PDMAsyncCompletionFileNormal.cpp@ 22997

Last change on this file since 22997 was 22977, checked in by vboxsync, 15 years ago
PDMASyncCompletionFile: uBitmaskAlignment fix.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 33.4 KB

Line
1	/* $Id: PDMAsyncCompletionFileNormal.cpp 22977 2009-09-13 13:00:19Z 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! pvBuf=%p uBitmaskAlignment=%p\n", pvBuf, pEpClassFile->uBitmaskAlignment));
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	pEndpoint->AioMgr.cRequestsActive += cRequests;
448	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, apReqs, cRequests);
449	if (RT_FAILURE(rc))
450	{
451	/* @todo implement */
452	AssertMsgFailed(("Implement\n"));
453	}
454
455	cRequests = 0;
456	}
457	break;
458	}
459	default:
460	AssertMsgFailed(("Invalid transfer type %d\n", pCurr->enmTransferType));
461	}
462	}
463
464	if (cRequests)
465	{
466	pAioMgr->cRequestsActive += cRequests;
467	pEndpoint->AioMgr.cRequestsActive += cRequests;
468	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, apReqs, cRequests);
469	AssertMsgReturn(RT_SUCCESS(rc), ("Could not submit %u requests %Rrc\n", cRequests, rc), rc);
470	}
471
472	if (RT_UNLIKELY(!cMaxRequests && pTaskHead && !pEndpoint->pFlushReq))
473	{
474	/*
475	* The I/O manager has no room left for more requests
476	* but there are still requests to process.
477	* Create a new I/O manager and let it handle some endpoints.
478	*/
479
480	/* Add the rest of the tasks to the pending list first */
481	if (!pEndpoint->AioMgr.pReqsPendingHead)
482	{
483	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
484	pEndpoint->AioMgr.pReqsPendingHead = pTaskHead;
485	}
486	else
487	{
488	Assert(pEndpoint->AioMgr.pReqsPendingTail);
489	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTaskHead;
490	}
491
492	/* Update the tail. */
493	while (pTaskHead->pNext)
494	pTaskHead = pTaskHead->pNext;
495
496	pEndpoint->AioMgr.pReqsPendingTail = pTaskHead;
497
498	pdmacFileAioMgrNormalBalanceLoad(pAioMgr);
499	}
500
501	return rc;
502	}
503
504	/**
505	* Adds all pending requests for the given endpoint
506	* until a flush request is encountered or there is no
507	* request anymore.
508	*
509	* @returns VBox status code.
510	* @param pAioMgr The async I/O manager for the endpoint
511	* @param pEndpoint The endpoint to get the requests from.
512	*/
513	static int pdmacFileAioMgrNormalQueueReqs(PPDMACEPFILEMGR pAioMgr,
514	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
515	{
516	int rc = VINF_SUCCESS;
517	PPDMACTASKFILE pTasksHead = NULL;
518
519	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
520	("Trying to process request lists of a non active endpoint!\n"));
521
522	Assert(!pEndpoint->pFlushReq);
523
524	/* Check the pending list first */
525	if (pEndpoint->AioMgr.pReqsPendingHead)
526	{
527	pTasksHead = pEndpoint->AioMgr.pReqsPendingHead;
528	/*
529	* Clear the list as the processing routine will insert them into the list
530	* again if it gets a flush request.
531	*/
532	pEndpoint->AioMgr.pReqsPendingHead = NULL;
533	pEndpoint->AioMgr.pReqsPendingTail = NULL;
534	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
535	AssertRC(rc);
536	}
537
538	if (!pEndpoint->pFlushReq)
539	{
540	/* Now the request queue. */
541	pTasksHead = pdmacFileEpGetNewTasks(pEndpoint);
542	if (pTasksHead)
543	{
544	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
545	AssertRC(rc);
546	}
547	}
548
549	return rc;
550	}
551
552	static int pdmacFileAioMgrNormalProcessBlockingEvent(PPDMACEPFILEMGR pAioMgr)
553	{
554	int rc = VINF_SUCCESS;
555	bool fNotifyWaiter = false;
556
557	Assert(pAioMgr->fBlockingEventPending);
558
559	switch (pAioMgr->enmBlockingEvent)
560	{
561	case PDMACEPFILEAIOMGRBLOCKINGEVENT_ADD_ENDPOINT:
562	{
563	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointNew = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint);
564	AssertMsg(VALID_PTR(pEndpointNew), ("Adding endpoint event without a endpoint to add\n"));
565
566	pEndpointNew->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE;
567
568	pEndpointNew->AioMgr.pEndpointNext = pAioMgr->pEndpointsHead;
569	pEndpointNew->AioMgr.pEndpointPrev = NULL;
570	if (pAioMgr->pEndpointsHead)
571	pAioMgr->pEndpointsHead->AioMgr.pEndpointPrev = pEndpointNew;
572	pAioMgr->pEndpointsHead = pEndpointNew;
573
574	/* Assign the completion point to this file. */
575	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pEndpointNew->File);
576	fNotifyWaiter = true;
577	pAioMgr->cEndpoints++;
578	break;
579	}
580	case PDMACEPFILEAIOMGRBLOCKINGEVENT_REMOVE_ENDPOINT:
581	{
582	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint);
583	AssertMsg(VALID_PTR(pEndpointRemove), ("Removing endpoint event without a endpoint to remove\n"));
584
585	pEndpointRemove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
586	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointRemove);
587	break;
588	}
589	case PDMACEPFILEAIOMGRBLOCKINGEVENT_CLOSE_ENDPOINT:
590	{
591	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointClose = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint);
592	AssertMsg(VALID_PTR(pEndpointClose), ("Close endpoint event without a endpoint to close\n"));
593
594	/* Make sure all tasks finished. Process the queues a last time first. */
595	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpointClose);
596	AssertRC(rc);
597
598	pEndpointClose->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING;
599	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointClose);
600	break;
601	}
602	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SHUTDOWN:
603	{
604	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SHUTDOWN;
605	if (!pAioMgr->cRequestsActive)
606	fNotifyWaiter = true;
607	break;
608	}
609	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SUSPEND:
610	{
611	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SUSPENDING;
612	break;
613	}
614	case PDMACEPFILEAIOMGRBLOCKINGEVENT_RESUME:
615	{
616	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
617	fNotifyWaiter = true;
618	break;
619	}
620	default:
621	AssertReleaseMsgFailed(("Invalid event type %d\n", pAioMgr->enmBlockingEvent));
622	}
623
624	if (fNotifyWaiter)
625	{
626	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
627	pAioMgr->enmBlockingEvent = PDMACEPFILEAIOMGRBLOCKINGEVENT_INVALID;
628
629	/* Release the waiting thread. */
630	LogFlow(("Signalling waiter\n"));
631	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
632	AssertRC(rc);
633	}
634
635	return rc;
636	}
637
638	/** Helper macro for checking for error codes. */
639	#define CHECK_RC(pAioMgr, rc) \
640	if (RT_FAILURE(rc)) \
641	{\
642	int rc2 = pdmacFileAioMgrNormalErrorHandler(pAioMgr, rc, RT_SRC_POS);\
643	return rc2;\
644	}
645
646	/**
647	* The normal I/O manager using the RTFileAio* API
648	*
649	* @returns VBox status code.
650	* @param ThreadSelf Handle of the thread.
651	* @param pvUser Opaque user data.
652	*/
653	int pdmacFileAioMgrNormal(RTTHREAD ThreadSelf, void *pvUser)
654	{
655	int rc = VINF_SUCCESS;
656	PPDMACEPFILEMGR pAioMgr = (PPDMACEPFILEMGR)pvUser;
657	uint64_t uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
658
659	while ( (pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING)
660	\|\| (pAioMgr->enmState == PDMACEPFILEMGRSTATE_SUSPENDING))
661	{
662	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, true);
663	if (!ASMAtomicReadBool(&pAioMgr->fWokenUp))
664	rc = RTSemEventWait(pAioMgr->EventSem, RT_INDEFINITE_WAIT);
665	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, false);
666	AssertRC(rc);
667
668	LogFlow(("Got woken up\n"));
669	ASMAtomicWriteBool(&pAioMgr->fWokenUp, false);
670
671	/* Check for an external blocking event first. */
672	if (pAioMgr->fBlockingEventPending)
673	{
674	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
675	CHECK_RC(pAioMgr, rc);
676	}
677
678	if (RT_LIKELY(pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING))
679	{
680	/* Check the assigned endpoints for new tasks if there isn't a flush request active at the moment. */
681	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint = pAioMgr->pEndpointsHead;
682
683	while (pEndpoint)
684	{
685	if (!pEndpoint->pFlushReq && (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE))
686	{
687	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
688	CHECK_RC(pAioMgr, rc);
689	}
690
691	pEndpoint = pEndpoint->AioMgr.pEndpointNext;
692	}
693
694	while (pAioMgr->cRequestsActive)
695	{
696	RTFILEAIOREQ apReqs[20];
697	uint32_t cReqsCompleted = 0;
698	size_t cReqsWait;
699
700	if (pAioMgr->cRequestsActive > RT_ELEMENTS(apReqs))
701	cReqsWait = RT_ELEMENTS(apReqs);
702	else
703	cReqsWait = pAioMgr->cRequestsActive;
704
705	rc = RTFileAioCtxWait(pAioMgr->hAioCtx,
706	cReqsWait,
707	RT_INDEFINITE_WAIT, apReqs,
708	RT_ELEMENTS(apReqs), &cReqsCompleted);
709	if (RT_FAILURE(rc) && (rc != VERR_INTERRUPTED))
710	CHECK_RC(pAioMgr, rc);
711
712	for (uint32_t i = 0; i < cReqsCompleted; i++)
713	{
714	size_t cbTransfered = 0;
715	int rcReq = RTFileAioReqGetRC(apReqs[i], &cbTransfered);
716	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(apReqs[i]);
717
718	pEndpoint = pTask->pEndpoint;
719
720	AssertMsg( RT_SUCCESS(rcReq)
721	&& ( (cbTransfered == pTask->DataSeg.cbSeg)
722	\|\| (pTask->fBounceBuffer)),
723	("Task didn't completed successfully (rc=%Rrc) or was incomplete (cbTransfered=%u)\n", rc, cbTransfered));
724
725	if (pTask->fPrefetch)
726	{
727	Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE);
728	Assert(pTask->fBounceBuffer);
729
730	memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
731	pTask->DataSeg.pvSeg,
732	pTask->DataSeg.cbSeg);
733
734	/* Write it now. */
735	pTask->fPrefetch = false;
736	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg, 512);
737	RTFOFF OffStart = pTask->Off & ~(RTFOFF)(512-1);
738
739	/* Grow the file if needed. */
740	if (RT_UNLIKELY((pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
741	{
742	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
743	RTFileSetSize(pEndpoint->File, pTask->Off + pTask->DataSeg.cbSeg);
744	}
745
746	rc = RTFileAioReqPrepareWrite(apReqs[i], pEndpoint->File,
747	OffStart, pTask->pvBounceBuffer, cbToTransfer, pTask);
748	AssertRC(rc);
749	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, &apReqs[i], 1);
750	AssertRC(rc);
751	}
752	else
753	{
754	if (pTask->fBounceBuffer)
755	{
756	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
757	memcpy(pTask->DataSeg.pvSeg,
758	((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
759	pTask->DataSeg.cbSeg);
760
761	RTMemPageFree(pTask->pvBounceBuffer);
762	}
763
764	/* Put the entry on the free array */
765	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i];
766	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) %pAioMgr->cReqEntries;
767
768	pAioMgr->cRequestsActive--;
769	pEndpoint->AioMgr.cRequestsActive--;
770	pEndpoint->AioMgr.cReqsProcessed++;
771
772	/* Call completion callback */
773	pTask->pfnCompleted(pTask, pTask->pvUser);
774	pdmacFileTaskFree(pEndpoint, pTask);
775
776	/*
777	* If there is no request left on the endpoint but a flush request is set
778	* it completed now and we notify the owner.
779	* Furthermore we look for new requests and continue.
780	*/
781	if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq)
782	{
783	/* Call completion callback */
784	pTask = pEndpoint->pFlushReq;
785	pEndpoint->pFlushReq = NULL;
786
787	AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n"));
788
789	pTask->pfnCompleted(pTask, pTask->pvUser);
790	pdmacFileTaskFree(pEndpoint, pTask);
791	}
792	}
793
794	if (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
795	{
796	if (!pEndpoint->pFlushReq)
797	{
798	/* Check if there are events on the endpoint. */
799	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
800	CHECK_RC(pAioMgr, rc);
801	}
802	}
803	else if (!pEndpoint->AioMgr.cRequestsActive)
804	{
805	/* Reopen the file so that the new endpoint can reassociate with the file */
806	RTFileClose(pEndpoint->File);
807	rc = RTFileOpen(&pEndpoint->File, pEndpoint->Core.pszUri, pEndpoint->fFlags);
808	AssertRC(rc);
809
810	if (pEndpoint->AioMgr.fMoving)
811	{
812	pEndpoint->AioMgr.fMoving = false;
813	pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
814	}
815	else
816	{
817	Assert(pAioMgr->fBlockingEventPending);
818	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
819
820	/* Release the waiting thread. */
821	LogFlow(("Signalling waiter\n"));
822	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
823	AssertRC(rc);
824	}
825	}
826	}
827
828	/* Check for an external blocking event before we go to sleep again. */
829	if (pAioMgr->fBlockingEventPending)
830	{
831	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
832	CHECK_RC(pAioMgr, rc);
833	}
834
835	/* Update load statistics. */
836	uint64_t uMillisCurr = RTTimeMilliTS();
837	if (uMillisCurr > uMillisEnd)
838	{
839	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointCurr = pAioMgr->pEndpointsHead;
840
841	/* Calculate timespan. */
842	uMillisCurr -= uMillisEnd;
843
844	while (pEndpointCurr)
845	{
846	pEndpointCurr->AioMgr.cReqsPerSec = pEndpointCurr->AioMgr.cReqsProcessed / (uMillisCurr + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD);
847	pEndpointCurr->AioMgr.cReqsProcessed = 0;
848	pEndpointCurr = pEndpointCurr->AioMgr.pEndpointNext;
849	}
850
851	/* Set new update interval */
852	uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
853	}
854	}
855	}
856	}
857
858	return rc;
859	}
860
861	#undef CHECK_RC
862

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

source: vbox/trunk/src/VBox/VMM/PDMAsyncCompletionFileNormal.cpp@ 22997

Download in other formats: