PDMAsyncCompletionFileNormal.cpp@ 22949

Last change on this file since 22949 was 22851, checked in by vboxsync, 15 years ago
Fixes for async I/O manager
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 33.0 KB

Line
1	/* $Id: PDMAsyncCompletionFileNormal.cpp 22851 2009-09-08 23:38:47Z vboxsync $ */
2	/** @file
3	* PDM Async I/O - Transport data asynchronous in R3 using EMT.
4	* Async File I/O manager.
5	*/
6
7	/*
8	* Copyright (C) 2006-2008 Sun Microsystems, Inc.
9	*
10	* This file is part of VirtualBox Open Source Edition (OSE), as
11	* available from http://www.virtualbox.org. This file is free software;
12	* you can redistribute it and/or modify it under the terms of the GNU
13	* General Public License (GPL) as published by the Free Software
14	* Foundation, in version 2 as it comes in the "COPYING" file of the
15	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
16	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
17	*
18	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
19	* Clara, CA 95054 USA or visit http://www.sun.com if you need
20	* additional information or have any questions.
21	*/
22	#define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION
23	#include <iprt/types.h>
24	#include <iprt/asm.h>
25	#include <iprt/file.h>
26	#include <iprt/mem.h>
27	#include <iprt/string.h>
28	#include <VBox/log.h>
29
30	#include "PDMAsyncCompletionFileInternal.h"
31
32	/** The update period for the I/O load statistics in ms. */
33	#define PDMACEPFILEMGR_LOAD_UPDATE_PERIOD 1000
34	/** Maximum number of requests a manager will handle. */
35	#define PDMACEPFILEMGR_REQS_MAX 512 /* @todo: Find better solution wrt. the request number*/
36
37	int pdmacFileAioMgrNormalInit(PPDMACEPFILEMGR pAioMgr)
38	{
39	int rc = VINF_SUCCESS;
40
41	rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, RTFILEAIO_UNLIMITED_REQS);
42	if (rc == VERR_OUT_OF_RANGE)
43	rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, PDMACEPFILEMGR_REQS_MAX);
44
45	if (RT_SUCCESS(rc))
46	{
47	/* Initialize request handle array. */
48	pAioMgr->iFreeEntryNext = 0;
49	pAioMgr->iFreeReqNext = 0;
50	pAioMgr->cReqEntries = PDMACEPFILEMGR_REQS_MAX + 1;
51	pAioMgr->pahReqsFree = (RTFILEAIOREQ )RTMemAllocZ(pAioMgr->cReqEntries sizeof(RTFILEAIOREQ));
52
53	if (pAioMgr->pahReqsFree)
54	{
55	return VINF_SUCCESS;
56	}
57	else
58	{
59	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
60	rc = VERR_NO_MEMORY;
61	}
62	}
63
64	return rc;
65	}
66
67	void pdmacFileAioMgrNormalDestroy(PPDMACEPFILEMGR pAioMgr)
68	{
69	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
70
71	while (pAioMgr->iFreeReqNext != pAioMgr->iFreeEntryNext)
72	{
73	RTFileAioReqDestroy(pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext]);
74	pAioMgr->iFreeReqNext = (pAioMgr->iFreeReqNext + 1) % pAioMgr->cReqEntries;
75	}
76
77	RTMemFree(pAioMgr->pahReqsFree);
78	}
79
80	/**
81	* Sorts the endpoint list with insertion sort.
82	*/
83	static void pdmacFileAioMgrNormalEndpointsSortByLoad(PPDMACEPFILEMGR pAioMgr)
84	{
85	PPDMASYNCCOMPLETIONENDPOINTFILE pEpPrev, pEpCurr, pEpNextToSort;
86
87	pEpPrev = pAioMgr->pEndpointsHead;
88	pEpCurr = pEpPrev->AioMgr.pEndpointNext;
89
90	while (pEpCurr)
91	{
92	/* Remember the next element to sort because the list might change. */
93	pEpNextToSort = pEpCurr->AioMgr.pEndpointNext;
94
95	/* Unlink the current element from the list. */
96	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
97	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
98
99	if (pPrev)
100	pPrev->AioMgr.pEndpointNext = pNext;
101	else
102	pAioMgr->pEndpointsHead = pNext;
103
104	if (pNext)
105	pNext->AioMgr.pEndpointPrev = pPrev;
106
107	/* Go back until we reached the place to insert the current endpoint into. */
108	while (pEpPrev && (pEpPrev->AioMgr.cReqsPerSec < pEpCurr->AioMgr.cReqsPerSec))
109	pEpPrev = pEpPrev->AioMgr.pEndpointPrev;
110
111	/* Link the endpoint into the list. */
112	if (pEpPrev)
113	pNext = pEpPrev->AioMgr.pEndpointNext;
114	else
115	pNext = pAioMgr->pEndpointsHead;
116
117	pEpCurr->AioMgr.pEndpointNext = pNext;
118	pEpCurr->AioMgr.pEndpointPrev = pEpPrev;
119	pNext->AioMgr.pEndpointPrev = pEpCurr;
120	if (pEpPrev)
121	pEpPrev->AioMgr.pEndpointNext = pEpCurr;
122	else
123	pAioMgr->pEndpointsHead = pEpCurr;
124
125	pEpCurr = pEpNextToSort;
126	}
127
128	#ifdef DEBUG
129	/* Validate sorting alogrithm */
130	unsigned cEndpoints = 0;
131	pEpCurr = pAioMgr->pEndpointsHead;
132
133	AssertMsg(pEpCurr, ("No endpoint in the list?\n"));
134	AssertMsg(!pEpCurr->AioMgr.pEndpointPrev, ("First element in the list points to previous element\n"));
135
136	while (pEpCurr)
137	{
138	cEndpoints++;
139
140	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
141	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
142
143	Assert(!pNext \|\| pNext->AioMgr.cReqsPerSec <= pEpCurr->AioMgr.cReqsPerSec);
144	Assert(!pPrev \|\| pPrev->AioMgr.cReqsPerSec >= pEpCurr->AioMgr.cReqsPerSec);
145
146	pEpCurr = pNext;
147	}
148
149	AssertMsg(cEndpoints == pAioMgr->cEndpoints, ("Endpoints lost during sort!\n"));
150
151	#endif
152	}
153
154	/**
155	* Removes an endpoint from the currently assigned manager.
156	*
157	* @returns TRUE if there are still requests pending on the current manager for this endpoint.
158	* FALSE otherwise.
159	* @param pEndpointRemove The endpoint to remove.
160	*/
161	static bool pdmacFileAioMgrNormalRemoveEndpoint(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove)
162	{
163	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEndpointRemove->AioMgr.pEndpointPrev;
164	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEndpointRemove->AioMgr.pEndpointNext;
165	PPDMACEPFILEMGR pAioMgr = pEndpointRemove->pAioMgr;
166
167	pAioMgr->cEndpoints--;
168
169	if (pPrev)
170	pPrev->AioMgr.pEndpointNext = pNext;
171	else
172	pAioMgr->pEndpointsHead = pNext;
173
174	if (pNext)
175	pNext->AioMgr.pEndpointPrev = pPrev;
176
177	/* Make sure that there is no request pending on this manager for the endpoint. */
178	if (!pEndpointRemove->AioMgr.cRequestsActive)
179	{
180	Assert(!pEndpointRemove->pFlushReq);
181
182	/* Reopen the file so that the new endpoint can reassociate with the file */
183	RTFileClose(pEndpointRemove->File);
184	int rc = RTFileOpen(&pEndpointRemove->File, pEndpointRemove->Core.pszUri, pEndpointRemove->fFlags);
185	AssertRC(rc);
186	return false;
187	}
188
189	return true;
190	}
191
192	/**
193	* Creates a new I/O manager and spreads the I/O load of the endpoints
194	* between the given I/O manager and the new one.
195	*
196	* @returns nothing.
197	* @param pAioMgr The I/O manager with high I/O load.
198	*/
199	static void pdmacFileAioMgrNormalBalanceLoad(PPDMACEPFILEMGR pAioMgr)
200	{
201	PPDMACEPFILEMGR pAioMgrNew = NULL;
202	int rc = VINF_SUCCESS;
203
204	/* Splitting can't be done with only one open endpoint. */
205	if (pAioMgr->cEndpoints > 1)
206	{
207	rc = pdmacFileAioMgrCreate((PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass,
208	&pAioMgrNew);
209	if (RT_SUCCESS(rc))
210	{
211	/* We will sort the list by request count per second. */
212	pdmacFileAioMgrNormalEndpointsSortByLoad(pAioMgr);
213
214	/* Now move some endpoints to the new manager. */
215	unsigned cReqsHere = pAioMgr->pEndpointsHead->AioMgr.cReqsPerSec;
216	unsigned cReqsOther = 0;
217	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead->AioMgr.pEndpointNext;
218
219	while (pCurr)
220	{
221	if (cReqsHere <= cReqsOther)
222	{
223	/*
224	* The other manager has more requests to handle now.
225	* We will keep the current endpoint.
226	*/
227	Log(("Keeping endpoint %#p{%s} with %u reqs/s\n", pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
228	cReqsHere += pCurr->AioMgr.cReqsPerSec;
229	pCurr = pCurr->AioMgr.pEndpointNext;
230	}
231	else
232	{
233	/* Move to other endpoint. */
234	Log(("Moving endpoint %#p{%s} with %u reqs/s to other manager\n", pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
235	cReqsOther += pCurr->AioMgr.cReqsPerSec;
236
237	PPDMASYNCCOMPLETIONENDPOINTFILE pMove = pCurr;
238
239	pCurr = pCurr->AioMgr.pEndpointNext;
240
241	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pMove);
242
243	if (fReqsPending)
244	{
245	pMove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
246	pMove->AioMgr.fMoving = true;
247	pMove->AioMgr.pAioMgrDst = pAioMgrNew;
248	}
249	else
250	{
251	pMove->AioMgr.fMoving = false;
252	pMove->AioMgr.pAioMgrDst = NULL;
253	pdmacFileAioMgrAddEndpoint(pAioMgrNew, pMove);
254	}
255	}
256	}
257	}
258	else
259	{
260	/* Don't process further but leave a log entry about reduced performance. */
261	LogRel(("AIOMgr: Could not create new I/O manager (rc=%Rrc). Expect reduced performance\n", rc));
262	}
263	}
264	}
265
266	/**
267	* Error handler which will create the failsafe managers and destroy the failed I/O manager.
268	*
269	* @returns VBox status code
270	* @param pAioMgr The I/O manager the error ocurred on.
271	* @param rc The error code.
272	*/
273	static int pdmacFileAioMgrNormalErrorHandler(PPDMACEPFILEMGR pAioMgr, int rc, RT_SRC_POS_DECL)
274	{
275	LogRel(("AIOMgr: I/O manager %#p encountered a critical error (rc=%Rrc) during operation. Falling back to failsafe mode. Expect reduced performance\n",
276	pAioMgr, rc));
277	LogRel(("AIOMgr: Error happened in %s:(%u){%s}\n", RT_SRC_POS_ARGS));
278	LogRel(("AIOMgr: Please contact the product vendor\n"));
279
280	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass;
281
282	pAioMgr->enmState = PDMACEPFILEMGRSTATE_FAULT;
283	ASMAtomicWriteBool(&pEpClassFile->fFailsafe, true);
284
285	AssertMsgFailed(("Implement\n"));
286	return VINF_SUCCESS;
287	}
288
289	static int pdmacFileAioMgrNormalProcessTaskList(PPDMACTASKFILE pTaskHead,
290	PPDMACEPFILEMGR pAioMgr,
291	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
292	{
293	RTFILEAIOREQ apReqs[20];
294	unsigned cRequests = 0;
295	unsigned cMaxRequests = PDMACEPFILEMGR_REQS_MAX - pAioMgr->cRequestsActive;
296	int rc = VINF_SUCCESS;
297	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
298
299	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
300	("Trying to process request lists of a non active endpoint!\n"));
301
302	/* Go through the list and queue the requests until we get a flush request */
303	while (pTaskHead && !pEndpoint->pFlushReq && (cMaxRequests > 0))
304	{
305	PPDMACTASKFILE pCurr = pTaskHead;
306
307	pTaskHead = pTaskHead->pNext;
308
309	AssertMsg(VALID_PTR(pCurr->pEndpoint) && (pCurr->pEndpoint == pEndpoint),
310	("Endpoints do not match\n"));
311
312	switch (pCurr->enmTransferType)
313	{
314	case PDMACTASKFILETRANSFER_FLUSH:
315	{
316	/* If there is no data transfer request this flush request finished immediately. */
317	if (!pEndpoint->AioMgr.cRequestsActive)
318	{
319	pCurr->pfnCompleted(pCurr, pCurr->pvUser);
320	pdmacFileTaskFree(pEndpoint, pCurr);
321	}
322	else
323	{
324	pEndpoint->pFlushReq = pCurr;
325
326	if (pTaskHead)
327	{
328	/* Add the rest of the tasks to the pending list */
329	if (!pEndpoint->AioMgr.pReqsPendingHead)
330	{
331	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
332	pEndpoint->AioMgr.pReqsPendingHead = pTaskHead;
333	}
334	else
335	{
336	Assert(pEndpoint->AioMgr.pReqsPendingTail);
337	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTaskHead;
338	}
339
340	/* Update the tail. */
341	while (pTaskHead->pNext)
342	pTaskHead = pTaskHead->pNext;
343
344	pEndpoint->AioMgr.pReqsPendingTail = pTaskHead;
345	}
346	}
347	break;
348	}
349	case PDMACTASKFILETRANSFER_READ:
350	case PDMACTASKFILETRANSFER_WRITE:
351	{
352	RTFILEAIOREQ hReq = NIL_RTFILEAIOREQ;
353	void *pvBuf = pCurr->DataSeg.pvSeg;
354
355	/* Get a request handle. */
356	if (pAioMgr->iFreeReqNext != pAioMgr->iFreeEntryNext)
357	{
358	hReq = pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext];
359	pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext] = NIL_RTFILEAIOREQ;
360	pAioMgr->iFreeReqNext = (pAioMgr->iFreeReqNext + 1) % pAioMgr->cReqEntries;
361	}
362	else
363	{
364	rc = RTFileAioReqCreate(&hReq);
365	AssertRC(rc);
366	}
367
368	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
369
370	/* Check if the alignment requirements are met.
371	* Offset, transfer size and buffer address
372	* need to be on a 512 boundary. */
373	size_t cbToTransfer = RT_ALIGN_Z(pCurr->DataSeg.cbSeg, 512);
374	RTFOFF OffStart = pCurr->Off & ~(RTFOFF)(512-1);
375	PDMACTASKFILETRANSFER enmTransferType = pCurr->enmTransferType;
376
377	AssertMsg(( (pCurr->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
378	\|\| (OffStart + cbToTransfer <= pEndpoint->cbFile)),
379	("Read exceeds file size OffStart=%RTfoff cbToTransfer=%d cbFile=%llu\n",
380	OffStart, cbToTransfer, pEndpoint->cbFile));
381
382	pCurr->fPrefetch = false;
383
384	if ( RT_UNLIKELY(cbToTransfer != pCurr->DataSeg.cbSeg)
385	\|\| RT_UNLIKELY(OffStart != pCurr->Off)
386	\|\| ((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) != (RTR3UINTPTR)pvBuf))
387	{
388	/* Create bounce buffer. */
389	pCurr->fBounceBuffer = true;
390
391	AssertMsg(pCurr->Off >= OffStart, ("Overflow in calculation Off=%llu OffStart=%llu\n",
392	pCurr->Off, OffStart));
393	pCurr->uBounceBufOffset = pCurr->Off - OffStart;
394
395	/** @todo: I think we need something like a RTMemAllocAligned method here.
396	* Current assumption is that the maximum alignment is 4096byte
397	* (GPT disk on Windows)
398	* so we can use RTMemPageAlloc here.
399	*/
400	pCurr->pvBounceBuffer = RTMemPageAlloc(cbToTransfer);
401	AssertPtr(pCurr->pvBounceBuffer);
402	pvBuf = pCurr->pvBounceBuffer;
403
404	if (pCurr->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
405	{
406	if ( RT_UNLIKELY(cbToTransfer != pCurr->DataSeg.cbSeg)
407	\|\| RT_UNLIKELY(OffStart != pCurr->Off))
408	{
409	/* We have to fill the buffer first before we can update the data. */
410	pCurr->fPrefetch = true;
411	enmTransferType = PDMACTASKFILETRANSFER_READ;
412	}
413	else
414	memcpy(pvBuf, pCurr->DataSeg.pvSeg, pCurr->DataSeg.cbSeg);
415	}
416	}
417	else
418	pCurr->fBounceBuffer = false;
419
420	AssertMsg((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) == (RTR3UINTPTR)pvBuf,
421	("AIO: Alignment restrictions not met!\n"));
422
423	if (enmTransferType == PDMACTASKFILETRANSFER_WRITE)
424	{
425	/* Grow the file if needed. */
426	if (RT_UNLIKELY((pCurr->Off + pCurr->DataSeg.cbSeg) > pEndpoint->cbFile))
427	{
428	ASMAtomicWriteU64(&pEndpoint->cbFile, pCurr->Off + pCurr->DataSeg.cbSeg);
429	RTFileSetSize(pEndpoint->File, pCurr->Off + pCurr->DataSeg.cbSeg);
430	}
431
432	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->File,
433	OffStart, pvBuf, cbToTransfer, pCurr);
434	}
435	else
436	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->File,
437	OffStart, pvBuf, cbToTransfer, pCurr);
438	AssertRC(rc);
439
440	apReqs[cRequests] = hReq;
441	pEndpoint->AioMgr.cReqsProcessed++;
442	cMaxRequests--;
443	cRequests++;
444	if (cRequests == RT_ELEMENTS(apReqs))
445	{
446	pAioMgr->cRequestsActive += cRequests;
447	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, apReqs, cRequests);
448	if (RT_FAILURE(rc))
449	{
450	/* @todo implement */
451	AssertMsgFailed(("Implement\n"));
452	}
453
454	cRequests = 0;
455	}
456	break;
457	}
458	default:
459	AssertMsgFailed(("Invalid transfer type %d\n", pCurr->enmTransferType));
460	}
461	}
462
463	if (cRequests)
464	{
465	pAioMgr->cRequestsActive += cRequests;
466	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, apReqs, cRequests);
467	AssertMsgReturn(RT_SUCCESS(rc), ("Could not submit %u requests %Rrc\n", cRequests, rc), rc);
468	}
469
470	if (RT_UNLIKELY(!cMaxRequests && pTaskHead && !pEndpoint->pFlushReq))
471	{
472	/*
473	* The I/O manager has no room left for more requests
474	* but there are still requests to process.
475	* Create a new I/O manager and let it handle some endpoints.
476	*/
477
478	/* Add the rest of the tasks to the pending list first */
479	if (!pEndpoint->AioMgr.pReqsPendingHead)
480	{
481	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
482	pEndpoint->AioMgr.pReqsPendingHead = pTaskHead;
483	}
484	else
485	{
486	Assert(pEndpoint->AioMgr.pReqsPendingTail);
487	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTaskHead;
488	}
489
490	/* Update the tail. */
491	while (pTaskHead->pNext)
492	pTaskHead = pTaskHead->pNext;
493
494	pEndpoint->AioMgr.pReqsPendingTail = pTaskHead;
495
496	pdmacFileAioMgrNormalBalanceLoad(pAioMgr);
497	}
498
499	return rc;
500	}
501
502	/**
503	* Adds all pending requests for the given endpoint
504	* until a flush request is encountered or there is no
505	* request anymore.
506	*
507	* @returns VBox status code.
508	* @param pAioMgr The async I/O manager for the endpoint
509	* @param pEndpoint The endpoint to get the requests from.
510	*/
511	static int pdmacFileAioMgrNormalQueueReqs(PPDMACEPFILEMGR pAioMgr,
512	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
513	{
514	int rc = VINF_SUCCESS;
515	PPDMACTASKFILE pTasksHead = NULL;
516
517	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
518	("Trying to process request lists of a non active endpoint!\n"));
519
520	Assert(!pEndpoint->pFlushReq);
521
522	/* Check the pending list first */
523	if (pEndpoint->AioMgr.pReqsPendingHead)
524	{
525	pTasksHead = pEndpoint->AioMgr.pReqsPendingHead;
526	/*
527	* Clear the list as the processing routine will insert them into the list
528	* again if it gets a flush request.
529	*/
530	pEndpoint->AioMgr.pReqsPendingHead = NULL;
531	pEndpoint->AioMgr.pReqsPendingTail = NULL;
532	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
533	AssertRC(rc);
534	}
535
536	if (!pEndpoint->pFlushReq)
537	{
538	/* Now the request queue. */
539	pTasksHead = pdmacFileEpGetNewTasks(pEndpoint);
540	if (pTasksHead)
541	{
542	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
543	AssertRC(rc);
544	}
545	}
546
547	return rc;
548	}
549
550	static int pdmacFileAioMgrNormalProcessBlockingEvent(PPDMACEPFILEMGR pAioMgr)
551	{
552	int rc = VINF_SUCCESS;
553	bool fNotifyWaiter = false;
554
555	Assert(pAioMgr->fBlockingEventPending);
556
557	switch (pAioMgr->enmBlockingEvent)
558	{
559	case PDMACEPFILEAIOMGRBLOCKINGEVENT_ADD_ENDPOINT:
560	{
561	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointNew = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint);
562	AssertMsg(VALID_PTR(pEndpointNew), ("Adding endpoint event without a endpoint to add\n"));
563
564	pEndpointNew->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE;
565
566	pEndpointNew->AioMgr.pEndpointNext = pAioMgr->pEndpointsHead;
567	pEndpointNew->AioMgr.pEndpointPrev = NULL;
568	if (pAioMgr->pEndpointsHead)
569	pAioMgr->pEndpointsHead->AioMgr.pEndpointPrev = pEndpointNew;
570	pAioMgr->pEndpointsHead = pEndpointNew;
571
572	/* Assign the completion point to this file. */
573	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pEndpointNew->File);
574	fNotifyWaiter = true;
575	pAioMgr->cEndpoints++;
576	break;
577	}
578	case PDMACEPFILEAIOMGRBLOCKINGEVENT_REMOVE_ENDPOINT:
579	{
580	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint);
581	AssertMsg(VALID_PTR(pEndpointRemove), ("Removing endpoint event without a endpoint to remove\n"));
582
583	pEndpointRemove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
584	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointRemove);
585	break;
586	}
587	case PDMACEPFILEAIOMGRBLOCKINGEVENT_CLOSE_ENDPOINT:
588	{
589	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointClose = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint);
590	AssertMsg(VALID_PTR(pEndpointClose), ("Close endpoint event without a endpoint to close\n"));
591
592	/* Make sure all tasks finished. Process the queues a last time first. */
593	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpointClose);
594	AssertRC(rc);
595
596	pEndpointClose->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING;
597	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointClose);
598	break;
599	}
600	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SHUTDOWN:
601	{
602	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SHUTDOWN;
603	if (!pAioMgr->cRequestsActive)
604	fNotifyWaiter = true;
605	break;
606	}
607	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SUSPEND:
608	{
609	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SUSPENDING;
610	break;
611	}
612	case PDMACEPFILEAIOMGRBLOCKINGEVENT_RESUME:
613	{
614	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
615	fNotifyWaiter = true;
616	break;
617	}
618	default:
619	AssertReleaseMsgFailed(("Invalid event type %d\n", pAioMgr->enmBlockingEvent));
620	}
621
622	if (fNotifyWaiter)
623	{
624	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
625	pAioMgr->enmBlockingEvent = PDMACEPFILEAIOMGRBLOCKINGEVENT_INVALID;
626
627	/* Release the waiting thread. */
628	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
629	AssertRC(rc);
630	}
631
632	return rc;
633	}
634
635	/** Helper macro for checking for error codes. */
636	#define CHECK_RC(pAioMgr, rc) \
637	if (RT_FAILURE(rc)) \
638	{\
639	int rc2 = pdmacFileAioMgrNormalErrorHandler(pAioMgr, rc, RT_SRC_POS);\
640	return rc2;\
641	}
642
643	/**
644	* The normal I/O manager using the RTFileAio* API
645	*
646	* @returns VBox status code.
647	* @param ThreadSelf Handle of the thread.
648	* @param pvUser Opaque user data.
649	*/
650	int pdmacFileAioMgrNormal(RTTHREAD ThreadSelf, void *pvUser)
651	{
652	int rc = VINF_SUCCESS;
653	PPDMACEPFILEMGR pAioMgr = (PPDMACEPFILEMGR)pvUser;
654	uint64_t uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
655
656	while ( (pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING)
657	\|\| (pAioMgr->enmState == PDMACEPFILEMGRSTATE_SUSPENDING))
658	{
659	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, true);
660	if (!ASMAtomicReadBool(&pAioMgr->fWokenUp))
661	rc = RTSemEventWait(pAioMgr->EventSem, RT_INDEFINITE_WAIT);
662	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, false);
663	AssertRC(rc);
664
665	LogFlow(("Got woken up\n"));
666	ASMAtomicWriteBool(&pAioMgr->fWokenUp, false);
667
668	/* Check for an external blocking event first. */
669	if (pAioMgr->fBlockingEventPending)
670	{
671	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
672	CHECK_RC(pAioMgr, rc);
673	}
674
675	if (RT_LIKELY(pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING))
676	{
677	/* Check the assigned endpoints for new tasks if there isn't a flush request active at the moment. */
678	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint = pAioMgr->pEndpointsHead;
679
680	while (pEndpoint)
681	{
682	if (!pEndpoint->pFlushReq && (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE))
683	{
684	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
685	CHECK_RC(pAioMgr, rc);
686	}
687
688	pEndpoint = pEndpoint->AioMgr.pEndpointNext;
689	}
690
691	while (pAioMgr->cRequestsActive)
692	{
693	RTFILEAIOREQ apReqs[20];
694	uint32_t cReqsCompleted = 0;
695	size_t cReqsWait;
696
697	if (pAioMgr->cRequestsActive > RT_ELEMENTS(apReqs))
698	cReqsWait = RT_ELEMENTS(apReqs);
699	else
700	cReqsWait = pAioMgr->cRequestsActive;
701
702	rc = RTFileAioCtxWait(pAioMgr->hAioCtx,
703	cReqsWait,
704	RT_INDEFINITE_WAIT, apReqs,
705	RT_ELEMENTS(apReqs), &cReqsCompleted);
706	if (RT_FAILURE(rc) && (rc != VERR_INTERRUPTED))
707	CHECK_RC(pAioMgr, rc);
708
709	for (uint32_t i = 0; i < cReqsCompleted; i++)
710	{
711	size_t cbTransfered = 0;
712	int rcReq = RTFileAioReqGetRC(apReqs[i], &cbTransfered);
713	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(apReqs[i]);
714
715	pEndpoint = pTask->pEndpoint;
716
717	AssertMsg( RT_SUCCESS(rcReq)
718	&& ( (cbTransfered == pTask->DataSeg.cbSeg)
719	\|\| (pTask->fBounceBuffer)),
720	("Task didn't completed successfully (rc=%Rrc) or was incomplete (cbTransfered=%u)\n", rc, cbTransfered));
721
722	if (pTask->fPrefetch)
723	{
724	Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE);
725	Assert(pTask->fBounceBuffer);
726
727	memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
728	pTask->DataSeg.pvSeg,
729	pTask->DataSeg.cbSeg);
730
731	/* Write it now. */
732	pTask->fPrefetch = false;
733	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg, 512);
734	RTFOFF OffStart = pTask->Off & ~(RTFOFF)(512-1);
735
736	/* Grow the file if needed. */
737	if (RT_UNLIKELY((pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
738	{
739	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
740	RTFileSetSize(pEndpoint->File, pTask->Off + pTask->DataSeg.cbSeg);
741	}
742
743	rc = RTFileAioReqPrepareWrite(apReqs[i], pEndpoint->File,
744	OffStart, pTask->pvBounceBuffer, cbToTransfer, pTask);
745	AssertRC(rc);
746	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, &apReqs[i], 1);
747	AssertRC(rc);
748	}
749	else
750	{
751	if (pTask->fBounceBuffer)
752	{
753	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
754	memcpy(pTask->DataSeg.pvSeg,
755	((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
756	pTask->DataSeg.cbSeg);
757
758	RTMemPageFree(pTask->pvBounceBuffer);
759	}
760
761	/* Put the entry on the free array */
762	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i];
763	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) %pAioMgr->cReqEntries;
764
765	pAioMgr->cRequestsActive--;
766	pEndpoint->AioMgr.cReqsProcessed++;
767
768	/* Call completion callback */
769	pTask->pfnCompleted(pTask, pTask->pvUser);
770	pdmacFileTaskFree(pEndpoint, pTask);
771
772	/*
773	* If there is no request left on the endpoint but a flush request is set
774	* it completed now and we notify the owner.
775	* Furthermore we look for new requests and continue.
776	*/
777	if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq)
778	{
779	/* Call completion callback */
780	pTask = pEndpoint->pFlushReq;
781	pEndpoint->pFlushReq = NULL;
782
783	AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n"));
784
785	pTask->pfnCompleted(pTask, pTask->pvUser);
786	pdmacFileTaskFree(pEndpoint, pTask);
787	}
788	}
789
790	if (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
791	{
792	if (!pEndpoint->pFlushReq)
793	{
794	/* Check if there are events on the endpoint. */
795	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
796	CHECK_RC(pAioMgr, rc);
797	}
798	}
799	else if (!pEndpoint->AioMgr.cRequestsActive)
800	{
801	/* Reopen the file so that the new endpoint can reassociate with the file */
802	RTFileClose(pEndpoint->File);
803	rc = RTFileOpen(&pEndpoint->File, pEndpoint->Core.pszUri, pEndpoint->fFlags);
804	AssertRC(rc);
805
806	if (pEndpoint->AioMgr.fMoving)
807	{
808	pEndpoint->AioMgr.fMoving = false;
809	pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
810	}
811	else
812	{
813	Assert(pAioMgr->fBlockingEventPending);
814	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
815
816	/* Release the waiting thread. */
817	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
818	AssertRC(rc);
819	}
820	}
821	}
822
823	/* Check for an external blocking event before we go to sleep again. */
824	if (pAioMgr->fBlockingEventPending)
825	{
826	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
827	CHECK_RC(pAioMgr, rc);
828	}
829
830	/* Update load statistics. */
831	uint64_t uMillisCurr = RTTimeMilliTS();
832	if (uMillisCurr > uMillisEnd)
833	{
834	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointCurr = pAioMgr->pEndpointsHead;
835
836	/* Calculate timespan. */
837	uMillisCurr -= uMillisEnd;
838
839	while (pEndpointCurr)
840	{
841	pEndpointCurr->AioMgr.cReqsPerSec = pEndpointCurr->AioMgr.cReqsProcessed / (uMillisCurr + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD);
842	pEndpointCurr->AioMgr.cReqsProcessed = 0;
843	pEndpointCurr = pEndpointCurr->AioMgr.pEndpointNext;
844	}
845
846	/* Set new update interval */
847	uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
848	}
849	}
850	}
851	}
852
853	return rc;
854	}
855
856	#undef CHECK_RC
857

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

source: vbox/trunk/src/VBox/VMM/PDMAsyncCompletionFileNormal.cpp@ 22949

Download in other formats: