PDMAsyncCompletionFileNormal.cpp@ 24355

Last change on this file since 24355 was 24355, checked in by vboxsync, 16 years ago
AsyncCompletion: Fix crash when doing load balancing
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 40.6 KB

Line
1	/* $Id: PDMAsyncCompletionFileNormal.cpp 24355 2009-11-04 20:30:10Z 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	#define RT_STRICT
24	#include <iprt/types.h>
25	#include <iprt/asm.h>
26	#include <iprt/file.h>
27	#include <iprt/mem.h>
28	#include <iprt/string.h>
29	#include <iprt/assert.h>
30	#include <VBox/log.h>
31
32	#include "PDMAsyncCompletionFileInternal.h"
33
34	/** The update period for the I/O load statistics in ms. */
35	#define PDMACEPFILEMGR_LOAD_UPDATE_PERIOD 1000
36	/** Maximum number of requests a manager will handle. */
37	#define PDMACEPFILEMGR_REQS_MAX 512 /* @todo: Find better solution wrt. the request number*/
38
39	int pdmacFileAioMgrNormalInit(PPDMACEPFILEMGR pAioMgr)
40	{
41	int rc = VINF_SUCCESS;
42
43	rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, RTFILEAIO_UNLIMITED_REQS);
44	if (rc == VERR_OUT_OF_RANGE)
45	rc = RTFileAioCtxCreate(&pAioMgr->hAioCtx, PDMACEPFILEMGR_REQS_MAX);
46
47	if (RT_SUCCESS(rc))
48	{
49	/* Initialize request handle array. */
50	pAioMgr->iFreeEntryNext = 0;
51	pAioMgr->iFreeReqNext = 0;
52	pAioMgr->cReqEntries = PDMACEPFILEMGR_REQS_MAX + 1;
53	pAioMgr->pahReqsFree = (RTFILEAIOREQ )RTMemAllocZ(pAioMgr->cReqEntries sizeof(RTFILEAIOREQ));
54
55	if (pAioMgr->pahReqsFree)
56	{
57	return VINF_SUCCESS;
58	}
59	else
60	{
61	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
62	rc = VERR_NO_MEMORY;
63	}
64	}
65
66	return rc;
67	}
68
69	void pdmacFileAioMgrNormalDestroy(PPDMACEPFILEMGR pAioMgr)
70	{
71	RTFileAioCtxDestroy(pAioMgr->hAioCtx);
72
73	while (pAioMgr->iFreeReqNext != pAioMgr->iFreeEntryNext)
74	{
75	RTFileAioReqDestroy(pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext]);
76	pAioMgr->iFreeReqNext = (pAioMgr->iFreeReqNext + 1) % pAioMgr->cReqEntries;
77	}
78
79	RTMemFree(pAioMgr->pahReqsFree);
80	}
81
82	/**
83	* Sorts the endpoint list with insertion sort.
84	*/
85	static void pdmacFileAioMgrNormalEndpointsSortByLoad(PPDMACEPFILEMGR pAioMgr)
86	{
87	PPDMASYNCCOMPLETIONENDPOINTFILE pEpPrev, pEpCurr, pEpNextToSort;
88
89	pEpPrev = pAioMgr->pEndpointsHead;
90	pEpCurr = pEpPrev->AioMgr.pEndpointNext;
91
92	while (pEpCurr)
93	{
94	/* Remember the next element to sort because the list might change. */
95	pEpNextToSort = pEpCurr->AioMgr.pEndpointNext;
96
97	/* Unlink the current element from the list. */
98	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
99	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
100
101	if (pPrev)
102	pPrev->AioMgr.pEndpointNext = pNext;
103	else
104	pAioMgr->pEndpointsHead = pNext;
105
106	if (pNext)
107	pNext->AioMgr.pEndpointPrev = pPrev;
108
109	/* Go back until we reached the place to insert the current endpoint into. */
110	while (pEpPrev && (pEpPrev->AioMgr.cReqsPerSec < pEpCurr->AioMgr.cReqsPerSec))
111	pEpPrev = pEpPrev->AioMgr.pEndpointPrev;
112
113	/* Link the endpoint into the list. */
114	if (pEpPrev)
115	pNext = pEpPrev->AioMgr.pEndpointNext;
116	else
117	pNext = pAioMgr->pEndpointsHead;
118
119	pEpCurr->AioMgr.pEndpointNext = pNext;
120	pEpCurr->AioMgr.pEndpointPrev = pEpPrev;
121
122	if (pNext)
123	pNext->AioMgr.pEndpointPrev = pEpCurr;
124
125	if (pEpPrev)
126	pEpPrev->AioMgr.pEndpointNext = pEpCurr;
127	else
128	pAioMgr->pEndpointsHead = pEpCurr;
129
130	pEpCurr = pEpNextToSort;
131	}
132
133	#ifdef DEBUG
134	/* Validate sorting alogrithm */
135	unsigned cEndpoints = 0;
136	pEpCurr = pAioMgr->pEndpointsHead;
137
138	AssertMsg(pEpCurr, ("No endpoint in the list?\n"));
139	AssertMsg(!pEpCurr->AioMgr.pEndpointPrev, ("First element in the list points to previous element\n"));
140
141	while (pEpCurr)
142	{
143	cEndpoints++;
144
145	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEpCurr->AioMgr.pEndpointNext;
146	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEpCurr->AioMgr.pEndpointPrev;
147
148	Assert(!pNext \|\| pNext->AioMgr.cReqsPerSec <= pEpCurr->AioMgr.cReqsPerSec);
149	Assert(!pPrev \|\| pPrev->AioMgr.cReqsPerSec >= pEpCurr->AioMgr.cReqsPerSec);
150
151	pEpCurr = pNext;
152	}
153
154	AssertMsg(cEndpoints == pAioMgr->cEndpoints, ("Endpoints lost during sort!\n"));
155
156	#endif
157	}
158
159	/**
160	* Removes an endpoint from the currently assigned manager.
161	*
162	* @returns TRUE if there are still requests pending on the current manager for this endpoint.
163	* FALSE otherwise.
164	* @param pEndpointRemove The endpoint to remove.
165	*/
166	static bool pdmacFileAioMgrNormalRemoveEndpoint(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove)
167	{
168	PPDMASYNCCOMPLETIONENDPOINTFILE pPrev = pEndpointRemove->AioMgr.pEndpointPrev;
169	PPDMASYNCCOMPLETIONENDPOINTFILE pNext = pEndpointRemove->AioMgr.pEndpointNext;
170	PPDMACEPFILEMGR pAioMgr = pEndpointRemove->pAioMgr;
171
172	pAioMgr->cEndpoints--;
173
174	if (pPrev)
175	pPrev->AioMgr.pEndpointNext = pNext;
176	else
177	pAioMgr->pEndpointsHead = pNext;
178
179	if (pNext)
180	pNext->AioMgr.pEndpointPrev = pPrev;
181
182	/* Make sure that there is no request pending on this manager for the endpoint. */
183	if (!pEndpointRemove->AioMgr.cRequestsActive)
184	{
185	Assert(!pEndpointRemove->pFlushReq);
186
187	/* Reopen the file so that the new endpoint can reassociate with the file */
188	RTFileClose(pEndpointRemove->File);
189	int rc = RTFileOpen(&pEndpointRemove->File, pEndpointRemove->Core.pszUri, pEndpointRemove->fFlags);
190	AssertRC(rc);
191	return false;
192	}
193
194	return true;
195	}
196
197	/**
198	* Creates a new I/O manager and spreads the I/O load of the endpoints
199	* between the given I/O manager and the new one.
200	*
201	* @returns nothing.
202	* @param pAioMgr The I/O manager with high I/O load.
203	*/
204	static void pdmacFileAioMgrNormalBalanceLoad(PPDMACEPFILEMGR pAioMgr)
205	{
206	PPDMACEPFILEMGR pAioMgrNew = NULL;
207	int rc = VINF_SUCCESS;
208
209	/* Splitting can't be done with only one open endpoint. */
210	if (pAioMgr->cEndpoints > 1)
211	{
212	rc = pdmacFileAioMgrCreate((PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass,
213	&pAioMgrNew, false);
214	if (RT_SUCCESS(rc))
215	{
216	/* We will sort the list by request count per second. */
217	pdmacFileAioMgrNormalEndpointsSortByLoad(pAioMgr);
218
219	/* Now move some endpoints to the new manager. */
220	unsigned cReqsHere = pAioMgr->pEndpointsHead->AioMgr.cReqsPerSec;
221	unsigned cReqsOther = 0;
222	PPDMASYNCCOMPLETIONENDPOINTFILE pCurr = pAioMgr->pEndpointsHead->AioMgr.pEndpointNext;
223
224	while (pCurr)
225	{
226	if (cReqsHere <= cReqsOther)
227	{
228	/*
229	* The other manager has more requests to handle now.
230	* We will keep the current endpoint.
231	*/
232	Log(("Keeping endpoint %#p{%s} with %u reqs/s\n", pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
233	cReqsHere += pCurr->AioMgr.cReqsPerSec;
234	pCurr = pCurr->AioMgr.pEndpointNext;
235	}
236	else
237	{
238	/* Move to other endpoint. */
239	Log(("Moving endpoint %#p{%s} with %u reqs/s to other manager\n", pCurr, pCurr->Core.pszUri, pCurr->AioMgr.cReqsPerSec));
240	cReqsOther += pCurr->AioMgr.cReqsPerSec;
241
242	PPDMASYNCCOMPLETIONENDPOINTFILE pMove = pCurr;
243
244	pCurr = pCurr->AioMgr.pEndpointNext;
245
246	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pMove);
247
248	if (fReqsPending)
249	{
250	pMove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
251	pMove->AioMgr.fMoving = true;
252	pMove->AioMgr.pAioMgrDst = pAioMgrNew;
253	}
254	else
255	{
256	pMove->AioMgr.fMoving = false;
257	pMove->AioMgr.pAioMgrDst = NULL;
258	pdmacFileAioMgrAddEndpoint(pAioMgrNew, pMove);
259	}
260	}
261	}
262	}
263	else
264	{
265	/* Don't process further but leave a log entry about reduced performance. */
266	LogRel(("AIOMgr: Could not create new I/O manager (rc=%Rrc). Expect reduced performance\n", rc));
267	}
268	}
269	}
270
271	/**
272	* Error handler which will create the failsafe managers and destroy the failed I/O manager.
273	*
274	* @returns VBox status code
275	* @param pAioMgr The I/O manager the error ocurred on.
276	* @param rc The error code.
277	*/
278	static int pdmacFileAioMgrNormalErrorHandler(PPDMACEPFILEMGR pAioMgr, int rc, RT_SRC_POS_DECL)
279	{
280	LogRel(("AIOMgr: I/O manager %#p encountered a critical error (rc=%Rrc) during operation. Falling back to failsafe mode. Expect reduced performance\n",
281	pAioMgr, rc));
282	LogRel(("AIOMgr: Error happened in %s:(%u){%s}\n", RT_SRC_POS_ARGS));
283	LogRel(("AIOMgr: Please contact the product vendor\n"));
284
285	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pAioMgr->pEndpointsHead->Core.pEpClass;
286
287	pAioMgr->enmState = PDMACEPFILEMGRSTATE_FAULT;
288	ASMAtomicWriteBool(&pEpClassFile->fFailsafe, true);
289
290	AssertMsgFailed(("Implement\n"));
291	return VINF_SUCCESS;
292	}
293
294	/**
295	* Put a list of tasks in the pending request list of an endpoint.
296	*/
297	DECLINLINE(void) pdmacFileAioMgrEpAddTaskList(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTaskHead)
298	{
299	/* Add the rest of the tasks to the pending list */
300	if (!pEndpoint->AioMgr.pReqsPendingHead)
301	{
302	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
303	pEndpoint->AioMgr.pReqsPendingHead = pTaskHead;
304	}
305	else
306	{
307	Assert(pEndpoint->AioMgr.pReqsPendingTail);
308	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTaskHead;
309	}
310
311	/* Update the tail. */
312	while (pTaskHead->pNext)
313	pTaskHead = pTaskHead->pNext;
314
315	pEndpoint->AioMgr.pReqsPendingTail = pTaskHead;
316	}
317
318	/**
319	* Put one task in the pending request list of an endpoint.
320	*/
321	DECLINLINE(void) pdmacFileAioMgrEpAddTask(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTask)
322	{
323	/* Add the rest of the tasks to the pending list */
324	if (!pEndpoint->AioMgr.pReqsPendingHead)
325	{
326	Assert(!pEndpoint->AioMgr.pReqsPendingTail);
327	pEndpoint->AioMgr.pReqsPendingHead = pTask;
328	}
329	else
330	{
331	Assert(pEndpoint->AioMgr.pReqsPendingTail);
332	pEndpoint->AioMgr.pReqsPendingTail->pNext = pTask;
333	}
334
335	pEndpoint->AioMgr.pReqsPendingTail = pTask;
336	}
337
338	/**
339	* Wrapper around RTFIleAioCtxSubmit() which is also doing error handling.
340	*/
341	static int pdmacFileAioMgrNormalReqsEnqueue(PPDMACEPFILEMGR pAioMgr,
342	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint,
343	PRTFILEAIOREQ pahReqs, size_t cReqs)
344	{
345	int rc;
346
347	pAioMgr->cRequestsActive += cReqs;
348	pEndpoint->AioMgr.cRequestsActive += cReqs;
349
350	LogFlow(("Enqueuing %d requests. I/O manager has a total of %d active requests now\n", cReqs, pAioMgr->cRequestsActive));
351	LogFlow(("Endpoint has a total of %d active requests now\n", pEndpoint->AioMgr.cRequestsActive));
352
353	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, pahReqs, cReqs);
354	if (RT_FAILURE(rc))
355	{
356	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
357	{
358	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
359
360	/*
361	* We run out of resources.
362	* Need to check which requests got queued
363	* and put the rest on the pending list again.
364	*/
365	if (RT_UNLIKELY(!pEpClass->fOutOfResourcesWarningPrinted))
366	{
367	pEpClass->fOutOfResourcesWarningPrinted = true;
368	LogRel(("AIOMgr: The operating system doesn't have enough resources "
369	"to handle the I/O load of the VM. Expect reduced I/O performance\n"));
370	}
371
372	for (size_t i = 0; i < cReqs; i++)
373	{
374	int rcReq = RTFileAioReqGetRC(pahReqs[i], NULL);
375
376	if (rcReq != VERR_FILE_AIO_IN_PROGRESS)
377	{
378	AssertMsg(rcReq == VERR_FILE_AIO_NOT_SUBMITTED,
379	("Request returned unexpected return code: rc=%Rrc\n", rcReq));
380
381	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(pahReqs[i]);
382
383	/* Put the entry on the free array */
384	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = pahReqs[i];
385	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries;
386
387	pdmacFileAioMgrEpAddTask(pEndpoint, pTask);
388	pAioMgr->cRequestsActive--;
389	pEndpoint->AioMgr.cRequestsActive--;
390	}
391	}
392	LogFlow(("Removed requests. I/O manager has a total of %d active requests now\n", pAioMgr->cRequestsActive));
393	LogFlow(("Endpoint has a total of %d active requests now\n", pEndpoint->AioMgr.cRequestsActive));
394	}
395	else
396	AssertMsgFailed(("Unexpected return code rc=%Rrc\n", rc));
397	}
398
399	return rc;
400	}
401
402	static int pdmacFileAioMgrNormalProcessTaskList(PPDMACTASKFILE pTaskHead,
403	PPDMACEPFILEMGR pAioMgr,
404	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
405	{
406	RTFILEAIOREQ apReqs[20];
407	unsigned cRequests = 0;
408	unsigned cMaxRequests = PDMACEPFILEMGR_REQS_MAX - pAioMgr->cRequestsActive;
409	int rc = VINF_SUCCESS;
410	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
411
412	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
413	("Trying to process request lists of a non active endpoint!\n"));
414
415	/* Go through the list and queue the requests until we get a flush request */
416	while ( pTaskHead
417	&& !pEndpoint->pFlushReq
418	&& (cMaxRequests > 0)
419	&& RT_SUCCESS(rc))
420	{
421	PPDMACTASKFILE pCurr = pTaskHead;
422
423	pTaskHead = pTaskHead->pNext;
424
425	pCurr->pNext = NULL;
426
427	AssertMsg(VALID_PTR(pCurr->pEndpoint) && (pCurr->pEndpoint == pEndpoint),
428	("Endpoints do not match\n"));
429
430	switch (pCurr->enmTransferType)
431	{
432	case PDMACTASKFILETRANSFER_FLUSH:
433	{
434	/* If there is no data transfer request this flush request finished immediately. */
435	if (!pEndpoint->AioMgr.cRequestsActive)
436	{
437	pCurr->pfnCompleted(pCurr, pCurr->pvUser);
438	pdmacFileTaskFree(pEndpoint, pCurr);
439	}
440	else
441	{
442	pEndpoint->pFlushReq = pCurr;
443	}
444	break;
445	}
446	case PDMACTASKFILETRANSFER_READ:
447	case PDMACTASKFILETRANSFER_WRITE:
448	{
449	RTFILEAIOREQ hReq = NIL_RTFILEAIOREQ;
450	void *pvBuf = pCurr->DataSeg.pvSeg;
451
452	/* Get a request handle. */
453	if (pAioMgr->iFreeReqNext != pAioMgr->iFreeEntryNext)
454	{
455	hReq = pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext];
456	pAioMgr->pahReqsFree[pAioMgr->iFreeReqNext] = NIL_RTFILEAIOREQ;
457	pAioMgr->iFreeReqNext = (pAioMgr->iFreeReqNext + 1) % pAioMgr->cReqEntries;
458	}
459	else
460	{
461	rc = RTFileAioReqCreate(&hReq);
462	AssertRC(rc);
463	}
464
465	AssertMsg(hReq != NIL_RTFILEAIOREQ, ("Out of request handles\n"));
466
467	/* Check if the alignment requirements are met.
468	* Offset, transfer size and buffer address
469	* need to be on a 512 boundary. */
470	RTFOFF offStart = pCurr->Off & ~(RTFOFF)(512-1);
471	size_t cbToTransfer = RT_ALIGN_Z(pCurr->DataSeg.cbSeg + (pCurr->Off - offStart), 512);
472	PDMACTASKFILETRANSFER enmTransferType = pCurr->enmTransferType;
473
474	AssertMsg( pCurr->enmTransferType == PDMACTASKFILETRANSFER_WRITE
475	\|\| (uint64_t)(offStart + cbToTransfer) <= pEndpoint->cbFile,
476	("Read exceeds file size offStart=%RTfoff cbToTransfer=%d cbFile=%llu\n",
477	offStart, cbToTransfer, pEndpoint->cbFile));
478
479	pCurr->fPrefetch = false;
480
481	if ( RT_UNLIKELY(cbToTransfer != pCurr->DataSeg.cbSeg)
482	\|\| RT_UNLIKELY(offStart != pCurr->Off)
483	\|\| ((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) != (RTR3UINTPTR)pvBuf))
484	{
485	LogFlow(("Using bounce buffer for task %#p cbToTransfer=%zd cbSeg=%zd offStart=%RTfoff off=%RTfoff\n",
486	pCurr, cbToTransfer, pCurr->DataSeg.cbSeg, offStart, pCurr->Off));
487
488	/* Create bounce buffer. */
489	pCurr->fBounceBuffer = true;
490
491	AssertMsg(pCurr->Off >= offStart, ("Overflow in calculation Off=%llu offStart=%llu\n",
492	pCurr->Off, offStart));
493	pCurr->uBounceBufOffset = pCurr->Off - offStart;
494
495	/** @todo: I think we need something like a RTMemAllocAligned method here.
496	* Current assumption is that the maximum alignment is 4096byte
497	* (GPT disk on Windows)
498	* so we can use RTMemPageAlloc here.
499	*/
500	pCurr->pvBounceBuffer = RTMemPageAlloc(cbToTransfer);
501	AssertPtr(pCurr->pvBounceBuffer);
502	pvBuf = pCurr->pvBounceBuffer;
503
504	if (pCurr->enmTransferType == PDMACTASKFILETRANSFER_WRITE)
505	{
506	if ( RT_UNLIKELY(cbToTransfer != pCurr->DataSeg.cbSeg)
507	\|\| RT_UNLIKELY(offStart != pCurr->Off))
508	{
509	/* We have to fill the buffer first before we can update the data. */
510	LogFlow(("Prefetching data for task %#p\n", pCurr));
511	pCurr->fPrefetch = true;
512	enmTransferType = PDMACTASKFILETRANSFER_READ;
513	}
514	else
515	memcpy(pvBuf, pCurr->DataSeg.pvSeg, pCurr->DataSeg.cbSeg);
516	}
517	}
518	else
519	pCurr->fBounceBuffer = false;
520
521	AssertMsg((pEpClassFile->uBitmaskAlignment & (RTR3UINTPTR)pvBuf) == (RTR3UINTPTR)pvBuf,
522	("AIO: Alignment restrictions not met! pvBuf=%p uBitmaskAlignment=%p\n", pvBuf, pEpClassFile->uBitmaskAlignment));
523
524	if (enmTransferType == PDMACTASKFILETRANSFER_WRITE)
525	{
526	/* Grow the file if needed. */
527	if (RT_UNLIKELY((uint64_t)(pCurr->Off + pCurr->DataSeg.cbSeg) > pEndpoint->cbFile))
528	{
529	ASMAtomicWriteU64(&pEndpoint->cbFile, pCurr->Off + pCurr->DataSeg.cbSeg);
530	RTFileSetSize(pEndpoint->File, pCurr->Off + pCurr->DataSeg.cbSeg);
531	}
532
533	rc = RTFileAioReqPrepareWrite(hReq, pEndpoint->File,
534	offStart, pvBuf, cbToTransfer, pCurr);
535	}
536	else
537	rc = RTFileAioReqPrepareRead(hReq, pEndpoint->File,
538	offStart, pvBuf, cbToTransfer, pCurr);
539	AssertRC(rc);
540
541	apReqs[cRequests] = hReq;
542	pEndpoint->AioMgr.cReqsProcessed++;
543	cMaxRequests--;
544	cRequests++;
545	if (cRequests == RT_ELEMENTS(apReqs))
546	{
547	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
548	cRequests = 0;
549	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
550	("Unexpected return code\n"));
551	}
552	break;
553	}
554	default:
555	AssertMsgFailed(("Invalid transfer type %d\n", pCurr->enmTransferType));
556	}
557	}
558
559	if (cRequests)
560	{
561	rc = pdmacFileAioMgrNormalReqsEnqueue(pAioMgr, pEndpoint, apReqs, cRequests);
562	AssertMsg(RT_SUCCESS(rc) \|\| (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES),
563	("Unexpected return code rc=%Rrc\n", rc));
564	}
565
566	if (pTaskHead)
567	{
568	/* Add the rest of the tasks to the pending list */
569	pdmacFileAioMgrEpAddTaskList(pEndpoint, pTaskHead);
570
571	if (RT_UNLIKELY(!cMaxRequests && !pEndpoint->pFlushReq))
572	{
573	/*
574	* The I/O manager has no room left for more requests
575	* but there are still requests to process.
576	* Create a new I/O manager and let it handle some endpoints.
577	*/
578	pdmacFileAioMgrNormalBalanceLoad(pAioMgr);
579	}
580	}
581
582	/* Insufficient resources are not fatal. */
583	if (rc == VERR_FILE_AIO_INSUFFICIENT_RESSOURCES)
584	rc = VINF_SUCCESS;
585
586	return rc;
587	}
588
589	/**
590	* Adds all pending requests for the given endpoint
591	* until a flush request is encountered or there is no
592	* request anymore.
593	*
594	* @returns VBox status code.
595	* @param pAioMgr The async I/O manager for the endpoint
596	* @param pEndpoint The endpoint to get the requests from.
597	*/
598	static int pdmacFileAioMgrNormalQueueReqs(PPDMACEPFILEMGR pAioMgr,
599	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
600	{
601	int rc = VINF_SUCCESS;
602	PPDMACTASKFILE pTasksHead = NULL;
603
604	AssertMsg(pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE,
605	("Trying to process request lists of a non active endpoint!\n"));
606
607	Assert(!pEndpoint->pFlushReq);
608
609	/* Check the pending list first */
610	if (pEndpoint->AioMgr.pReqsPendingHead)
611	{
612	LogFlow(("Queuing pending requests first\n"));
613
614	pTasksHead = pEndpoint->AioMgr.pReqsPendingHead;
615	/*
616	* Clear the list as the processing routine will insert them into the list
617	* again if it gets a flush request.
618	*/
619	pEndpoint->AioMgr.pReqsPendingHead = NULL;
620	pEndpoint->AioMgr.pReqsPendingTail = NULL;
621	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
622	AssertRC(rc);
623	}
624
625	if (!pEndpoint->pFlushReq && !pEndpoint->AioMgr.pReqsPendingHead)
626	{
627	/* Now the request queue. */
628	pTasksHead = pdmacFileEpGetNewTasks(pEndpoint);
629	if (pTasksHead)
630	{
631	rc = pdmacFileAioMgrNormalProcessTaskList(pTasksHead, pAioMgr, pEndpoint);
632	AssertRC(rc);
633	}
634	}
635
636	return rc;
637	}
638
639	static int pdmacFileAioMgrNormalProcessBlockingEvent(PPDMACEPFILEMGR pAioMgr)
640	{
641	int rc = VINF_SUCCESS;
642	bool fNotifyWaiter = false;
643
644	LogFlowFunc((": Enter\n"));
645
646	Assert(pAioMgr->fBlockingEventPending);
647
648	switch (pAioMgr->enmBlockingEvent)
649	{
650	case PDMACEPFILEAIOMGRBLOCKINGEVENT_ADD_ENDPOINT:
651	{
652	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointNew = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint);
653	AssertMsg(VALID_PTR(pEndpointNew), ("Adding endpoint event without a endpoint to add\n"));
654
655	pEndpointNew->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE;
656
657	pEndpointNew->AioMgr.pEndpointNext = pAioMgr->pEndpointsHead;
658	pEndpointNew->AioMgr.pEndpointPrev = NULL;
659	if (pAioMgr->pEndpointsHead)
660	pAioMgr->pEndpointsHead->AioMgr.pEndpointPrev = pEndpointNew;
661	pAioMgr->pEndpointsHead = pEndpointNew;
662
663	/* Assign the completion point to this file. */
664	rc = RTFileAioCtxAssociateWithFile(pAioMgr->hAioCtx, pEndpointNew->File);
665	fNotifyWaiter = true;
666	pAioMgr->cEndpoints++;
667	break;
668	}
669	case PDMACEPFILEAIOMGRBLOCKINGEVENT_REMOVE_ENDPOINT:
670	{
671	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointRemove = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint);
672	AssertMsg(VALID_PTR(pEndpointRemove), ("Removing endpoint event without a endpoint to remove\n"));
673
674	pEndpointRemove->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_REMOVING;
675	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointRemove);
676	break;
677	}
678	case PDMACEPFILEAIOMGRBLOCKINGEVENT_CLOSE_ENDPOINT:
679	{
680	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointClose = (PPDMASYNCCOMPLETIONENDPOINTFILE)ASMAtomicReadPtr((void * volatile *)&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint);
681	AssertMsg(VALID_PTR(pEndpointClose), ("Close endpoint event without a endpoint to close\n"));
682
683	LogFlowFunc((": Closing endpoint %#p{%s}\n", pEndpointClose, pEndpointClose->Core.pszUri));
684
685	/* Make sure all tasks finished. Process the queues a last time first. */
686	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpointClose);
687	AssertRC(rc);
688
689	pEndpointClose->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_CLOSING;
690	fNotifyWaiter = !pdmacFileAioMgrNormalRemoveEndpoint(pEndpointClose);
691	break;
692	}
693	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SHUTDOWN:
694	{
695	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SHUTDOWN;
696	if (!pAioMgr->cRequestsActive)
697	fNotifyWaiter = true;
698	break;
699	}
700	case PDMACEPFILEAIOMGRBLOCKINGEVENT_SUSPEND:
701	{
702	pAioMgr->enmState = PDMACEPFILEMGRSTATE_SUSPENDING;
703	break;
704	}
705	case PDMACEPFILEAIOMGRBLOCKINGEVENT_RESUME:
706	{
707	pAioMgr->enmState = PDMACEPFILEMGRSTATE_RUNNING;
708	fNotifyWaiter = true;
709	break;
710	}
711	default:
712	AssertReleaseMsgFailed(("Invalid event type %d\n", pAioMgr->enmBlockingEvent));
713	}
714
715	if (fNotifyWaiter)
716	{
717	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
718	pAioMgr->enmBlockingEvent = PDMACEPFILEAIOMGRBLOCKINGEVENT_INVALID;
719
720	/* Release the waiting thread. */
721	LogFlow(("Signalling waiter\n"));
722	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
723	AssertRC(rc);
724	}
725
726	LogFlowFunc((": Leave\n"));
727	return rc;
728	}
729
730	/**
731	* Checks all endpoints for pending events or new requests.
732	*
733	* @returns VBox status code.
734	* @param pAioMgr The I/O manager handle.
735	*/
736	static int pdmacFileAioMgrNormalCheckEndpoints(PPDMACEPFILEMGR pAioMgr)
737	{
738	/* Check the assigned endpoints for new tasks if there isn't a flush request active at the moment. */
739	int rc = VINF_SUCCESS;
740	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint = pAioMgr->pEndpointsHead;
741
742	while (pEndpoint)
743	{
744	if (!pEndpoint->pFlushReq && (pEndpoint->enmState == PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE))
745	{
746	rc = pdmacFileAioMgrNormalQueueReqs(pAioMgr, pEndpoint);
747	if (RT_FAILURE(rc))
748	return rc;
749	}
750	else if (!pEndpoint->AioMgr.cRequestsActive)
751	{
752	/* Reopen the file so that the new endpoint can reassociate with the file */
753	RTFileClose(pEndpoint->File);
754	rc = RTFileOpen(&pEndpoint->File, pEndpoint->Core.pszUri, pEndpoint->fFlags);
755	AssertRC(rc);
756
757	if (pEndpoint->AioMgr.fMoving)
758	{
759	pEndpoint->AioMgr.fMoving = false;
760	pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
761	}
762	else
763	{
764	Assert(pAioMgr->fBlockingEventPending);
765	ASMAtomicWriteBool(&pAioMgr->fBlockingEventPending, false);
766
767	/* Release the waiting thread. */
768	LogFlow(("Signalling waiter\n"));
769	rc = RTSemEventSignal(pAioMgr->EventSemBlock);
770	AssertRC(rc);
771	}
772	}
773
774	pEndpoint = pEndpoint->AioMgr.pEndpointNext;
775	}
776
777	return rc;
778	}
779
780	/** Helper macro for checking for error codes. */
781	#define CHECK_RC(pAioMgr, rc) \
782	if (RT_FAILURE(rc)) \
783	{\
784	int rc2 = pdmacFileAioMgrNormalErrorHandler(pAioMgr, rc, RT_SRC_POS);\
785	return rc2;\
786	}
787
788	/**
789	* The normal I/O manager using the RTFileAio* API
790	*
791	* @returns VBox status code.
792	* @param ThreadSelf Handle of the thread.
793	* @param pvUser Opaque user data.
794	*/
795	int pdmacFileAioMgrNormal(RTTHREAD ThreadSelf, void *pvUser)
796	{
797	int rc = VINF_SUCCESS;
798	PPDMACEPFILEMGR pAioMgr = (PPDMACEPFILEMGR)pvUser;
799	uint64_t uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
800
801	while ( (pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING)
802	\|\| (pAioMgr->enmState == PDMACEPFILEMGRSTATE_SUSPENDING))
803	{
804	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, true);
805	if (!ASMAtomicReadBool(&pAioMgr->fWokenUp))
806	rc = RTSemEventWait(pAioMgr->EventSem, RT_INDEFINITE_WAIT);
807	ASMAtomicWriteBool(&pAioMgr->fWaitingEventSem, false);
808	AssertRC(rc);
809
810	LogFlow(("Got woken up\n"));
811	ASMAtomicWriteBool(&pAioMgr->fWokenUp, false);
812
813	/* Check for an external blocking event first. */
814	if (pAioMgr->fBlockingEventPending)
815	{
816	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
817	CHECK_RC(pAioMgr, rc);
818	}
819
820	if (RT_LIKELY(pAioMgr->enmState == PDMACEPFILEMGRSTATE_RUNNING))
821	{
822	/* We got woken up because an endpoint issued new requests. Queue them. */
823	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
824	CHECK_RC(pAioMgr, rc);
825
826	while (pAioMgr->cRequestsActive)
827	{
828	RTFILEAIOREQ apReqs[20];
829	uint32_t cReqsCompleted = 0;
830	size_t cReqsWait;
831
832	if (pAioMgr->cRequestsActive > RT_ELEMENTS(apReqs))
833	cReqsWait = RT_ELEMENTS(apReqs);
834	else
835	cReqsWait = pAioMgr->cRequestsActive;
836
837	LogFlow(("Waiting for %d of %d tasks to complete\n", pAioMgr->cRequestsActive, cReqsWait));
838
839	rc = RTFileAioCtxWait(pAioMgr->hAioCtx,
840	cReqsWait,
841	RT_INDEFINITE_WAIT, apReqs,
842	RT_ELEMENTS(apReqs), &cReqsCompleted);
843	if (RT_FAILURE(rc) && (rc != VERR_INTERRUPTED))
844	CHECK_RC(pAioMgr, rc);
845
846	LogFlow(("%d tasks completed\n", cReqsCompleted));
847
848	for (uint32_t i = 0; i < cReqsCompleted; i++)
849	{
850	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint;
851	size_t cbTransfered = 0;
852	int rcReq = RTFileAioReqGetRC(apReqs[i], &cbTransfered);
853	PPDMACTASKFILE pTask = (PPDMACTASKFILE)RTFileAioReqGetUser(apReqs[i]);
854
855	pEndpoint = pTask->pEndpoint;
856
857	/*
858	* It is possible that the request failed on Linux with kernels < 2.6.23
859	* if the passed buffer was allocated with remap_pfn_range or if the file
860	* is on an NFS endpoint which does not support async and direct I/O at the same time.
861	* The endpoint will be migrated to a failsafe manager in case a request fails.
862	*/
863	if (RT_FAILURE(rcReq))
864	{
865	/* Free bounce buffers and the IPRT request. */
866	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i];
867	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries;
868
869	pAioMgr->cRequestsActive--;
870	pEndpoint->AioMgr.cRequestsActive--;
871	pEndpoint->AioMgr.cReqsProcessed++;
872
873	if (pTask->fBounceBuffer)
874	RTMemFree(pTask->pvBounceBuffer);
875
876	/* Queue the request on the pending list. */
877	pTask->pNext = pEndpoint->AioMgr.pReqsPendingHead;
878	pEndpoint->AioMgr.pReqsPendingHead = pTask;
879
880	/* Create a new failsafe manager if neccessary. */
881	if (!pEndpoint->AioMgr.fMoving)
882	{
883	PPDMACEPFILEMGR pAioMgrFailsafe;
884
885	LogRel(("%s: Request %#p failed with rc=%Rrc, migrating endpoint %s to failsafe manager.\n",
886	RTThreadGetName(pAioMgr->Thread), pTask, rcReq, pEndpoint->Core.pszUri));
887
888	pEndpoint->AioMgr.fMoving = true;
889
890	rc = pdmacFileAioMgrCreate((PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass,
891	&pAioMgrFailsafe, true);
892	AssertRC(rc);
893
894	pEndpoint->AioMgr.pAioMgrDst = pAioMgrFailsafe;
895
896	/* Update the flags to open the file with. Disable async I/O and enable the host cache. */
897	pEndpoint->fFlags &= ~(RTFILE_O_ASYNC_IO \| RTFILE_O_NO_CACHE);
898	}
899
900	/* If this was the last request for the endpoint migrate it to the new manager. */
901	if (!pEndpoint->AioMgr.cRequestsActive)
902	{
903	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pEndpoint);
904	Assert(!fReqsPending);
905
906	rc = pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
907	AssertRC(rc);
908	}
909	}
910	else
911	{
912	AssertMsg(( (cbTransfered == pTask->DataSeg.cbSeg)
913	\|\| (pTask->fBounceBuffer && (cbTransfered >= pTask->DataSeg.cbSeg))),
914	("Task didn't completed successfully (rc=%Rrc) or was incomplete (cbTransfered=%u)\n", rcReq, cbTransfered));
915
916	if (pTask->fPrefetch)
917	{
918	Assert(pTask->enmTransferType == PDMACTASKFILETRANSFER_WRITE);
919	Assert(pTask->fBounceBuffer);
920
921	memcpy(((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
922	pTask->DataSeg.pvSeg,
923	pTask->DataSeg.cbSeg);
924
925	/* Write it now. */
926	pTask->fPrefetch = false;
927	size_t cbToTransfer = RT_ALIGN_Z(pTask->DataSeg.cbSeg, 512);
928	RTFOFF offStart = pTask->Off & ~(RTFOFF)(512-1);
929
930	/* Grow the file if needed. */
931	if (RT_UNLIKELY((uint64_t)(pTask->Off + pTask->DataSeg.cbSeg) > pEndpoint->cbFile))
932	{
933	ASMAtomicWriteU64(&pEndpoint->cbFile, pTask->Off + pTask->DataSeg.cbSeg);
934	RTFileSetSize(pEndpoint->File, pTask->Off + pTask->DataSeg.cbSeg);
935	}
936
937	rc = RTFileAioReqPrepareWrite(apReqs[i], pEndpoint->File,
938	offStart, pTask->pvBounceBuffer, cbToTransfer, pTask);
939	AssertRC(rc);
940	rc = RTFileAioCtxSubmit(pAioMgr->hAioCtx, &apReqs[i], 1);
941	AssertRC(rc);
942	}
943	else
944	{
945	if (pTask->fBounceBuffer)
946	{
947	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
948	memcpy(pTask->DataSeg.pvSeg,
949	((uint8_t *)pTask->pvBounceBuffer) + pTask->uBounceBufOffset,
950	pTask->DataSeg.cbSeg);
951
952	RTMemPageFree(pTask->pvBounceBuffer);
953	}
954
955	/* Put the entry on the free array */
956	pAioMgr->pahReqsFree[pAioMgr->iFreeEntryNext] = apReqs[i];
957	pAioMgr->iFreeEntryNext = (pAioMgr->iFreeEntryNext + 1) % pAioMgr->cReqEntries;
958
959	pAioMgr->cRequestsActive--;
960	pEndpoint->AioMgr.cRequestsActive--;
961	pEndpoint->AioMgr.cReqsProcessed++;
962
963	/* Call completion callback */
964	pTask->pfnCompleted(pTask, pTask->pvUser);
965	pdmacFileTaskFree(pEndpoint, pTask);
966
967	/*
968	* If there is no request left on the endpoint but a flush request is set
969	* it completed now and we notify the owner.
970	* Furthermore we look for new requests and continue.
971	*/
972	if (!pEndpoint->AioMgr.cRequestsActive && pEndpoint->pFlushReq)
973	{
974	/* Call completion callback */
975	pTask = pEndpoint->pFlushReq;
976	pEndpoint->pFlushReq = NULL;
977
978	AssertMsg(pTask->pEndpoint == pEndpoint, ("Endpoint of the flush request does not match assigned one\n"));
979
980	pTask->pfnCompleted(pTask, pTask->pvUser);
981	pdmacFileTaskFree(pEndpoint, pTask);
982	}
983	else if (RT_UNLIKELY(!pEndpoint->AioMgr.cRequestsActive && pEndpoint->AioMgr.fMoving))
984	{
985	/* If the endpoint is about to be migrated do it now. */
986	bool fReqsPending = pdmacFileAioMgrNormalRemoveEndpoint(pEndpoint);
987	Assert(!fReqsPending);
988
989	rc = pdmacFileAioMgrAddEndpoint(pEndpoint->AioMgr.pAioMgrDst, pEndpoint);
990	AssertRC(rc);
991	}
992	}
993	} /* request completed successfully */
994	} /* for every completed request */
995
996	/* Check for an external blocking event before we go to sleep again. */
997	if (pAioMgr->fBlockingEventPending)
998	{
999	rc = pdmacFileAioMgrNormalProcessBlockingEvent(pAioMgr);
1000	CHECK_RC(pAioMgr, rc);
1001	}
1002
1003	/* Update load statistics. */
1004	uint64_t uMillisCurr = RTTimeMilliTS();
1005	if (uMillisCurr > uMillisEnd)
1006	{
1007	PPDMASYNCCOMPLETIONENDPOINTFILE pEndpointCurr = pAioMgr->pEndpointsHead;
1008
1009	/* Calculate timespan. */
1010	uMillisCurr -= uMillisEnd;
1011
1012	while (pEndpointCurr)
1013	{
1014	pEndpointCurr->AioMgr.cReqsPerSec = pEndpointCurr->AioMgr.cReqsProcessed / (uMillisCurr + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD);
1015	pEndpointCurr->AioMgr.cReqsProcessed = 0;
1016	pEndpointCurr = pEndpointCurr->AioMgr.pEndpointNext;
1017	}
1018
1019	/* Set new update interval */
1020	uMillisEnd = RTTimeMilliTS() + PDMACEPFILEMGR_LOAD_UPDATE_PERIOD;
1021	}
1022
1023	/* Check endpoints for new requests. */
1024	rc = pdmacFileAioMgrNormalCheckEndpoints(pAioMgr);
1025	CHECK_RC(pAioMgr, rc);
1026	} /* while requests are active. */
1027	} /* if still running */
1028	} /* while running */
1029
1030	return rc;
1031	}
1032
1033	#undef CHECK_RC
1034

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

source: vbox/trunk/src/VBox/VMM/PDMAsyncCompletionFileNormal.cpp@ 24355

Download in other formats: