(("AIOMgr: Please contact the product vendor\n"));
460
461	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass;
462
463	pAioMgr->enmState = PDMACEPFILEMGRSTATE_FAULT;
464	ASMAtomicWriteU32((volatile uint32_t *)&pEpClassFile->enmMgrTypeOverride, PDMACEPFILEMGRTYPE_SIMPLE);
465
466	AssertMsgFailed(("Implement\n"));
467	return VINF_SUCCESS;
468	}
469
470	/**
471	* Put a list of tasks in the pending request list of an endpoint.
472	*/
473	DECLINLINE(void) pdmacFileAioMgrEpAddTaskList(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTaskHead)
474	{
475	/* Add the rest of the tasks to the pending list */
476	if (!pEndpoint->AioMgr.pReqsPendingHead)
477	{
478	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
479	pEndpoint->AioMgr.pReqsPendingHead = pTaskHead;
480	}
481	else
482	{
483	Assert(pEndpoint->AioMgr.pReqsPendingTail);
484	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTaskHead;
485	}
486
487	/* Update the tail. */
488	while (pTaskHead->pNext)
489	pTaskHead = pTaskHead->pNext;
490
491	pEndpoint->AioMgr.pReqsPendingTail = pTaskHead;
492	pTaskHead->pNext = NULL;
493	}
494
495	/**
496	* Put one task in the pending request list of an endpoint.
497	*/
498	DECLINLINE(void) pdmacFileAioMgrEpAddTask(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTask)
499	{
500	/* Add the rest of the tasks to the pending list */
501	if (!pEndpoint->AioMgr.pReqsPendingHead)
502	{
503	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
504	pEndpoint->AioMgr.pReqsPendingHead = pTask;
505	}
506	else
507	{
508	Assert(pEndpoint->AioMgr.pReqsPendingTail);
509	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTask;
510	}
511
512	pEndpoint->AioMgr.pReqsPendingTail = pTask;
513	pTask->pNext = NULL;
514	}
515
516	/**
517	* Allocates a async I/O request.
518	*
519	* @returns Handle to the request.
520	* @param pAioMgr The I/O manager.
521	*/
522	static RTFILEAIOREQ pdmacFileAioMgrNormalRequestAlloc(PPDMACEPFILEMGR pAioMgr)
523	{
524	/* Get a request handle. */
525	RTFILEAIOREQ hReq;
526	if (pAioMgr->iFreeEntry > 0)
527	{
528	pAioMgr->iFreeEntry--;
529	hReq = pAioMgr->pahReqsFree[pAioMgr->iFreeEntry];
530	pAioMgr->pahReqsFree[pAioMgr->iFreeEntry] = NIL_RTFILEAIOREQ;
531	Assert(hReq != NIL_RTFILEAIOREQ);
532	}
533	else
534	{
535	int rc = RTFileAioReqCreate(&hReq);
536	AssertRCReturn(rc, NIL_RTFILEAIOREQ);
537	}
538
539	return hReq;
540	}
541
542	/**
543	* Frees a async I/O request handle.
544	*
545	* @returns nothing.
546	* @param pAioMgr The I/O manager.
547	* @param hReq The I/O request handle to free.
548	*/
549	static void pdmacFileAioMgrNormalRequestFree(PPDMACEPFILEMGR pAioMgr, RTFILEAIOREQ hReq)
550	{
551	Assert(pAioMgr->iFreeEntry < pAioMgr->cReqEntries);
552	Assert(pAioMgr->pahReqsFree[pAioMgr->iFreeEntry] == NIL_RTFILEAIOREQ);
553
554	pAioMgr->pahReqsFree[pAioMgr->iFreeEntry] = hReq;
555	pAioMgr->iFreeEntry++;
556	}
557
558	/**
559	* Wrapper around RTFIleAioCtxSubmit() which is also doing error handling.
560	*/
561	static int pdmacFileAioMgrNormalReqsEnqueue(PPDMACEPFILEMGR pAioMgr,
562	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
563	PRTFILEAIOREQ pahReqs, unsigned cReqs)
564	{
565	pAioMgr->cRequestsActive += cReqs;
566	pEndpoint->AioMgr.cRequestsActive += cReqs;
567
568	LogFlow(("Enqueuing %d requests. I/O manager has a total of %d active requests now\n", cReqs, pAioMgr->cRequestsActive));
569	LogFlow(("Endpoint has a total of %d active requests now\n", pEndpoint->AioMgr.cRequestsActive));
570
571	int rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, pahReqs, cReqs);
572	if (RT_FAILURE(rc))
573	{
574	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
575	{
576	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
577
578	/* Append any not submitted task to the waiting list. */
579	for (size_t i = 0; i < cReqs; i++)
580	{
581	int rcReq = RTFileAioReqGetRC(pahReqs[i], NULL);
582
583	if (rcReq != VERR_FILE_AIO_IN_PROGRESS)
584	{
585	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(pahReqs[i]);
586
587	Assert(pTask->hReq == pahReqs[i]);
588	pdmacFileAioMgrEpAddTask(pEndpoint, pTask);
589	pAioMgr->cRequestsActive--;
590	pEndpoint->AioMgr.cRequestsActive--;
591
592	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_FLUSH)
593	{
594	/* Clear the pending flush */
595	Assert(pEndpoint->pFlushReq == pTask);
596	pEndpoint->pFlushReq = NULL;
597	}
598	}
599	}
600
601	pAioMgr->cRequestsActiveMax = pAioMgr->cRequestsActive;
602
603	/* Print an entry in the release log */
604	if (RT_UNLIKELY(!pEpClass->fOutOfResourcesWarningPrinted))
605	{
606	pEpClass->fOutOfResourcesWarningPrinted = true;
607	LogRel(("AIOMgr: Host limits number of active IO requests to %u. Expect a performance impact.\n",
608	pAioMgr->cRequestsActive));
609	}
610
611	LogFlow(("Removed requests. I/O manager has a total of %u active requests now\n", pAioMgr->cRequestsActive));
612	LogFlow(("Endpoint has a total of %u active requests now\n", pEndpoint->AioMgr.cRequestsActive));
613	rc = VINF_SUCCESS;
614	}
615	else /* Another kind of error happened (full disk, ...) */
616	{
617	/* An error happened. Find out which one caused the error and resubmit all other tasks. */
618	for (size_t i = 0; i < cReqs; i++)
619	{
620	int rcReq = RTFileAioReqGetRC(pahReqs[i], NULL);
621
622	if (rcReq == VERR_FILE_AIO_NOT_SUBMITTED)
623	{
624	/* We call ourself again to do any error handling which might come up now. */
625	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, &pahReqs[i], 1);
626	AssertRC(rc);
627	}
628	else if (rcReq != VERR_FILE_AIO_IN_PROGRESS)
629	pdmacFileAioMgrNormalReqCompleteRc(pAioMgr, pahReqs[i], rcReq, 0);
630	}
631
632
633	if ( pEndpoint->pFlushReq
634	&& !pAioMgr->cRequestsActive
635	&& !pEndpoint->fAsyncFlushSupported)
636	{
637	/*
638	* Complete a pending flush if we don't have requests enqueued and the host doesn't support
639	* the async flush API.
640	* Happens only if this we just noticed that this is not supported
641	* and the only active request was a flush.
642	*/
643	PPDMACTASKFILE pFlush = pEndpoint->pFlushReq;
644	pEndpoint->pFlushReq = NULL;
645	pFlush->pfnCompleted(pFlush, pFlush->pvUser, VINF_SUCCESS);
646	pdmacFileTaskFree(pEndpoint, pFlush);
647	}
648	}
649	}
650
651	return VINF_SUCCESS;
652	}
653
654	static bool pdmacFileAioMgrNormalIsRangeLocked(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
655	RTFOFF offStart, size_t cbRange,
656	PPDMACTASKFILE pTask, bool fAlignedReq)
657	{
658	AssertMsg( pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE
659	\|\| pTask->enmTransferType == PDMACTASKFILETRANSFER_READ,
660	("Invalid task type %d\n", pTask->enmTransferType));
661
662	/*
663	* If there is no unaligned request active and the current one is aligned
664	* just pass it through.
665	*/
666	if (!pEndpoint->AioMgr.cLockedReqsActive && fAlignedReq)
667	return false;
668
669	PPDMACFILERANGELOCK pRangeLock;
670	pRangeLock = (PPDMACFILERANGELOCK)RTAvlrFileOffsetRangeGet(pEndpoint->AioMgr.pTreeRangesLocked, offStart);
671	if (!pRangeLock)
672	{
673	pRangeLock = (PPDMACFILERANGELOCK)RTAvlrFileOffsetGetBestFit(pEndpoint->AioMgr.pTreeRangesLocked, offStart, true);
674	/* Check if we intersect with the range. */
675	if ( !pRangeLock
676	\|\| !( (pRangeLock->Core.Key) <= (offStart + (RTFOFF)cbRange - 1)
677	&& (pRangeLock->Core.KeyLast) >= offStart))
678	{
679	pRangeLock = NULL; /* False alarm */
680	}
681	}
682
683	/* Check whether we have one of the situations explained below */
684	if (pRangeLock)
685	{
686	/* Add to the list. */
687	pTask->pNext = NULL;
688
689	if (!pRangeLock->pWaitingTasksHead)
690	{
691	Assert(!pRangeLock->pWaitingTasksTail);
692	pRangeLock->pWaitingTasksHead = pTask;
693	pRangeLock->pWaitingTasksTail = pTask;
694	}
695	else
696	{
697	AssertPtr(pRangeLock->pWaitingTasksTail);
698	pRangeLock->pWaitingTasksTail->pNext = pTask;
699	pRangeLock->pWaitingTasksTail = pTask;
700	}
701	return true;
702	}
703
704	return false;
705	}
706
707	static int pdmacFileAioMgrNormalRangeLock(PPDMACEPFILEMGR pAioMgr,
708	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
709	RTFOFF offStart, size_t cbRange,
710	PPDMACTASKFILE pTask, bool fAlignedReq)
711	{
712	LogFlowFunc(("pAioMgr=%#p pEndpoint=%#p offStart=%RTfoff cbRange=%zu pTask=%#p\n",
713	pAioMgr, pEndpoint, offStart, cbRange, pTask));
714
715	AssertMsg(!pdmacFileAioMgrNormalIsRangeLocked(pEndpoint, offStart, cbRange, pTask, fAlignedReq),
716	("Range is already locked offStart=%RTfoff cbRange=%u\n",
717	offStart, cbRange));
718
719	/*
720	* If there is no unaligned request active and the current one is aligned
721	* just don't use the lock.
722	*/
723	if (!pEndpoint->AioMgr.cLockedReqsActive && fAlignedReq)
724	{
725	pTask->pRangeLock = NULL;
726	return VINF_SUCCESS;
727	}
728
729	PPDMACFILERANGELOCK pRangeLock = (PPDMACFILERANGELOCK)RTMemCacheAlloc(pAioMgr->hMemCacheRangeLocks);
730	if (!pRangeLock)
731	return VERR_NO_MEMORY;
732
733	/* Init the lock. */
734	pRangeLock->Core.Key = offStart;
735	pRangeLock->Core.KeyLast = offStart + cbRange - 1;
736	pRangeLock->cRefs = 1;
737	pRangeLock->fReadLock = pTask->enmTransferType == PDMACTASKFILETRANSFER_READ;
738	pRangeLock->pWaitingTasksHead = NULL;
739	pRangeLock->pWaitingTasksTail = NULL;
740
741	bool fInserted = RTAvlrFileOffsetInsert(pEndpoint->AioMgr.pTreeRangesLocked, &pRangeLock->Core);
742	AssertMsg(fInserted, ("Range lock was not inserted!\n")); NOREF(fInserted);
743
744	/* Let the task point to its lock. */
745	pTask->pRangeLock = pRangeLock;
746	pEndpoint->AioMgr.cLockedReqsActive++;
747
748	return VINF_SUCCESS;
749	}
750
751	static PPDMACTASKFILE pdmacFileAioMgrNormalRangeLockFree(PPDMACEPFILEMGR pAioMgr,
752	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
753	PPDMACFILERANGELOCK pRangeLock)
754	{
755	PPDMACTASKFILE pTasksWaitingHead;
756
757	LogFlowFunc(("pAioMgr=%#p pEndpoint=%#p pRangeLock=%#p\n",
758	pAioMgr, pEndpoint, pRangeLock));
759
760	/* pRangeLock can be NULL if there was no lock assigned with the task. */
761	if (!pRangeLock)
762	return NULL;
763
764	Assert(pRangeLock->cRefs == 1);
765
766	RTAvlrFileOffsetRemove(pEndpoint->AioMgr.pTreeRangesLocked, pRangeLock->Core.Key);
767	pTasksWaitingHead = pRangeLock->pWaitingTasksHead;
768	pRangeLock->pWaitingTasksHead = NULL;
769	pRangeLock->pWaitingTasksTail = NULL;
770	RTMemCacheFree(pAioMgr->hMemCacheRangeLocks, pRangeLock);
771	pEndpoint->AioMgr.cLockedReqsActive--;
772
773	return pTasksWaitingHead;
774	}
775
776	static int pdmacFileAioMgrNormalTaskPrepareBuffered(PPDMACEPFILEMGR pAioMgr,
777	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
778	PPDMACTASKFILE pTask, PRTFILEAIOREQ phReq)
779	{
780	AssertMsg( pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE
781	\|\| (uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) <= pEndpoint->cbFile,
782	("Read exceeds file size offStart=%RTfoff cbToTransfer=%d cbFile=%llu\n",
783	pTask->Off, pTask->DataSeg.cbSeg, pEndpoint->cbFile));
784
785	pTask->fPrefetch = false;
786	pTask->cbBounceBuffer = 0;
787
788	/*
789	* Before we start to setup the request we have to check whether there is a task
790	* already active which range intersects with ours. We have to defer execution
791	* of this task in two cases:
792	* - The pending task is a write and the current is either read or write
793	* - The pending task is a read and the current task is a write task.
794	*
795	* To check whether a range is currently "locked" we use the AVL tree where every pending task
796	* is stored by its file offset range. The current task will be added to the active task
797	* and will be executed when the active one completes. (The method below
798	* which checks whether a range is already used will add the task)
799	*
800	* This is necessary because of the requirement to align all requests to a 512 boundary
801	* which is enforced by the host OS (Linux and Windows atm). It is possible that
802	* we have to process unaligned tasks and need to align them using bounce buffers.
803	* While the data is fetched from the file another request might arrive writing to
804	* the same range. This will result in data corruption if both are executed concurrently.
805	*/
806	int rc = VINF_SUCCESS;
807	bool fLocked = pdmacFileAioMgrNormalIsRangeLocked(pEndpoint, pTask->Off, pTask->DataSeg.cbSeg, pTask,
808	true /* fAlignedReq */);
809	if (!fLocked)
810	{
811	/* Get a request handle. */
812	RTFILEAIOREQ hReq = pdmacFileAioMgrNormalRequestAlloc(pAioMgr);
813	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
814
815	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
816	{
817	/* Grow the file if needed. */
818	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
819	{
820	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
821	RTFileSetSize(pEndpoint->hFile, pTask->Off + pTask->DataSeg.cbSeg);
822	}
823
824	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->hFile,
825	pTask->Off, pTask->DataSeg.pvSeg,
826	pTask->DataSeg.cbSeg, pTask);
827	}
828	else
829	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->hFile,
830	pTask->Off, pTask->DataSeg.pvSeg,
831	pTask->DataSeg.cbSeg, pTask);
832	AssertRC(rc);
833
834	rc = pdmacFileAioMgrNormalRangeLock(pAioMgr, pEndpoint, pTask->Off,
835	pTask->DataSeg.cbSeg,
836	pTask, true /* fAlignedReq */);
837
838	if (RT_SUCCESS(rc))
839	{
840	pTask->hReq = hReq;
841	*phReq = hReq;
842	}
843	}
844	else
845	LogFlow(("Task %#p was deferred because the access range is locked\n", pTask));
846
847	return rc;
848	}
849
850	static int pdmacFileAioMgrNormalTaskPrepareNonBuffered(PPDMACEPFILEMGR pAioMgr,
851	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
852	PPDMACTASKFILE pTask, PRTFILEAIOREQ phReq)
853	{
854	/*
855	* Check if the alignment requirements are met.
856	* Offset, transfer size and buffer address
857	* need to be on a 512 boundary.
858	*/
859	RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1);
860	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg + (pTask->Off - offStart), 512);
861	PDMACTASKFILETRANSFER enmTransferType = pTask->enmTransferType;
862	bool fAlignedReq = cbToTransfer == pTask->DataSeg.cbSeg
863	&& offStart == pTask->Off;
864
865	AssertMsg( pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE
866	\|\| (uint64_t)(offStart + cbToTransfer) <= pEndpoint->cbFile,
867	("Read exceeds file size offStart=%RTfoff cbToTransfer=%d cbFile=%llu\n",
868	offStart, cbToTransfer, pEndpoint->cbFile));
869
870	pTask->fPrefetch = false;
871
872	/*
873	* Before we start to setup the request we have to check whether there is a task
874	* already active which range intersects with ours. We have to defer execution
875	* of this task in two cases:
876	* - The pending task is a write and the current is either read or write
877	* - The pending task is a read and the current task is a write task.
878	*
879	* To check whether a range is currently "locked" we use the AVL tree where every pending task
880	* is stored by its file offset range. The current task will be added to the active task
881	* and will be executed when the active one completes. (The method below
882	* which checks whether a range is already used will add the task)
883	*
884	* This is necessary because of the requirement to align all requests to a 512 boundary
885	* which is enforced by the host OS (Linux and Windows atm). It is possible that
886	* we have to process unaligned tasks and need to align them using bounce buffers.
887	* While the data is fetched from the file another request might arrive writing to
888	* the same range. This will result in data corruption if both are executed concurrently.
889	*/
890	int rc = VINF_SUCCESS;
891	bool fLocked = pdmacFileAioMgrNormalIsRangeLocked(pEndpoint, offStart, cbToTransfer, pTask, fAlignedReq);
892	if (!fLocked)
893	{
894	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
895	void *pvBuf = pTask->DataSeg.pvSeg;
896
897	/* Get a request handle. */
898	RTFILEAIOREQ hReq = pdmacFileAioMgrNormalRequestAlloc(pAioMgr);
899	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
900
901	if ( !fAlignedReq
902	\|\| ((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) != (RTR3UINTPTR)pvBuf))
903	{
904	LogFlow(("Using bounce buffer for task %#p cbToTransfer=%zd cbSeg=%zd offStart=%RTfoff off=%RTfoff\n",
905	pTask, cbToTransfer, pTask->DataSeg.cbSeg, offStart, pTask->Off));
906
907	/* Create bounce buffer. */
908	pTask->cbBounceBuffer = cbToTransfer;
909
910	AssertMsg(pTask->Off >= offStart, ("Overflow in calculation Off=%llu offStart=%llu\n",
911	pTask->Off, offStart));
912	pTask->offBounceBuffer = pTask->Off - offStart;
913
914	/** @todo I think we need something like a RTMemAllocAligned method here.
915	* Current assumption is that the maximum alignment is 4096byte
916	* (GPT disk on Windows)
917	* so we can use RTMemPageAlloc here.
918	*/
919	pTask->pvBounceBuffer = RTMemPageAlloc(cbToTransfer);
920	if (RT_LIKELY(pTask->pvBounceBuffer))
921	{
922	pvBuf = pTask->pvBounceBuffer;
923
924	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
925	{
926	if ( RT_UNLIKELY(cbToTransfer != pTask->DataSeg.cbSeg)
927	\|\| RT_UNLIKELY(offStart != pTask->Off))
928	{
929	/* We have to fill the buffer first before we can update the data. */
930	LogFlow(("Prefetching data for task %#p\n", pTask));
931	pTask->fPrefetch = true;
932	enmTransferType = PDMACTASKFILETRANSFER_READ;
933	}
934	else
935	memcpy(pvBuf, pTask->DataSeg.pvSeg, pTask->DataSeg.cbSeg);
936	}
937	}
938	else
939	rc = VERR_NO_MEMORY;
940	}
941	else
942	pTask->cbBounceBuffer = 0;
943
944	if (RT_SUCCESS(rc))
945	{
946	AssertMsg((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) == (RTR3UINTPTR)pvBuf,
947	("AIO: Alignment restrictions not met! pvBuf=%p uBitmaskAlignment=%p\n", pvBuf, pEpClassFile->uBitmaskAlignment));
948
949	if (enmTransferType == PDMACTASKFILETRANSFER_WRITE)
950	{
951	/* Grow the file if needed. */
952	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
953	{
954	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
955	RTFileSetSize(pEndpoint->hFile, pTask->Off + pTask->DataSeg.cbSeg);
956	}
957
958	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->hFile,
959	offStart, pvBuf, cbToTransfer, pTask);
960	}
961	else
962	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->hFile,
963	offStart, pvBuf, cbToTransfer, pTask);
964	AssertRC(rc);
965
966	rc = pdmacFileAioMgrNormalRangeLock(pAioMgr, pEndpoint, offStart, cbToTransfer, pTask, fAlignedReq);
967	if (RT_SUCCESS(rc))
968	{
969	pTask->hReq = hReq;
970	*phReq = hReq;
971	}
972	else
973	{
974	/* Cleanup */
975	if (pTask->cbBounceBuffer)
976	RTMemPageFree(pTask->pvBounceBuffer, pTask->cbBounceBuffer);
977	}
978	}
979	}
980	else
981	LogFlow(("Task %#p was deferred because the access range is locked\n", pTask));
982
983	return rc;
984	}
985
986	static int pdmacFileAioMgrNormalProcessTaskList(PPDMACTASKFILE pTaskHead,
987	PPDMACEPFILEMGR pAioMgr,
988	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
989	{
990	RTFILEAIOREQ apReqs[20];
991	unsigned cRequests = 0;
992	int rc = VINF_SUCCESS;
993
994	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
995	("Trying to process request lists of a non active endpoint!\n"));
996
997	/* Go through the list and queue the requests until we get a flush request */
998	while ( pTaskHead
999	&& !pEndpoint->pFlushReq
1000	&& (pAioMgr->cRequestsActive + cRequests < pAioMgr->cRequestsActiveMax)
1001	&& RT_SUCCESS(rc))
1002	{
1003	RTMSINTERVAL msWhenNext;
1004	PPDMACTASKFILE pCurr = pTaskHead;
1005
1006	if (!pdmacEpIsTransferAllowed(&pEndpoint->Core, (uint32_t)pCurr->DataSeg.cbSeg, &msWhenNext))
1007	{
1008	pAioMgr->msBwLimitExpired = RT_MIN(pAioMgr->msBwLimitExpired, msWhenNext);
1009	break;
1010	}
1011
1012	pTaskHead = pTaskHead->pNext;
1013
1014	pCurr->pNext = NULL;
1015
1016	AssertMsg(RT_VALID_PTR(pCurr->pEndpoint) && pCurr->pEndpoint == pEndpoint,
1017	("Endpoints do not match\n"));
1018
1019	switch (pCurr->enmTransferType)
1020	{
1021	case PDMACTASKFILETRANSFER_FLUSH:
1022	{
1023	/* If there is no data transfer request this flush request finished immediately. */
1024	if (pEndpoint->fAsyncFlushSupported)
1025	{
1026	/* Issue a flush to the host. */
1027	RTFILEAIOREQ hReq = pdmacFileAioMgrNormalRequestAlloc(pAioMgr);
1028	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
1029
1030	LogFlow(("Flush request %#p\n", hReq));
1031
1032	rc = RTFileAioReqPrepareFlush(hReq, pEndpoint->hFile, pCurr);
1033	if (RT_FAILURE(rc))
1034	{
1035	if (rc == VERR_NOT_SUPPORTED)
1036	LogRel(("AIOMgr: Async flushes not supported\n"));
1037	else
1038	LogRel(("AIOMgr: Preparing flush failed with %Rrc, disabling async flushes\n", rc));
1039	pEndpoint->fAsyncFlushSupported = false;
1040	pdmacFileAioMgrNormalRequestFree(pAioMgr, hReq);
1041	rc = VINF_SUCCESS; /* Fake success */
1042	}
1043	else
1044	{
1045	pCurr->hReq = hReq;
1046	apReqs[cRequests] = hReq;
1047	pEndpoint->AioMgr.cReqsProcessed++;
1048	cRequests++;
1049	}
1050	}
1051
1052	if ( !pEndpoint->AioMgr.cRequestsActive
1053	&& !pEndpoint->fAsyncFlushSupported)
1054	{
1055	pCurr->pfnCompleted(pCurr, pCurr->pvUser, VINF_SUCCESS);
1056	pdmacFileTaskFree(pEndpoint, pCurr);
1057	}
1058	else
1059	{
1060	Assert(!pEndpoint->pFlushReq);
1061	pEndpoint->pFlushReq = pCurr;
1062	}
1063	break;
1064	}
1065	case PDMACTASKFILETRANSFER_READ:
1066	case PDMACTASKFILETRANSFER_WRITE:
1067	{
1068	RTFILEAIOREQ hReq = NIL_RTFILEAIOREQ;
1069
1070	if (pCurr->hReq == NIL_RTFILEAIOREQ)
1071	{
1072	if (pEndpoint->enmBackendType == PDMACFILEEPBACKEND_BUFFERED)
1073	rc = pdmacFileAioMgrNormalTaskPrepareBuffered(pAioMgr, pEndpoint, pCurr, &hReq);
1074	else if (pEndpoint->enmBackendType == PDMACFILEEPBACKEND_NON_BUFFERED)
1075	rc = pdmacFileAioMgrNormalTaskPrepareNonBuffered(pAioMgr, pEndpoint, pCurr, &hReq);
1076	else
1077	AssertMsgFailed(("Invalid backend type %d\n", pEndpoint->enmBackendType));
1078
1079	AssertRC(rc);
1080	}
1081	else
1082	{
1083	LogFlow(("Task %#p has I/O request %#p already\n", pCurr, pCurr->hReq));
1084	hReq = pCurr->hReq;
1085	}
1086
1087	LogFlow(("Read/Write request %#p\n", hReq));
1088
1089	if (hReq != NIL_RTFILEAIOREQ)
1090	{
1091	apReqs[cRequests] = hReq;
1092	cRequests++;
1093	}
1094	break;
1095	}
1096	default:
1097	AssertMsgFailed(("Invalid transfer type %d\n", pCurr->enmTransferType));
1098	} /* switch transfer type */
1099
1100	/* Queue the requests if the array is full. */
1101	if (cRequests == RT_ELEMENTS(apReqs))
1102	{
1103	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
1104	cRequests = 0;
1105	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
1106	("Unexpected return code\n"));
1107	}
1108	}
1109
1110	if (cRequests)
1111	{
1112	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
1113	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
1114	("Unexpected return code rc=%Rrc\n", rc));
1115	}
1116
1117	if (pTaskHead)
1118	{
1119	/* Add the rest of the tasks to the pending list */
1120	pdmacFileAioMgrEpAddTaskList(pEndpoint, pTaskHead);
1121
1122	if (RT_UNLIKELY( pAioMgr->cRequestsActiveMax == pAioMgr->cRequestsActive
1123	&& !pEndpoint->pFlushReq))
1124	{
1125	#if 0
1126	/*
1127	* The I/O manager has no room left for more requests
1128	* but there are still requests to process.
1129	* Create a new I/O manager and let it handle some endpoints.
1130	*/
1131	pdmacFileAioMgrNormalBalanceLoad(pAioMgr);
1132	#else
1133	/* Grow the I/O manager */
1134	pAioMgr->enmState = PDMACEPFILEMGRSTATE_GROWING;
1135	#endif
1136	}
1137	}
1138
1139	/* Insufficient resources are not fatal. */
1140	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
1141	rc = VINF_SUCCESS;
1142
1143	return rc;
1144	}
1145
1146	/**
1147	* Adds all pending requests for the given endpoint
1148	* until a flush request is encountered or there is no
1149	* request anymore.
1150	*
1151	* @returns VBox status code.
1152	* @param pAioMgr The async I/O manager for the endpoint
1153	* @param pEndpoint The endpoint to get the requests from.
1154	*/
1155	static int pdmacFileAioMgrNormalQueueReqs(PPDMACEPFILEMGR pAioMgr,
1156	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
1157	{
1158	int rc = VINF_SUCCESS;
1159	PPDMACTASKFILE pTasksHead = NULL;
1160
1161	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
1162	("Trying to process request lists of a non active endpoint!\n"));
1163
1164	Assert(!pEndpoint->pFlushReq);
1165
1166	/* Check the pending list first */
1167	if (pEndpoint->AioMgr.pReqsPendingHead)
1168	{
1169	LogFlow(("Queuing pending requests first\n"));
1170
1171	pTasksHead = pEndpoint->AioMgr.pReqsPendingHead;
1172	/*
1173	* Clear the list as the processing routine will insert them into the list
1174	* again if it gets a flush request.
1175	*/
1176	pEndpoint->AioMgr.pReqsPendingHead = NULL;
1177	pEndpoint->AioMgr.pReqsPendingTail = NULL;
1178	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
1179	AssertRC(rc); /** @todo r=bird: status code potentially overwritten. */
1180	}
1181
1182	if (!pEndpoint->pFlushReq && !pEndpoint->AioMgr.pReqsPendingHead)
1183	{
1184	/* Now the request queue. */
1185	pTasksHead = pdmacFileEpGetNewTasks(pEndpoint);
1186	if (pTasksHead)
1187	{
1188	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
1189	AssertRC(rc);
1190	}
1191	}
1192
1193	return rc;
1194	}
1195
1196	static int pdmacFileAioMgrNormalProcessBlockingEvent(PPDMACEPFILEMGR pAioMgr)
1197	{
1198	int rc = VINF_SUCCESS;
1199	bool fNotifyWaiter = false;
1200
1201	LogFlowFunc((": Enter\n"));
1202
1203	Assert(pAioMgr->fBlockingEventPending);
1204
1205	switch (pAioMgr->enmBlockingEvent)
1206	{
1207	case PDMACEPFILEAIOMGRBLOCKINGEVENT_ADD_ENDPOINT:
1208	{
1209	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointNew = ASMAtomicReadPtrT(&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint, PPDMASYNCCOMPLETIONENDPOINTFILE);
1210	AssertMsg(RT_VALID_PTR(pEndpointNew), ("Adding endpoint event without a endpoint to add\n"));
1211
1212	pEndpointNew->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE;
1213
1214	pEndpointNew->AioMgr.pEndpointNext = pAioMgr->pEndpointsHead;
1215	pEndpointNew->AioMgr.pEndpointPrev = NULL;
1216	if (pAioMgr->pEndpointsHead)
1217	pAioMgr->pEndpointsHead->AioMgr.pEndpointPrev = pEndpointNew;
1218	pAioMgr->pEndpointsHead = pEndpointNew;
1219
1220	/* Assign the completion point to this file. */
1221	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pEndpointNew->hFile);
1222	fNotifyWaiter = true;
1223	pAioMgr->cEndpoints++;
1224	break;
1225	}
1226	case PDMACEPFILEAIOMGRBLOCKINGEVENT_REMOVE_ENDPOINT:
1227	{
1228	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove = ASMAtomicReadPtrT(&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint, PPDMASYNCCOMPLETIONENDPOINTFILE);
1229	AssertMsg(RT_VALID_PTR(pEndpointRemove), ("Removing endpoint event without a endpoint to remove\n"));
1230
1231	pEndpointRemove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
1232	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointRemove);
1233	break;
1234	}
1235	case PDMACEPFILEAIOMGRBLOCKINGEVENT_CLOSE_ENDPOINT:
1236	{
1237	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointClose = ASMAtomicReadPtrT(&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint, PPDMASYNCCOMPLETIONENDPOINTFILE);
1238	AssertMsg(RT_VALID_PTR(pEndpointClose), ("Close endpoint event without a endpoint to close\n"));
1239
1240	if (pEndpointClose->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
1241	{
1242	LogFlowFunc((": Closing endpoint %#p{%s}\n", pEndpointClose, pEndpointClose->Core.pszUri));
1243
1244	/* Make sure all tasks finished. Process the queues a last time first. */
1245	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpointClose);
1246	AssertRC(rc);
1247
1248	pEndpointClose->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING;
1249	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointClose);
1250	}
1251	else if ( (pEndpointClose->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING)
1252	&& (!pEndpointClose->AioMgr.cRequestsActive))
1253	fNotifyWaiter = true;
1254	break;
1255	}
1256	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SHUTDOWN:
1257	{
1258	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SHUTDOWN;
1259	if (!pAioMgr->cRequestsActive)
1260	fNotifyWaiter = true;
1261	break;
1262	}
1263	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SUSPEND:
1264	{
1265	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SUSPENDING;
1266	break;
1267	}
1268	case PDMACEPFILEAIOMGRBLOCKINGEVENT_RESUME:
1269	{
1270	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
1271	fNotifyWaiter = true;
1272	break;
1273	}
1274	default:
1275	AssertReleaseMsgFailed(("Invalid event type %d\n", pAioMgr->enmBlockingEvent));
1276	}
1277
1278	if (fNotifyWaiter)
1279	{
1280	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
1281	pAioMgr->enmBlockingEvent = PDMACEPFILEAIOMGRBLOCKINGEVENT_INVALID;
1282
1283	/* Release the waiting thread. */
1284	LogFlow(("Signalling waiter\n"));
1285	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
1286	AssertRC(rc);
1287	}
1288
1289	LogFlowFunc((": Leave\n"));
1290	return rc;
1291	}
1292
1293	/**
1294	* Checks all endpoints for pending events or new requests.
1295	*
1296	* @returns VBox status code.
1297	* @param pAioMgr The I/O manager handle.
1298	*/
1299	static int pdmacFileAioMgrNormalCheckEndpoints(PPDMACEPFILEMGR pAioMgr)
1300	{
1301	/* Check the assigned endpoints for new tasks if there isn't a flush request active at the moment. */
1302	int rc = VINF_SUCCESS;
1303	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint = pAioMgr->pEndpointsHead;
1304
1305	pAioMgr->msBwLimitExpired = RT_INDEFINITE_WAIT;
1306
1307	while (pEndpoint)
1308	{
1309	if (!pEndpoint->pFlushReq
1310	&& (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
1311	&& !pEndpoint->AioMgr.fMoving)
1312	{
1313	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
1314	if (RT_FAILURE(rc))
1315	return rc;
1316	}
1317	else if ( !pEndpoint->AioMgr.cRequestsActive
1318	&& pEndpoint->enmState != PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
1319	{
1320	/* Reopen the file so that the new endpoint can re-associate with the file */
1321	RTFileClose(pEndpoint->hFile);
1322	rc = RTFileOpen(&pEndpoint->hFile, pEndpoint->Core.pszUri, pEndpoint->fFlags);
1323	AssertRC(rc);
1324
1325	if (pEndpoint->AioMgr.fMoving)
1326	{
1327	pEndpoint->AioMgr.fMoving = false;
1328	pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
1329	}
1330	else
1331	{
1332	Assert(pAioMgr->fBlockingEventPending);
1333	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
1334
1335	/* Release the waiting thread. */
1336	LogFlow(("Signalling waiter\n"));
1337	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
1338	AssertRC(rc);
1339	}
1340	}
1341
1342	pEndpoint = pEndpoint->AioMgr.pEndpointNext;
1343	}
1344
1345	return rc;
1346	}
1347
1348	/**
1349	* Wrapper around pdmacFileAioMgrNormalReqCompleteRc().
1350	*/
1351	static void pdmacFileAioMgrNormalReqComplete(PPDMACEPFILEMGR pAioMgr, RTFILEAIOREQ hReq)
1352	{
1353	size_t cbTransfered = 0;
1354	int rcReq = RTFileAioReqGetRC(hReq, &cbTransfered);
1355
1356	pdmacFileAioMgrNormalReqCompleteRc(pAioMgr, hReq, rcReq, cbTransfered);
1357	}
1358
1359	static void pdmacFileAioMgrNormalReqCompleteRc(PPDMACEPFILEMGR pAioMgr, RTFILEAIOREQ hReq,
1360	int rcReq, size_t cbTransfered)
1361	{
1362	int rc = VINF_SUCCESS;
1363	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint;
1364	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(hReq);
1365	PPDMACTASKFILE pTasksWaiting;
1366
1367	LogFlowFunc(("pAioMgr=%#p hReq=%#p\n", pAioMgr, hReq));
1368
1369	pEndpoint = pTask->pEndpoint;
1370
1371	pTask->hReq = NIL_RTFILEAIOREQ;
1372
1373	pAioMgr->cRequestsActive--;
1374	pEndpoint->AioMgr.cRequestsActive--;
1375	pEndpoint->AioMgr.cReqsProcessed++;
1376
1377	/*
1378	* It is possible that the request failed on Linux with kernels < 2.6.23
1379	* if the passed buffer was allocated with remap_pfn_range or if the file
1380	* is on an NFS endpoint which does not support async and direct I/O at the same time.
1381	* The endpoint will be migrated to a failsafe manager in case a request fails.
1382	*/
1383	if (RT_FAILURE(rcReq))
1384	{
1385	/* Free bounce buffers and the IPRT request. */
1386	pdmacFileAioMgrNormalRequestFree(pAioMgr, hReq);
1387
1388	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_FLUSH)
1389	{
1390	LogRel(("AIOMgr: Flush failed with %Rrc, disabling async flushes\n", rcReq));
1391	pEndpoint->fAsyncFlushSupported = false;
1392	AssertMsg(pEndpoint->pFlushReq == pTask, ("Failed flush request doesn't match active one\n"));
1393	/* The other method will take over now. */
1394
1395	pEndpoint->pFlushReq = NULL;
1396	/* Call completion callback */
1397	LogFlow(("Flush task=%#p completed with %Rrc\n", pTask, VINF_SUCCESS));
1398	pTask->pfnCompleted(pTask, pTask->pvUser, VINF_SUCCESS);
1399	pdmacFileTaskFree(pEndpoint, pTask);
1400	}
1401	else
1402	{
1403	/* Free the lock and process pending tasks if necessary */
1404	pTasksWaiting = pdmacFileAioMgrNormalRangeLockFree(pAioMgr, pEndpoint, pTask->pRangeLock);
1405	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksWaiting, pAioMgr, pEndpoint);
1406	AssertRC(rc);
1407
1408	if (pTask->cbBounceBuffer)
1409	RTMemPageFree(pTask->pvBounceBuffer, pTask->cbBounceBuffer);
1410
1411	/*
1412	* Fatal errors are reported to the guest and non-fatal errors
1413	* will cause a migration to the failsafe manager in the hope
1414	* that the error disappears.
1415	*/
1416	if (!pdmacFileAioMgrNormalRcIsFatal(rcReq))
1417	{
1418	/* Queue the request on the pending list. */
1419	pTask->pNext = pEndpoint->AioMgr.pReqsPendingHead;
1420	pEndpoint->AioMgr.pReqsPendingHead = pTask;
1421
1422	/* Create a new failsafe manager if necessary. */
1423	if (!pEndpoint->AioMgr.fMoving)
1424	{
1425	PPDMACEPFILEMGR pAioMgrFailsafe;
1426
1427	LogRel(("%s: Request %#p failed with rc=%Rrc, migrating endpoint %s to failsafe manager.\n",
1428	RTThreadGetName(pAioMgr->Thread), pTask, rcReq, pEndpoint->Core.pszUri));
1429
1430	pEndpoint->AioMgr.fMoving = true;
1431
1432	rc = pdmacFileAioMgrCreate((PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass,
1433	&pAioMgrFailsafe, PDMACEPFILEMGRTYPE_SIMPLE);
1434	AssertRC(rc);
1435
1436	pEndpoint->AioMgr.pAioMgrDst = pAioMgrFailsafe;
1437
1438	/* Update the flags to open the file with. Disable async I/O and enable the host cache. */
1439	pEndpoint->fFlags &= ~(RTFILE_O_ASYNC_IO \| RTFILE_O_NO_CACHE);
1440	}
1441
1442	/* If this was the last request for the endpoint migrate it to the new manager. */
1443	if (!pEndpoint->AioMgr.cRequestsActive)
1444	{
1445	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pEndpoint);
1446	Assert(!fReqsPending); NOREF(fReqsPending);
1447
1448	rc = pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
1449	AssertRC(rc);
1450	}
1451	}
1452	else
1453	{
1454	pTask->pfnCompleted(pTask, pTask->pvUser, rcReq);
1455	pdmacFileTaskFree(pEndpoint, pTask);
1456	}
1457	}
1458	}
1459	else
1460	{
1461	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_FLUSH)
1462	{
1463	/* Clear pending flush */
1464	AssertMsg(pEndpoint->pFlushReq == pTask, ("Completed flush request doesn't match active one\n"));
1465	pEndpoint->pFlushReq = NULL;
1466	pdmacFileAioMgrNormalRequestFree(pAioMgr, hReq);
1467
1468	/* Call completion callback */
1469	LogFlow(("Flush task=%#p completed with %Rrc\n", pTask, rcReq));
1470	pTask->pfnCompleted(pTask, pTask->pvUser, rcReq);
1471	pdmacFileTaskFree(pEndpoint, pTask);
1472	}
1473	else
1474	{
1475	/*
1476	* Restart an incomplete transfer.
1477	* This usually means that the request will return an error now
1478	* but to get the cause of the error (disk full, file too big, I/O error, ...)
1479	* the transfer needs to be continued.
1480	*/
1481	pTask->cbTransfered += cbTransfered;
1482
1483	if (RT_UNLIKELY( pTask->cbTransfered < pTask->DataSeg.cbSeg
1484	\|\| ( pTask->cbBounceBuffer
1485	&& pTask->cbTransfered < pTask->cbBounceBuffer)))
1486	{
1487	RTFOFF offStart;
1488	size_t cbToTransfer;
1489	uint8_t *pbBuf = NULL;
1490
1491	LogFlow(("Restarting incomplete transfer %#p (%zu bytes transferred)\n",
1492	pTask, cbTransfered));
1493	Assert(cbTransfered % 512 == 0);
1494
1495	if (pTask->cbBounceBuffer)
1496	{
1497	AssertPtr(pTask->pvBounceBuffer);
1498	offStart = (pTask->Off & ~((RTFOFF)512-1)) + pTask->cbTransfered;
1499	cbToTransfer = pTask->cbBounceBuffer - pTask->cbTransfered;
1500	pbBuf = (uint8_t *)pTask->pvBounceBuffer + pTask->cbTransfered;
1501	}
1502	else
1503	{
1504	Assert(!pTask->pvBounceBuffer);
1505	offStart = pTask->Off + pTask->cbTransfered;
1506	cbToTransfer = pTask->DataSeg.cbSeg - pTask->cbTransfered;
1507	pbBuf = (uint8_t *)pTask->DataSeg.pvSeg + pTask->cbTransfered;
1508	}
1509
1510	if (pTask->fPrefetch \|\| pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
1511	{
1512	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->hFile, offStart,
1513	pbBuf, cbToTransfer, pTask);
1514	}
1515	else
1516	{
1517	AssertMsg(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE,
1518	("Invalid transfer type\n"));
1519	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->hFile, offStart,
1520	pbBuf, cbToTransfer, pTask);
1521	}
1522	AssertRC(rc);
1523
1524	pTask->hReq = hReq;
1525	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, &hReq, 1);
1526	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
1527	("Unexpected return code rc=%Rrc\n", rc));
1528	}
1529	else if (pTask->fPrefetch)
1530	{
1531	Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE);
1532	Assert(pTask->cbBounceBuffer);
1533
1534	memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->offBounceBuffer,
1535	pTask->DataSeg.pvSeg,
1536	pTask->DataSeg.cbSeg);
1537
1538	/* Write it now. */
1539	pTask->fPrefetch = false;
1540	RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1);
1541	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg + (pTask->Off - offStart), 512);
1542
1543	pTask->cbTransfered = 0;
1544
1545	/* Grow the file if needed. */
1546	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
1547	{
1548	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
1549	RTFileSetSize(pEndpoint->hFile, pTask->Off + pTask->DataSeg.cbSeg);
1550	}
1551
1552	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->hFile,
1553	offStart, pTask->pvBounceBuffer, cbToTransfer, pTask);
1554	AssertRC(rc);
1555	pTask->hReq = hReq;
1556	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, &hReq, 1);
1557	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
1558	("Unexpected return code rc=%Rrc\n", rc));
1559	}
1560	else
1561	{
1562	if (RT_SUCCESS(rc) && pTask->cbBounceBuffer)
1563	{
1564	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
1565	memcpy(pTask->DataSeg.pvSeg,
1566	((uint8_t *)pTask->pvBounceBuffer) + pTask->offBounceBuffer,
1567	pTask->DataSeg.cbSeg);
1568
1569	RTMemPageFree(pTask->pvBounceBuffer, pTask->cbBounceBuffer);
1570	}
1571
1572	pdmacFileAioMgrNormalRequestFree(pAioMgr, hReq);
1573
1574	/* Free the lock and process pending tasks if necessary */
1575	pTasksWaiting = pdmacFileAioMgrNormalRangeLockFree(pAioMgr, pEndpoint, pTask->pRangeLock);
1576	if (pTasksWaiting)
1577	{
1578	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksWaiting, pAioMgr, pEndpoint);
1579	AssertRC(rc);
1580	}
1581
1582	/* Call completion callback */
1583	LogFlow(("Task=%#p completed with %Rrc\n", pTask, rcReq));
1584	pTask->pfnCompleted(pTask, pTask->pvUser, rcReq);
1585	pdmacFileTaskFree(pEndpoint, pTask);
1586
1587	/*
1588	* If there is no request left on the endpoint but a flush request is set
1589	* it completed now and we notify the owner.
1590	* Furthermore we look for new requests and continue.
1591	*/
1592	if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq)
1593	{
1594	/* Call completion callback */
1595	pTask = pEndpoint->pFlushReq;
1596	pEndpoint->pFlushReq = NULL;
1597
1598	AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n"));
1599
1600	pTask->pfnCompleted(pTask, pTask->pvUser, VINF_SUCCESS);
1601	pdmacFileTaskFree(pEndpoint, pTask);
1602	}
1603	else if (RT_UNLIKELY(!pEndpoint->AioMgr.cRequestsActive && pEndpoint->AioMgr.fMoving))
1604	{
1605	/* If the endpoint is about to be migrated do it now. */
1606	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pEndpoint);
1607	Assert(!fReqsPending); NOREF(fReqsPending);
1608
1609	rc = pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
1610	AssertRC(rc);
1611	}
1612	}
1613	} /* Not a flush request */
1614	} /* request completed successfully */
1615	}
1616
1617	/** Helper macro for checking for error codes. */
1618	#define CHECK_RC(pAioMgr, rc) \
1619	if (RT_FAILURE(rc)) \
1620	{\
1621	int rc2 = pdmacFileAioMgrNormalErrorHandler(pAioMgr, rc, RT_SRC_POS);\
1622	return rc2;\
1623	}
1624
1625	/**
1626	* The normal I/O manager using the RTFileAio* API
1627	*
1628	* @returns VBox status code.
1629	* @param hThreadSelf Handle of the thread.
1630	* @param pvUser Opaque user data.
1631	*/
1632	DECLCALLBACK(int) pdmacFileAioMgrNormal(RTTHREAD hThreadSelf, void *pvUser)
1633	{
1634	int rc = VINF_SUCCESS;
1635	PPDMACEPFILEMGR pAioMgr = (PPDMACEPFILEMGR)pvUser;
1636	uint64_t uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
1637	NOREF(hThreadSelf);
1638
1639	while ( pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING
1640	\|\| pAioMgr->enmState == PDMACEPFILEMGRSTATE_SUSPENDING
1641	\|\| pAioMgr->enmState == PDMACEPFILEMGRSTATE_GROWING)
1642	{
1643	if (!pAioMgr->cRequestsActive)
1644	{
1645	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, true);
1646	if (!ASMAtomicReadBool(&pAioMgr->fWokenUp))
1647	rc = RTSemEventWait(pAioMgr->EventSem, pAioMgr->msBwLimitExpired);
1648	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, false);
1649	Assert(RT_SUCCESS(rc) \|\| rc == VERR_TIMEOUT);
1650
1651	LogFlow(("Got woken up\n"));
1652	ASMAtomicWriteBool(&pAioMgr->fWokenUp, false);
1653	}
1654
1655	/* Check for an external blocking event first. */
1656	if (pAioMgr->fBlockingEventPending)
1657	{
1658	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
1659	CHECK_RC(pAioMgr, rc);
1660	}
1661
1662	if (RT_LIKELY( pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING
1663	\|\| pAioMgr->enmState == PDMACEPFILEMGRSTATE_GROWING))
1664	{
1665	/* We got woken up because an endpoint issued new requests. Queue them. */
1666	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
1667	CHECK_RC(pAioMgr, rc);
1668
1669	while (pAioMgr->cRequestsActive)
1670	{
1671	RTFILEAIOREQ apReqs[20];
1672	uint32_t cReqsCompleted = 0;
1673	size_t cReqsWait;
1674
1675	if (pAioMgr->cRequestsActive > RT_ELEMENTS(apReqs))
1676	cReqsWait = RT_ELEMENTS(apReqs);
1677	else
1678	cReqsWait = pAioMgr->cRequestsActive;
1679
1680	LogFlow(("Waiting for %d of %d tasks to complete\n", 1, cReqsWait));
1681
1682	rc = RTFileAioCtxWait(pAioMgr->hAioCtx,
1683	1,
1684	RT_INDEFINITE_WAIT, apReqs,
1685	cReqsWait, &cReqsCompleted);
1686	if (RT_FAILURE(rc) && (rc != VERR_INTERRUPTED))
1687	CHECK_RC(pAioMgr, rc);
1688
1689	LogFlow(("%d tasks completed\n", cReqsCompleted));
1690
1691	for (uint32_t i = 0; i < cReqsCompleted; i++)
1692	pdmacFileAioMgrNormalReqComplete(pAioMgr, apReqs[i]);
1693
1694	/* Check for an external blocking event before we go to sleep again. */
1695	if (pAioMgr->fBlockingEventPending)
1696	{
1697	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
1698	CHECK_RC(pAioMgr, rc);
1699	}
1700
1701	/* Update load statistics. */
1702	uint64_t uMillisCurr = RTTimeMilliTS();
1703	if (uMillisCurr > uMillisEnd)
1704	{
1705	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointCurr = pAioMgr->pEndpointsHead;
1706
1707	/* Calculate timespan. */
1708	uMillisCurr -= uMillisEnd;
1709
1710	while (pEndpointCurr)
1711	{
1712	pEndpointCurr->AioMgr.cReqsPerSec = pEndpointCurr->AioMgr.cReqsProcessed / (uMillisCurr + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD);
1713	pEndpointCurr->AioMgr.cReqsProcessed = 0;
1714	pEndpointCurr = pEndpointCurr->AioMgr.pEndpointNext;
1715	}
1716
1717	/* Set new update interval */
1718	uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
1719	}
1720
1721	/* Check endpoints for new requests. */
1722	if (pAioMgr->enmState != PDMACEPFILEMGRSTATE_GROWING)
1723	{
1724	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
1725	CHECK_RC(pAioMgr, rc);
1726	}
1727	} /* while requests are active. */
1728
1729	if (pAioMgr->enmState == PDMACEPFILEMGRSTATE_GROWING)
1730	{
1731	rc = pdmacFileAioMgrNormalGrow(pAioMgr);
1732	AssertRC(rc);
1733	Assert(pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING);
1734
1735	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
1736	CHECK_RC(pAioMgr, rc);
1737	}
1738	} /* if still running */
1739	} /* while running */
1740
1741	LogFlowFunc(("rc=%Rrc\n", rc));
1742	return rc;
1743	}
1744
1745	#undef CHECK_RC
1746

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PDMAsyncCompletionFileNormal.cpp@ 96854

Download in other formats: