PDMAsyncCompletionFileNormal.cpp@ 23495

Last change on this file since 23495 was 23404, checked in by vboxsync, 15 years ago
AsyncCompletion: Handle out of resources case properly. Should fix assertion on Darwin
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 35.8 KB

Line
1	/* $Id: PDMAsyncCompletionFileNormal.cpp 23404 2009-09-29 10:18:37Z 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	/**
290	* Put a list of tasks in the pending request list of an endpoint.
291	*/
292	DECLINLINE(void) pdmacFileAioMgrEpAddTaskList(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTaskHead)
293	{
294	/* Add the rest of the tasks to the pending list */
295	if (!pEndpoint->AioMgr.pReqsPendingHead)
296	{
297	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
298	pEndpoint->AioMgr.pReqsPendingHead = pTaskHead;
299	}
300	else
301	{
302	Assert(pEndpoint->AioMgr.pReqsPendingTail);
303	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTaskHead;
304	}
305
306	/* Update the tail. */
307	while (pTaskHead->pNext)
308	pTaskHead = pTaskHead->pNext;
309
310	pEndpoint->AioMgr.pReqsPendingTail = pTaskHead;
311	}
312
313	/**
314	* Put one task in the pending request list of an endpoint.
315	*/
316	DECLINLINE(void) pdmacFileAioMgrEpAddTask(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTask)
317	{
318	/* Add the rest of the tasks to the pending list */
319	if (!pEndpoint->AioMgr.pReqsPendingHead)
320	{
321	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
322	pEndpoint->AioMgr.pReqsPendingHead = pTask;
323	}
324	else
325	{
326	Assert(pEndpoint->AioMgr.pReqsPendingTail);
327	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTask;
328	}
329
330	pEndpoint->AioMgr.pReqsPendingTail = pTask;
331	}
332
333	/**
334	* Wrapper around RTFIleAioCtxSubmit() which is also doing error handling.
335	*/
336	static int pdmacFileAioMgrNormalReqsEnqueue(PPDMACEPFILEMGR pAioMgr,
337	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
338	PRTFILEAIOREQ pahReqs, size_t cReqs)
339	{
340	int rc;
341
342	pAioMgr->cRequestsActive += cReqs;
343	pEndpoint->AioMgr.cRequestsActive += cReqs;
344
345	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, pahReqs, cReqs);
346	if (RT_FAILURE(rc))
347	{
348	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
349	{
350	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
351
352	/*
353	* We run out of resources.
354	* Need to check which requests got queued
355	* and put the rest on the pending list again.
356	*/
357	if (RT_UNLIKELY(!pEpClass->fOutOfResourcesWarningPrinted))
358	{
359	pEpClass->fOutOfResourcesWarningPrinted = true;
360	LogRel(("AIOMgr: The operating system doesn't have enough resources "
361	"to handle the I/O load of the VM. Expect reduced I/O performance\n"));
362	}
363
364	for (size_t i = 0; i < cReqs; i++)
365	{
366	int rcReq = RTFileAioReqGetRC(pahReqs[i], NULL);
367
368	if (rcReq != VERR_FILE_AIO_IN_PROGRESS)
369	{
370	AssertMsg(rcReq == VERR_FILE_AIO_NOT_SUBMITTED,
371	("Request returned unexpected return code: rc=%Rrc\n", rcReq));
372
373	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(pahReqs[i]);
374
375	/* Put the entry on the free array */
376	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = pahReqs[i];
377	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries;
378
379	pdmacFileAioMgrEpAddTask(pEndpoint, pTask);
380	pAioMgr->cRequestsActive--;
381	pEndpoint->AioMgr.cRequestsActive--;
382	}
383	}
384	}
385	else
386	AssertMsgFailed(("Unexpected return code rc=%Rrc\n", rc));
387	}
388
389	return rc;
390	}
391
392	static int pdmacFileAioMgrNormalProcessTaskList(PPDMACTASKFILE pTaskHead,
393	PPDMACEPFILEMGR pAioMgr,
394	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
395	{
396	RTFILEAIOREQ apReqs[20];
397	unsigned cRequests = 0;
398	unsigned cMaxRequests = PDMACEPFILEMGR_REQS_MAX - pAioMgr->cRequestsActive;
399	int rc = VINF_SUCCESS;
400	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
401
402	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
403	("Trying to process request lists of a non active endpoint!\n"));
404
405	/* Go through the list and queue the requests until we get a flush request */
406	while (pTaskHead && !pEndpoint->pFlushReq && (cMaxRequests > 0))
407	{
408	PPDMACTASKFILE pCurr = pTaskHead;
409
410	pTaskHead = pTaskHead->pNext;
411
412	pCurr->pNext = NULL;
413
414	AssertMsg(VALID_PTR(pCurr->pEndpoint) && (pCurr->pEndpoint == pEndpoint),
415	("Endpoints do not match\n"));
416
417	switch (pCurr->enmTransferType)
418	{
419	case PDMACTASKFILETRANSFER_FLUSH:
420	{
421	/* If there is no data transfer request this flush request finished immediately. */
422	if (!pEndpoint->AioMgr.cRequestsActive)
423	{
424	pCurr->pfnCompleted(pCurr, pCurr->pvUser);
425	pdmacFileTaskFree(pEndpoint, pCurr);
426	}
427	else
428	{
429	pEndpoint->pFlushReq = pCurr;
430
431	/* Do not process the task list further until the flush finished. */
432	pdmacFileAioMgrEpAddTaskList(pEndpoint, pTaskHead);
433	}
434	break;
435	}
436	case PDMACTASKFILETRANSFER_READ:
437	case PDMACTASKFILETRANSFER_WRITE:
438	{
439	RTFILEAIOREQ hReq = NIL_RTFILEAIOREQ;
440	void *pvBuf = pCurr->DataSeg.pvSeg;
441
442	/* Get a request handle. */
443	if (pAioMgr->iFreeReqNext != pAioMgr->iFreeEntryNext)
444	{
445	hReq = pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext];
446	pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext] = NIL_RTFILEAIOREQ;
447	pAioMgr->iFreeReqNext = (pAioMgr->iFreeReqNext + 1) % pAioMgr->cReqEntries;
448	}
449	else
450	{
451	rc = RTFileAioReqCreate(&hReq);
452	AssertRC(rc);
453	}
454
455	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
456
457	/* Check if the alignment requirements are met.
458	* Offset, transfer size and buffer address
459	* need to be on a 512 boundary. */
460	size_t cbToTransfer = RT_ALIGN_Z(pCurr->DataSeg.cbSeg, 512);
461	RTFOFF offStart = pCurr->Off & ~(RTFOFF)(512-1);
462	PDMACTASKFILETRANSFER enmTransferType = pCurr->enmTransferType;
463
464	AssertMsg( pCurr->enmTransferType == PDMACTASKFILETRANSFER_WRITE
465	\|\| (uint64_t)(offStart + cbToTransfer) <= pEndpoint->cbFile,
466	("Read exceeds file size offStart=%RTfoff cbToTransfer=%d cbFile=%llu\n",
467	offStart, cbToTransfer, pEndpoint->cbFile));
468
469	pCurr->fPrefetch = false;
470
471	if ( RT_UNLIKELY(cbToTransfer != pCurr->DataSeg.cbSeg)
472	\|\| RT_UNLIKELY(offStart != pCurr->Off)
473	\|\| ((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) != (RTR3UINTPTR)pvBuf))
474	{
475	/* Create bounce buffer. */
476	pCurr->fBounceBuffer = true;
477
478	AssertMsg(pCurr->Off >= offStart, ("Overflow in calculation Off=%llu offStart=%llu\n",
479	pCurr->Off, offStart));
480	pCurr->uBounceBufOffset = pCurr->Off - offStart;
481
482	/** @todo: I think we need something like a RTMemAllocAligned method here.
483	* Current assumption is that the maximum alignment is 4096byte
484	* (GPT disk on Windows)
485	* so we can use RTMemPageAlloc here.
486	*/
487	pCurr->pvBounceBuffer = RTMemPageAlloc(cbToTransfer);
488	AssertPtr(pCurr->pvBounceBuffer);
489	pvBuf = pCurr->pvBounceBuffer;
490
491	if (pCurr->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
492	{
493	if ( RT_UNLIKELY(cbToTransfer != pCurr->DataSeg.cbSeg)
494	\|\| RT_UNLIKELY(offStart != pCurr->Off))
495	{
496	/* We have to fill the buffer first before we can update the data. */
497	pCurr->fPrefetch = true;
498	enmTransferType = PDMACTASKFILETRANSFER_READ;
499	}
500	else
501	memcpy(pvBuf, pCurr->DataSeg.pvSeg, pCurr->DataSeg.cbSeg);
502	}
503	}
504	else
505	pCurr->fBounceBuffer = false;
506
507	AssertMsg((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) == (RTR3UINTPTR)pvBuf,
508	("AIO: Alignment restrictions not met! pvBuf=%p uBitmaskAlignment=%p\n", pvBuf, pEpClassFile->uBitmaskAlignment));
509
510	if (enmTransferType == PDMACTASKFILETRANSFER_WRITE)
511	{
512	/* Grow the file if needed. */
513	if (RT_UNLIKELY((uint64_t)(pCurr->Off + pCurr->DataSeg.cbSeg) > pEndpoint->cbFile))
514	{
515	ASMAtomicWriteU64(&pEndpoint->cbFile, pCurr->Off + pCurr->DataSeg.cbSeg);
516	RTFileSetSize(pEndpoint->File, pCurr->Off + pCurr->DataSeg.cbSeg);
517	}
518
519	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->File,
520	offStart, pvBuf, cbToTransfer, pCurr);
521	}
522	else
523	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->File,
524	offStart, pvBuf, cbToTransfer, pCurr);
525	AssertRC(rc);
526
527	apReqs[cRequests] = hReq;
528	pEndpoint->AioMgr.cReqsProcessed++;
529	cMaxRequests--;
530	cRequests++;
531	if (cRequests == RT_ELEMENTS(apReqs))
532	{
533	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
534	cRequests = 0;
535	if (RT_FAILURE(rc))
536	{
537	AssertMsg(rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES, ("Unexpected return code\n"));
538	break;
539	}
540	}
541	break;
542	}
543	default:
544	AssertMsgFailed(("Invalid transfer type %d\n", pCurr->enmTransferType));
545	}
546	}
547
548	if (cRequests)
549	{
550	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
551	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
552	("Unexpected return code rc=%Rrc\n", rc));
553	}
554
555	if (pTaskHead)
556	{
557	/* Add the rest of the tasks to the pending list */
558	pdmacFileAioMgrEpAddTaskList(pEndpoint, pTaskHead);
559
560
561	if (RT_UNLIKELY(!cMaxRequests && !pEndpoint->pFlushReq))
562	{
563	/*
564	* The I/O manager has no room left for more requests
565	* but there are still requests to process.
566	* Create a new I/O manager and let it handle some endpoints.
567	*/
568	pdmacFileAioMgrNormalBalanceLoad(pAioMgr);
569	}
570	}
571
572	/* Insufficient resources are not fatal. */
573	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
574	rc = VINF_SUCCESS;
575
576	return rc;
577	}
578
579	/**
580	* Adds all pending requests for the given endpoint
581	* until a flush request is encountered or there is no
582	* request anymore.
583	*
584	* @returns VBox status code.
585	* @param pAioMgr The async I/O manager for the endpoint
586	* @param pEndpoint The endpoint to get the requests from.
587	*/
588	static int pdmacFileAioMgrNormalQueueReqs(PPDMACEPFILEMGR pAioMgr,
589	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
590	{
591	int rc = VINF_SUCCESS;
592	PPDMACTASKFILE pTasksHead = NULL;
593
594	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
595	("Trying to process request lists of a non active endpoint!\n"));
596
597	Assert(!pEndpoint->pFlushReq);
598
599	/* Check the pending list first */
600	if (pEndpoint->AioMgr.pReqsPendingHead)
601	{
602	pTasksHead = pEndpoint->AioMgr.pReqsPendingHead;
603	/*
604	* Clear the list as the processing routine will insert them into the list
605	* again if it gets a flush request.
606	*/
607	pEndpoint->AioMgr.pReqsPendingHead = NULL;
608	pEndpoint->AioMgr.pReqsPendingTail = NULL;
609	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
610	AssertRC(rc);
611	}
612
613	if (!pEndpoint->pFlushReq && !pEndpoint->AioMgr.pReqsPendingHead)
614	{
615	/* Now the request queue. */
616	pTasksHead = pdmacFileEpGetNewTasks(pEndpoint);
617	if (pTasksHead)
618	{
619	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
620	AssertRC(rc);
621	}
622	}
623
624	return rc;
625	}
626
627	static int pdmacFileAioMgrNormalProcessBlockingEvent(PPDMACEPFILEMGR pAioMgr)
628	{
629	int rc = VINF_SUCCESS;
630	bool fNotifyWaiter = false;
631
632	Assert(pAioMgr->fBlockingEventPending);
633
634	switch (pAioMgr->enmBlockingEvent)
635	{
636	case PDMACEPFILEAIOMGRBLOCKINGEVENT_ADD_ENDPOINT:
637	{
638	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointNew = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint);
639	AssertMsg(VALID_PTR(pEndpointNew), ("Adding endpoint event without a endpoint to add\n"));
640
641	pEndpointNew->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE;
642
643	pEndpointNew->AioMgr.pEndpointNext = pAioMgr->pEndpointsHead;
644	pEndpointNew->AioMgr.pEndpointPrev = NULL;
645	if (pAioMgr->pEndpointsHead)
646	pAioMgr->pEndpointsHead->AioMgr.pEndpointPrev = pEndpointNew;
647	pAioMgr->pEndpointsHead = pEndpointNew;
648
649	/* Assign the completion point to this file. */
650	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pEndpointNew->File);
651	fNotifyWaiter = true;
652	pAioMgr->cEndpoints++;
653	break;
654	}
655	case PDMACEPFILEAIOMGRBLOCKINGEVENT_REMOVE_ENDPOINT:
656	{
657	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint);
658	AssertMsg(VALID_PTR(pEndpointRemove), ("Removing endpoint event without a endpoint to remove\n"));
659
660	pEndpointRemove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
661	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointRemove);
662	break;
663	}
664	case PDMACEPFILEAIOMGRBLOCKINGEVENT_CLOSE_ENDPOINT:
665	{
666	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointClose = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint);
667	AssertMsg(VALID_PTR(pEndpointClose), ("Close endpoint event without a endpoint to close\n"));
668
669	/* Make sure all tasks finished. Process the queues a last time first. */
670	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpointClose);
671	AssertRC(rc);
672
673	pEndpointClose->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING;
674	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointClose);
675	break;
676	}
677	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SHUTDOWN:
678	{
679	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SHUTDOWN;
680	if (!pAioMgr->cRequestsActive)
681	fNotifyWaiter = true;
682	break;
683	}
684	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SUSPEND:
685	{
686	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SUSPENDING;
687	break;
688	}
689	case PDMACEPFILEAIOMGRBLOCKINGEVENT_RESUME:
690	{
691	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
692	fNotifyWaiter = true;
693	break;
694	}
695	default:
696	AssertReleaseMsgFailed(("Invalid event type %d\n", pAioMgr->enmBlockingEvent));
697	}
698
699	if (fNotifyWaiter)
700	{
701	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
702	pAioMgr->enmBlockingEvent = PDMACEPFILEAIOMGRBLOCKINGEVENT_INVALID;
703
704	/* Release the waiting thread. */
705	LogFlow(("Signalling waiter\n"));
706	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
707	AssertRC(rc);
708	}
709
710	return rc;
711	}
712
713	/** Helper macro for checking for error codes. */
714	#define CHECK_RC(pAioMgr, rc) \
715	if (RT_FAILURE(rc)) \
716	{\
717	int rc2 = pdmacFileAioMgrNormalErrorHandler(pAioMgr, rc, RT_SRC_POS);\
718	return rc2;\
719	}
720
721	/**
722	* The normal I/O manager using the RTFileAio* API
723	*
724	* @returns VBox status code.
725	* @param ThreadSelf Handle of the thread.
726	* @param pvUser Opaque user data.
727	*/
728	int pdmacFileAioMgrNormal(RTTHREAD ThreadSelf, void *pvUser)
729	{
730	int rc = VINF_SUCCESS;
731	PPDMACEPFILEMGR pAioMgr = (PPDMACEPFILEMGR)pvUser;
732	uint64_t uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
733
734	while ( (pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING)
735	\|\| (pAioMgr->enmState == PDMACEPFILEMGRSTATE_SUSPENDING))
736	{
737	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, true);
738	if (!ASMAtomicReadBool(&pAioMgr->fWokenUp))
739	rc = RTSemEventWait(pAioMgr->EventSem, RT_INDEFINITE_WAIT);
740	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, false);
741	AssertRC(rc);
742
743	LogFlow(("Got woken up\n"));
744	ASMAtomicWriteBool(&pAioMgr->fWokenUp, false);
745
746	/* Check for an external blocking event first. */
747	if (pAioMgr->fBlockingEventPending)
748	{
749	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
750	CHECK_RC(pAioMgr, rc);
751	}
752
753	if (RT_LIKELY(pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING))
754	{
755	/* Check the assigned endpoints for new tasks if there isn't a flush request active at the moment. */
756	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint = pAioMgr->pEndpointsHead;
757
758	while (pEndpoint)
759	{
760	if (!pEndpoint->pFlushReq && (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE))
761	{
762	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
763	CHECK_RC(pAioMgr, rc);
764	}
765
766	pEndpoint = pEndpoint->AioMgr.pEndpointNext;
767	}
768
769	while (pAioMgr->cRequestsActive)
770	{
771	RTFILEAIOREQ apReqs[20];
772	uint32_t cReqsCompleted = 0;
773	size_t cReqsWait;
774
775	if (pAioMgr->cRequestsActive > RT_ELEMENTS(apReqs))
776	cReqsWait = RT_ELEMENTS(apReqs);
777	else
778	cReqsWait = pAioMgr->cRequestsActive;
779
780	rc = RTFileAioCtxWait(pAioMgr->hAioCtx,
781	cReqsWait,
782	RT_INDEFINITE_WAIT, apReqs,
783	RT_ELEMENTS(apReqs), &cReqsCompleted);
784	if (RT_FAILURE(rc) && (rc != VERR_INTERRUPTED))
785	CHECK_RC(pAioMgr, rc);
786
787	for (uint32_t i = 0; i < cReqsCompleted; i++)
788	{
789	size_t cbTransfered = 0;
790	int rcReq = RTFileAioReqGetRC(apReqs[i], &cbTransfered);
791	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(apReqs[i]);
792
793	pEndpoint = pTask->pEndpoint;
794
795	AssertMsg( RT_SUCCESS(rcReq)
796	&& ( (cbTransfered == pTask->DataSeg.cbSeg)
797	\|\| (pTask->fBounceBuffer)),
798	("Task didn't completed successfully (rc=%Rrc) or was incomplete (cbTransfered=%u)\n", rc, cbTransfered));
799
800	if (pTask->fPrefetch)
801	{
802	Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE);
803	Assert(pTask->fBounceBuffer);
804
805	memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
806	pTask->DataSeg.pvSeg,
807	pTask->DataSeg.cbSeg);
808
809	/* Write it now. */
810	pTask->fPrefetch = false;
811	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg, 512);
812	RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1);
813
814	/* Grow the file if needed. */
815	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
816	{
817	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
818	RTFileSetSize(pEndpoint->File, pTask->Off + pTask->DataSeg.cbSeg);
819	}
820
821	rc = RTFileAioReqPrepareWrite(apReqs[i], pEndpoint->File,
822	offStart, pTask->pvBounceBuffer, cbToTransfer, pTask);
823	AssertRC(rc);
824	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, &apReqs[i], 1);
825	AssertRC(rc);
826	}
827	else
828	{
829	if (pTask->fBounceBuffer)
830	{
831	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
832	memcpy(pTask->DataSeg.pvSeg,
833	((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
834	pTask->DataSeg.cbSeg);
835
836	RTMemPageFree(pTask->pvBounceBuffer);
837	}
838
839	/* Put the entry on the free array */
840	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i];
841	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries;
842
843	pAioMgr->cRequestsActive--;
844	pEndpoint->AioMgr.cRequestsActive--;
845	pEndpoint->AioMgr.cReqsProcessed++;
846
847	/* Call completion callback */
848	pTask->pfnCompleted(pTask, pTask->pvUser);
849	pdmacFileTaskFree(pEndpoint, pTask);
850
851	/*
852	* If there is no request left on the endpoint but a flush request is set
853	* it completed now and we notify the owner.
854	* Furthermore we look for new requests and continue.
855	*/
856	if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq)
857	{
858	/* Call completion callback */
859	pTask = pEndpoint->pFlushReq;
860	pEndpoint->pFlushReq = NULL;
861
862	AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n"));
863
864	pTask->pfnCompleted(pTask, pTask->pvUser);
865	pdmacFileTaskFree(pEndpoint, pTask);
866	}
867	}
868
869	if (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE)
870	{
871	if (!pEndpoint->pFlushReq)
872	{
873	/* Check if there are events on the endpoint. */
874	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
875	CHECK_RC(pAioMgr, rc);
876	}
877	}
878	else if (!pEndpoint->AioMgr.cRequestsActive)
879	{
880	/* Reopen the file so that the new endpoint can reassociate with the file */
881	RTFileClose(pEndpoint->File);
882	rc = RTFileOpen(&pEndpoint->File, pEndpoint->Core.pszUri, pEndpoint->fFlags);
883	AssertRC(rc);
884
885	if (pEndpoint->AioMgr.fMoving)
886	{
887	pEndpoint->AioMgr.fMoving = false;
888	pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
889	}
890	else
891	{
892	Assert(pAioMgr->fBlockingEventPending);
893	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
894
895	/* Release the waiting thread. */
896	LogFlow(("Signalling waiter\n"));
897	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
898	AssertRC(rc);
899	}
900	}
901	}
902
903	/* Check for an external blocking event before we go to sleep again. */
904	if (pAioMgr->fBlockingEventPending)
905	{
906	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
907	CHECK_RC(pAioMgr, rc);
908	}
909
910	/* Update load statistics. */
911	uint64_t uMillisCurr = RTTimeMilliTS();
912	if (uMillisCurr > uMillisEnd)
913	{
914	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointCurr = pAioMgr->pEndpointsHead;
915
916	/* Calculate timespan. */
917	uMillisCurr -= uMillisEnd;
918
919	while (pEndpointCurr)
920	{
921	pEndpointCurr->AioMgr.cReqsPerSec = pEndpointCurr->AioMgr.cReqsProcessed / (uMillisCurr + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD);
922	pEndpointCurr->AioMgr.cReqsProcessed = 0;
923	pEndpointCurr = pEndpointCurr->AioMgr.pEndpointNext;
924	}
925
926	/* Set new update interval */
927	uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
928	}
929	}
930	}
931	}
932
933	return rc;
934	}
935
936	#undef CHECK_RC
937

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

source: vbox/trunk/src/VBox/VMM/PDMAsyncCompletionFileNormal.cpp@ 23495

Download in other formats: