PDMAsyncCompletionFile.cpp@ 106952

Last change on this file since 106952 was 106061, checked in by vboxsync, 2 months ago
Copyright year updates by scm.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 47.9 KB

Line
1	/* $Id: PDMAsyncCompletionFile.cpp 106061 2024-09-16 14:03:52Z vboxsync $ */
2	/** @file
3	* PDM Async I/O - Transport data asynchronous in R3 using EMT.
4	*/
5
6	/*
7	* Copyright (C) 2006-2024 Oracle and/or its affiliates.
8	*
9	* This file is part of VirtualBox base platform packages, as
10	* available from https://www.virtualbox.org.
11	*
12	* This program is free software; you can redistribute it and/or
13	* modify it under the terms of the GNU General Public License
14	* as published by the Free Software Foundation, in version 3 of the
15	* License.
16	*
17	* This program is distributed in the hope that it will be useful, but
18	* WITHOUT ANY WARRANTY; without even the implied warranty of
19	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20	* General Public License for more details.
21	*
22	* You should have received a copy of the GNU General Public License
23	* along with this program; if not, see <https://www.gnu.org/licenses>.
24	*
25	* SPDX-License-Identifier: GPL-3.0-only
26	*/
27
28
29	/*********************************************************************************************************************************
30	* Header Files *
31	*********************************************************************************************************************************/
32	#define LOG_GROUP LOG_GROUP_PDM_ASYNC_COMPLETION
33	#include "PDMInternal.h"
34	#include <VBox/vmm/pdm.h>
35	#include <VBox/vmm/mm.h>
36	#include <VBox/vmm/vm.h>
37	#include <VBox/err.h>
38	#include <VBox/log.h>
39	#include <VBox/dbg.h>
40	#include <VBox/vmm/uvm.h>
41	#include <VBox/vmm/tm.h>
42
43	#include <iprt/asm.h>
44	#include <iprt/assert.h>
45	#include <iprt/critsect.h>
46	#include <iprt/env.h>
47	#include <iprt/file.h>
48	#include <iprt/mem.h>
49	#include <iprt/semaphore.h>
50	#include <iprt/string.h>
51	#include <iprt/thread.h>
52	#include <iprt/path.h>
53	#include <iprt/rand.h>
54
55	#include "PDMAsyncCompletionFileInternal.h"
56
57
58	/*********************************************************************************************************************************
59	* Internal Functions *
60	*********************************************************************************************************************************/
61	#ifdef VBOX_WITH_DEBUGGER
62	static FNDBGCCMD pdmacEpFileErrorInject;
63	# ifdef PDM_ASYNC_COMPLETION_FILE_WITH_DELAY
64	static FNDBGCCMD pdmacEpFileDelayInject;
65	# endif
66	#endif
67
68
69	/*********************************************************************************************************************************
70	* Global Variables *
71	*********************************************************************************************************************************/
72	#ifdef VBOX_WITH_DEBUGGER
73	static const DBGCVARDESC g_aInjectErrorArgs[] =
74	{
75	/* cTimesMin, cTimesMax, enmCategory, fFlags, pszName, pszDescription */
76	{ 1, 1, DBGCVAR_CAT_STRING, 0, "direction", "write/read." },
77	{ 1, 1, DBGCVAR_CAT_STRING, 0, "filename", "Filename." },
78	{ 1, 1, DBGCVAR_CAT_NUMBER, 0, "errcode", "VBox status code." },
79	};
80
81	# ifdef PDM_ASYNC_COMPLETION_FILE_WITH_DELAY
82	static const DBGCVARDESC g_aInjectDelayArgs[] =
83	{
84	/* cTimesMin, cTimesMax, enmCategory, fFlags, pszName, pszDescription */
85	{ 1, 1, DBGCVAR_CAT_STRING, 0, "direction", "write\|read\|flush\|any." },
86	{ 1, 1, DBGCVAR_CAT_STRING, 0, "filename", "Filename." },
87	{ 1, 1, DBGCVAR_CAT_NUMBER, 0, "delay", "Delay in milliseconds." },
88	{ 1, 1, DBGCVAR_CAT_NUMBER, 0, "jitter", "Jitter of the delay." },
89	{ 1, 1, DBGCVAR_CAT_NUMBER, 0, "reqs", "Number of requests to delay." }
90
91	};
92	# endif
93
94	/** Command descriptors. */
95	static const DBGCCMD g_aCmds[] =
96	{
97	/* pszCmd, cArgsMin, cArgsMax, paArgDesc, cArgDescs, fFlags, pfnHandler pszSyntax,.pszDescription */
98	{ "injecterror", 3, 3, &g_aInjectErrorArgs[0], 3, 0, pdmacEpFileErrorInject, "", "Inject error into I/O subsystem." }
99	# ifdef PDM_ASYNC_COMPLETION_FILE_WITH_DELAY
100	,{ "injectdelay", 3, 5, &g_aInjectDelayArgs[0], RT_ELEMENTS(g_aInjectDelayArgs), 0, pdmacEpFileDelayInject, "", "Inject a delay of a request." }
101	# endif
102	};
103	#endif
104
105
106	/**
107	* Frees a task.
108	*
109	* @param pEndpoint Pointer to the endpoint the segment was for.
110	* @param pTask The task to free.
111	*/
112	void pdmacFileTaskFree(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTask)
113	{
114	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
115
116	LogFlowFunc((": pEndpoint=%p pTask=%p\n", pEndpoint, pTask));
117
118	/* Try the per endpoint cache first. */
119	if (pEndpoint->cTasksCached < pEpClass->cTasksCacheMax)
120	{
121	/* Add it to the list. */
122	pEndpoint->pTasksFreeTail->pNext = pTask;
123	pEndpoint->pTasksFreeTail = pTask;
124	ASMAtomicIncU32(&pEndpoint->cTasksCached);
125	}
126	else
127	{
128	Log(("Freeing task %p because all caches are full\n", pTask));
129	MMR3HeapFree(pTask);
130	}
131	}
132
133	/**
134	* Allocates a task segment
135	*
136	* @returns Pointer to the new task segment or NULL
137	* @param pEndpoint Pointer to the endpoint
138	*/
139	PPDMACTASKFILE pdmacFileTaskAlloc(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
140	{
141	PPDMACTASKFILE pTask = NULL;
142
143	/* Try the small per endpoint cache first. */
144	if (pEndpoint->pTasksFreeHead == pEndpoint->pTasksFreeTail)
145	{
146	/* Try the bigger endpoint class cache. */
147	PPDMASYNCCOMPLETIONEPCLASSFILE pEndpointClass = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->Core.pEpClass;
148
149	/*
150	* Allocate completely new.
151	* If this fails we return NULL.
152	*/
153	int rc = MMR3HeapAllocZEx(pEndpointClass->Core.pVM, MM_TAG_PDM_ASYNC_COMPLETION,
154	sizeof(PDMACTASKFILE),
155	(void **)&pTask);
156	if (RT_FAILURE(rc))
157	pTask = NULL;
158
159	LogFlow(("Allocated task %p -> %Rrc\n", pTask, rc));
160	}
161	else
162	{
163	/* Grab a free task from the head. */
164	AssertMsg(pEndpoint->cTasksCached > 0, ("No tasks cached but list contains more than one element\n"));
165
166	pTask = pEndpoint->pTasksFreeHead;
167	pEndpoint->pTasksFreeHead = pTask->pNext;
168	ASMAtomicDecU32(&pEndpoint->cTasksCached);
169	pTask->pNext = NULL;
170	}
171
172	return pTask;
173	}
174
175	PPDMACTASKFILE pdmacFileEpGetNewTasks(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
176	{
177	/*
178	* Get pending tasks.
179	*/
180	PPDMACTASKFILE pTasks = ASMAtomicXchgPtrT(&pEndpoint->pTasksNewHead, NULL, PPDMACTASKFILE);
181
182	/* Reverse the list to process in FIFO order. */
183	if (pTasks)
184	{
185	PPDMACTASKFILE pTask = pTasks;
186
187	pTasks = NULL;
188
189	while (pTask)
190	{
191	PPDMACTASKFILE pCur = pTask;
192	pTask = pTask->pNext;
193	pCur->pNext = pTasks;
194	pTasks = pCur;
195	}
196	}
197
198	return pTasks;
199	}
200
201	static void pdmacFileAioMgrWakeup(PPDMACEPFILEMGR pAioMgr)
202	{
203	bool fWokenUp = ASMAtomicXchgBool(&pAioMgr->fWokenUp, true);
204	if (!fWokenUp)
205	{
206	bool fWaitingEventSem = ASMAtomicReadBool(&pAioMgr->fWaitingEventSem);
207	if (fWaitingEventSem)
208	{
209	int rc = RTSemEventSignal(pAioMgr->EventSem);
210	AssertRC(rc);
211	}
212	}
213	}
214
215	static int pdmacFileAioMgrWaitForBlockingEvent(PPDMACEPFILEMGR pAioMgr, PDMACEPFILEAIOMGRBLOCKINGEVENT enmEvent)
216	{
217	ASMAtomicWriteU32((volatile uint32_t *)&pAioMgr->enmBlockingEvent, enmEvent);
218	Assert(!pAioMgr->fBlockingEventPending);
219	ASMAtomicXchgBool(&pAioMgr->fBlockingEventPending, true);
220
221	/* Wakeup the async I/O manager */
222	pdmacFileAioMgrWakeup(pAioMgr);
223
224	/* Wait for completion. */
225	int rc = RTSemEventWait(pAioMgr->EventSemBlock, RT_INDEFINITE_WAIT);
226	AssertRC(rc);
227
228	ASMAtomicXchgBool(&pAioMgr->fBlockingEventPending, false);
229	ASMAtomicWriteU32((volatile uint32_t *)&pAioMgr->enmBlockingEvent, PDMACEPFILEAIOMGRBLOCKINGEVENT_INVALID);
230
231	return rc;
232	}
233
234	int pdmacFileAioMgrAddEndpoint(PPDMACEPFILEMGR pAioMgr, PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
235	{
236	LogFlowFunc(("pAioMgr=%#p pEndpoint=%#p{%s}\n", pAioMgr, pEndpoint, pEndpoint->Core.pszUri));
237
238	/* Update the assigned I/O manager. */
239	ASMAtomicWritePtr(&pEndpoint->pAioMgr, pAioMgr);
240
241	int rc = RTCritSectEnter(&pAioMgr->CritSectBlockingEvent);
242	AssertRCReturn(rc, rc);
243
244	ASMAtomicWritePtr(&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint, pEndpoint);
245	rc = pdmacFileAioMgrWaitForBlockingEvent(pAioMgr, PDMACEPFILEAIOMGRBLOCKINGEVENT_ADD_ENDPOINT);
246	ASMAtomicWriteNullPtr(&pAioMgr->BlockingEventData.AddEndpoint.pEndpoint);
247
248	RTCritSectLeave(&pAioMgr->CritSectBlockingEvent);
249
250	return rc;
251	}
252
253	#ifdef SOME_UNUSED_FUNCTION
254	static int pdmacFileAioMgrRemoveEndpoint(PPDMACEPFILEMGR pAioMgr, PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
255	{
256	int rc = RTCritSectEnter(&pAioMgr->CritSectBlockingEvent);
257	AssertRCReturn(rc, rc);
258
259	ASMAtomicWritePtr(&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint, pEndpoint);
260	rc = pdmacFileAioMgrWaitForBlockingEvent(pAioMgr, PDMACEPFILEAIOMGRBLOCKINGEVENT_REMOVE_ENDPOINT);
261	ASMAtomicWriteNullPtr(&pAioMgr->BlockingEventData.RemoveEndpoint.pEndpoint);
262
263	RTCritSectLeave(&pAioMgr->CritSectBlockingEvent);
264
265	return rc;
266	}
267	#endif
268
269	static int pdmacFileAioMgrCloseEndpoint(PPDMACEPFILEMGR pAioMgr, PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint)
270	{
271	int rc = RTCritSectEnter(&pAioMgr->CritSectBlockingEvent);
272	AssertRCReturn(rc, rc);
273
274	ASMAtomicWritePtr(&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint, pEndpoint);
275	rc = pdmacFileAioMgrWaitForBlockingEvent(pAioMgr, PDMACEPFILEAIOMGRBLOCKINGEVENT_CLOSE_ENDPOINT);
276	ASMAtomicWriteNullPtr(&pAioMgr->BlockingEventData.CloseEndpoint.pEndpoint);
277
278	RTCritSectLeave(&pAioMgr->CritSectBlockingEvent);
279
280	return rc;
281	}
282
283	static int pdmacFileAioMgrShutdown(PPDMACEPFILEMGR pAioMgr)
284	{
285	int rc = RTCritSectEnter(&pAioMgr->CritSectBlockingEvent);
286	AssertRCReturn(rc, rc);
287
288	rc = pdmacFileAioMgrWaitForBlockingEvent(pAioMgr, PDMACEPFILEAIOMGRBLOCKINGEVENT_SHUTDOWN);
289
290	RTCritSectLeave(&pAioMgr->CritSectBlockingEvent);
291
292	return rc;
293	}
294
295	int pdmacFileEpAddTask(PPDMASYNCCOMPLETIONENDPOINTFILE pEndpoint, PPDMACTASKFILE pTask)
296	{
297	PPDMACTASKFILE pNext;
298	do
299	{
300	pNext = pEndpoint->pTasksNewHead;
301	pTask->pNext = pNext;
302	} while (!ASMAtomicCmpXchgPtr(&pEndpoint->pTasksNewHead, pTask, pNext));
303
304	pdmacFileAioMgrWakeup(ASMAtomicReadPtrT(&pEndpoint->pAioMgr, PPDMACEPFILEMGR));
305
306	return VINF_SUCCESS;
307	}
308
309	static DECLCALLBACK(void) pdmacFileEpTaskCompleted(PPDMACTASKFILE pTask, void *pvUser, int rc)
310	{
311	PPDMASYNCCOMPLETIONTASKFILE pTaskFile = (PPDMASYNCCOMPLETIONTASKFILE)pvUser;
312
313	LogFlowFunc(("pTask=%#p pvUser=%#p rc=%Rrc\n", pTask, pvUser, rc));
314
315	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_FLUSH)
316	pdmR3AsyncCompletionCompleteTask(&pTaskFile->Core, rc, true);
317	else
318	{
319	Assert((uint32_t)pTask->DataSeg.cbSeg == pTask->DataSeg.cbSeg && (int32_t)pTask->DataSeg.cbSeg >= 0);
320	uint32_t uOld = ASMAtomicSubS32(&pTaskFile->cbTransferLeft, (int32_t)pTask->DataSeg.cbSeg);
321
322	/* The first error will be returned. */
323	if (RT_FAILURE(rc))
324	ASMAtomicCmpXchgS32(&pTaskFile->rc, rc, VINF_SUCCESS);
325	#ifdef VBOX_WITH_DEBUGGER
326	else
327	{
328	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pTaskFile->Core.pEndpoint;
329
330	/* Overwrite with injected error code. */
331	if (pTask->enmTransferType == PDMACTASKFILETRANSFER_READ)
332	rc = ASMAtomicXchgS32(&pEpFile->rcReqRead, VINF_SUCCESS);
333	else
334	rc = ASMAtomicXchgS32(&pEpFile->rcReqWrite, VINF_SUCCESS);
335
336	if (RT_FAILURE(rc))
337	ASMAtomicCmpXchgS32(&pTaskFile->rc, rc, VINF_SUCCESS);
338	}
339	#endif
340
341	if (!(uOld - pTask->DataSeg.cbSeg)
342	&& !ASMAtomicXchgBool(&pTaskFile->fCompleted, true))
343	{
344	#ifdef PDM_ASYNC_COMPLETION_FILE_WITH_DELAY
345	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pTaskFile->Core.pEndpoint;
346	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEpFile->Core.pEpClass;
347
348	/* Check if we should delay completion of the request. */
349	if ( ASMAtomicReadU32(&pEpFile->msDelay) > 0
350	&& ASMAtomicReadU32(&pEpFile->cReqsDelay) > 0)
351	{
352	uint64_t tsDelay = pEpFile->msDelay;
353
354	if (pEpFile->msJitter)
355	tsDelay = (RTRandU32() % 100) > 50 ? pEpFile->msDelay + (RTRandU32() % pEpFile->msJitter)
356	: pEpFile->msDelay - (RTRandU32() % pEpFile->msJitter);
357	ASMAtomicDecU32(&pEpFile->cReqsDelay);
358
359	/* Arm the delay. */
360	pTaskFile->tsDelayEnd = RTTimeProgramMilliTS() + tsDelay;
361
362	/* Append to the list. */
363	PPDMASYNCCOMPLETIONTASKFILE pHead = NULL;
364	do
365	{
366	pHead = ASMAtomicReadPtrT(&pEpFile->pDelayedHead, PPDMASYNCCOMPLETIONTASKFILE);
367	pTaskFile->pDelayedNext = pHead;
368	} while (!ASMAtomicCmpXchgPtr(&pEpFile->pDelayedHead, pTaskFile, pHead));
369
370	if (tsDelay < pEpClassFile->cMilliesNext)
371	{
372	ASMAtomicWriteU64(&pEpClassFile->cMilliesNext, tsDelay);
373	TMTimerSetMillies(pVM, pEpClassFile->hTimer, tsDelay);
374	}
375
376	LogRel(("AIOMgr: Delaying request %#p for %u ms\n", pTaskFile, tsDelay));
377	}
378	else
379	#endif
380	pdmR3AsyncCompletionCompleteTask(&pTaskFile->Core, pTaskFile->rc, true);
381	}
382	}
383	}
384
385	DECLINLINE(void) pdmacFileEpTaskInit(PPDMASYNCCOMPLETIONTASK pTask, size_t cbTransfer)
386	{
387	PPDMASYNCCOMPLETIONTASKFILE pTaskFile = (PPDMASYNCCOMPLETIONTASKFILE)pTask;
388
389	Assert((uint32_t)cbTransfer == cbTransfer && (int32_t)cbTransfer >= 0);
390	ASMAtomicWriteS32(&pTaskFile->cbTransferLeft, (int32_t)cbTransfer);
391	ASMAtomicWriteBool(&pTaskFile->fCompleted, false);
392	ASMAtomicWriteS32(&pTaskFile->rc, VINF_SUCCESS);
393	}
394
395	static int pdmacFileEpTaskInitiate(PPDMASYNCCOMPLETIONTASK pTask,
396	PPDMASYNCCOMPLETIONENDPOINT pEndpoint, RTFOFF off,
397	PCRTSGSEG paSegments, size_t cSegments,
398	size_t cbTransfer, PDMACTASKFILETRANSFER enmTransfer)
399	{
400	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEndpoint;
401	PPDMASYNCCOMPLETIONTASKFILE pTaskFile = (PPDMASYNCCOMPLETIONTASKFILE)pTask;
402
403	Assert( (enmTransfer == PDMACTASKFILETRANSFER_READ)
404	\|\| (enmTransfer == PDMACTASKFILETRANSFER_WRITE));
405
406	for (size_t i = 0; i < cSegments; i++)
407	{
408	PPDMACTASKFILE pIoTask = pdmacFileTaskAlloc(pEpFile);
409	AssertPtr(pIoTask);
410
411	pIoTask->pEndpoint = pEpFile;
412	pIoTask->enmTransferType = enmTransfer;
413	pIoTask->Off = off;
414	pIoTask->DataSeg.cbSeg = paSegments[i].cbSeg;
415	pIoTask->DataSeg.pvSeg = paSegments[i].pvSeg;
416	pIoTask->pvUser = pTaskFile;
417	pIoTask->pfnCompleted = pdmacFileEpTaskCompleted;
418
419	/* Send it off to the I/O manager. */
420	pdmacFileEpAddTask(pEpFile, pIoTask);
421	off += paSegments[i].cbSeg;
422	#ifdef RT_STRICT
423	cbTransfer -= paSegments[i].cbSeg;
424	#else
425	RT_NOREF(cbTransfer);
426	#endif
427	}
428
429	AssertMsg(!cbTransfer, ("Incomplete transfer %u bytes left\n", cbTransfer));
430
431	return VINF_AIO_TASK_PENDING;
432	}
433
434	/**
435	* Creates a new async I/O manager.
436	*
437	* @returns VBox status code.
438	* @param pEpClass Pointer to the endpoint class data.
439	* @param ppAioMgr Where to store the pointer to the new async I/O manager on success.
440	* @param enmMgrType Wanted manager type - can be overwritten by the global override.
441	*/
442	int pdmacFileAioMgrCreate(PPDMASYNCCOMPLETIONEPCLASSFILE pEpClass, PPPDMACEPFILEMGR ppAioMgr,
443	PDMACEPFILEMGRTYPE enmMgrType)
444	{
445	LogFlowFunc((": Entered\n"));
446
447	PPDMACEPFILEMGR pAioMgrNew;
448	int rc = MMR3HeapAllocZEx(pEpClass->Core.pVM, MM_TAG_PDM_ASYNC_COMPLETION, sizeof(PDMACEPFILEMGR), (void **)&pAioMgrNew);
449	if (RT_SUCCESS(rc))
450	{
451	if (enmMgrType < pEpClass->enmMgrTypeOverride)
452	pAioMgrNew->enmMgrType = enmMgrType;
453	else
454	pAioMgrNew->enmMgrType = pEpClass->enmMgrTypeOverride;
455
456	pAioMgrNew->msBwLimitExpired = RT_INDEFINITE_WAIT;
457
458	rc = RTSemEventCreate(&pAioMgrNew->EventSem);
459	if (RT_SUCCESS(rc))
460	{
461	rc = RTSemEventCreate(&pAioMgrNew->EventSemBlock);
462	if (RT_SUCCESS(rc))
463	{
464	rc = RTCritSectInit(&pAioMgrNew->CritSectBlockingEvent);
465	if (RT_SUCCESS(rc))
466	{
467	/* Init the rest of the manager. */
468	if (pAioMgrNew->enmMgrType != PDMACEPFILEMGRTYPE_SIMPLE)
469	rc = pdmacFileAioMgrNormalInit(pAioMgrNew);
470
471	if (RT_SUCCESS(rc))
472	{
473	pAioMgrNew->enmState = PDMACEPFILEMGRSTATE_RUNNING;
474
475	rc = RTThreadCreateF(&pAioMgrNew->Thread,
476	pAioMgrNew->enmMgrType == PDMACEPFILEMGRTYPE_SIMPLE
477	? pdmacFileAioMgrFailsafe
478	: pdmacFileAioMgrNormal,
479	pAioMgrNew,
480	0,
481	RTTHREADTYPE_IO,
482	0,
483	"AioMgr%d-%s", pEpClass->cAioMgrs,
484	pAioMgrNew->enmMgrType == PDMACEPFILEMGRTYPE_SIMPLE
485	? "F"
486	: "N");
487	if (RT_SUCCESS(rc))
488	{
489	/* Link it into the list. */
490	RTCritSectEnter(&pEpClass->CritSect);
491	pAioMgrNew->pNext = pEpClass->pAioMgrHead;
492	if (pEpClass->pAioMgrHead)
493	pEpClass->pAioMgrHead->pPrev = pAioMgrNew;
494	pEpClass->pAioMgrHead = pAioMgrNew;
495	pEpClass->cAioMgrs++;
496	RTCritSectLeave(&pEpClass->CritSect);
497
498	*ppAioMgr = pAioMgrNew;
499
500	Log(("PDMAC: Successfully created new file AIO Mgr {%s}\n", RTThreadGetName(pAioMgrNew->Thread)));
501	return VINF_SUCCESS;
502	}
503	pdmacFileAioMgrNormalDestroy(pAioMgrNew);
504	}
505	RTCritSectDelete(&pAioMgrNew->CritSectBlockingEvent);
506	}
507	RTSemEventDestroy(pAioMgrNew->EventSem);
508	}
509	RTSemEventDestroy(pAioMgrNew->EventSemBlock);
510	}
511	MMR3HeapFree(pAioMgrNew);
512	}
513
514	LogFlowFunc((": Leave rc=%Rrc\n", rc));
515
516	return rc;
517	}
518
519	/**
520	* Destroys a async I/O manager.
521	*
522	* @param pEpClassFile Pointer to globals for the file endpoint class.
523	* @param pAioMgr The async I/O manager to destroy.
524	*/
525	static void pdmacFileAioMgrDestroy(PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile, PPDMACEPFILEMGR pAioMgr)
526	{
527	int rc = pdmacFileAioMgrShutdown(pAioMgr);
528	AssertRC(rc);
529
530	/* Unlink from the list. */
531	rc = RTCritSectEnter(&pEpClassFile->CritSect);
532	AssertRC(rc);
533
534	PPDMACEPFILEMGR pPrev = pAioMgr->pPrev;
535	PPDMACEPFILEMGR pNext = pAioMgr->pNext;
536
537	if (pPrev)
538	pPrev->pNext = pNext;
539	else
540	pEpClassFile->pAioMgrHead = pNext;
541
542	if (pNext)
543	pNext->pPrev = pPrev;
544
545	pEpClassFile->cAioMgrs--;
546	rc = RTCritSectLeave(&pEpClassFile->CritSect);
547	AssertRC(rc);
548
549	/* Free the resources. */
550	RTCritSectDelete(&pAioMgr->CritSectBlockingEvent);
551	RTSemEventDestroy(pAioMgr->EventSem);
552	RTSemEventDestroy(pAioMgr->EventSemBlock);
553	if (pAioMgr->enmMgrType != PDMACEPFILEMGRTYPE_SIMPLE)
554	pdmacFileAioMgrNormalDestroy(pAioMgr);
555
556	MMR3HeapFree(pAioMgr);
557	}
558
559	static int pdmacFileMgrTypeFromName(const char *pszVal, PPDMACEPFILEMGRTYPE penmMgrType)
560	{
561	int rc = VINF_SUCCESS;
562
563	if (!RTStrCmp(pszVal, "Simple"))
564	*penmMgrType = PDMACEPFILEMGRTYPE_SIMPLE;
565	else if (!RTStrCmp(pszVal, "Async"))
566	*penmMgrType = PDMACEPFILEMGRTYPE_ASYNC;
567	else
568	rc = VERR_CFGM_CONFIG_UNKNOWN_VALUE;
569
570	return rc;
571	}
572
573	static const char *pdmacFileMgrTypeToName(PDMACEPFILEMGRTYPE enmMgrType)
574	{
575	if (enmMgrType == PDMACEPFILEMGRTYPE_SIMPLE)
576	return "Simple";
577	if (enmMgrType == PDMACEPFILEMGRTYPE_ASYNC)
578	return "Async";
579
580	return NULL;
581	}
582
583	static int pdmacFileBackendTypeFromName(const char *pszVal, PPDMACFILEEPBACKEND penmBackendType)
584	{
585	int rc = VINF_SUCCESS;
586
587	if (!RTStrCmp(pszVal, "Buffered"))
588	*penmBackendType = PDMACFILEEPBACKEND_BUFFERED;
589	else if (!RTStrCmp(pszVal, "NonBuffered"))
590	*penmBackendType = PDMACFILEEPBACKEND_NON_BUFFERED;
591	else
592	rc = VERR_CFGM_CONFIG_UNKNOWN_VALUE;
593
594	return rc;
595	}
596
597	static const char *pdmacFileBackendTypeToName(PDMACFILEEPBACKEND enmBackendType)
598	{
599	if (enmBackendType == PDMACFILEEPBACKEND_BUFFERED)
600	return "Buffered";
601	if (enmBackendType == PDMACFILEEPBACKEND_NON_BUFFERED)
602	return "NonBuffered";
603
604	return NULL;
605	}
606
607	#ifdef VBOX_WITH_DEBUGGER
608
609	/**
610	* @callback_method_impl{FNDBGCCMD, The '.injecterror' command.}
611	*/
612	static DECLCALLBACK(int) pdmacEpFileErrorInject(PCDBGCCMD pCmd, PDBGCCMDHLP pCmdHlp, PUVM pUVM, PCDBGCVAR pArgs, unsigned cArgs)
613	{
614	/*
615	* Validate input.
616	*/
617	DBGC_CMDHLP_REQ_UVM_RET(pCmdHlp, pCmd, pUVM);
618	DBGC_CMDHLP_ASSERT_PARSER_RET(pCmdHlp, pCmd, -1, cArgs == 3);
619	DBGC_CMDHLP_ASSERT_PARSER_RET(pCmdHlp, pCmd, 0, pArgs[0].enmType == DBGCVAR_TYPE_STRING);
620	DBGC_CMDHLP_ASSERT_PARSER_RET(pCmdHlp, pCmd, 1, pArgs[1].enmType == DBGCVAR_TYPE_STRING);
621	DBGC_CMDHLP_ASSERT_PARSER_RET(pCmdHlp, pCmd, 2, pArgs[2].enmType == DBGCVAR_TYPE_NUMBER);
622
623	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile;
624	pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pUVM->pdm.s.apAsyncCompletionEndpointClass[PDMASYNCCOMPLETIONEPCLASSTYPE_FILE];
625
626	/* Syntax is "read\|write <filename> <status code>" */
627	bool fWrite;
628	if (!RTStrCmp(pArgs[0].u.pszString, "read"))
629	fWrite = false;
630	else if (!RTStrCmp(pArgs[0].u.pszString, "write"))
631	fWrite = true;
632	else
633	return DBGCCmdHlpFail(pCmdHlp, pCmd, "invalid transfer direction '%s'", pArgs[0].u.pszString);
634
635	int32_t rcToInject = (int32_t)pArgs[2].u.u64Number;
636	if ((uint64_t)rcToInject != pArgs[2].u.u64Number)
637	return DBGCCmdHlpFail(pCmdHlp, pCmd, "The status code '%lld' is out of range", pArgs[0].u.u64Number);
638
639	/*
640	* Search for the matching endpoint.
641	*/
642	RTCritSectEnter(&pEpClassFile->Core.CritSect);
643
644	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEpClassFile->Core.pEndpointsHead;
645	while (pEpFile)
646	{
647	if (!RTStrCmp(pArgs[1].u.pszString, RTPathFilename(pEpFile->Core.pszUri)))
648	break;
649	pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEpFile->Core.pNext;
650	}
651
652	if (pEpFile)
653	{
654	/*
655	* Do the job.
656	*/
657	if (fWrite)
658	ASMAtomicXchgS32(&pEpFile->rcReqWrite, rcToInject);
659	else
660	ASMAtomicXchgS32(&pEpFile->rcReqRead, rcToInject);
661
662	DBGCCmdHlpPrintf(pCmdHlp, "Injected %Rrc into '%s' for %s\n",
663	(int)rcToInject, pArgs[1].u.pszString, pArgs[0].u.pszString);
664	}
665
666	RTCritSectLeave(&pEpClassFile->Core.CritSect);
667
668	if (!pEpFile)
669	return DBGCCmdHlpFail(pCmdHlp, pCmd, "No file with name '%s' found", pArgs[1].u.pszString);
670	return VINF_SUCCESS;
671	}
672
673	# ifdef PDM_ASYNC_COMPLETION_FILE_WITH_DELAY
674	/**
675	* @callback_method_impl{FNDBGCCMD, The '.injectdelay' command.}
676	*/
677	static DECLCALLBACK(int) pdmacEpFileDelayInject(PCDBGCCMD pCmd, PDBGCCMDHLP pCmdHlp, PUVM pUVM, PCDBGCVAR pArgs, unsigned cArgs)
678	{
679	/*
680	* Validate input.
681	*/
682	DBGC_CMDHLP_REQ_UVM_RET(pCmdHlp, pCmd, pUVM);
683	DBGC_CMDHLP_ASSERT_PARSER_RET(pCmdHlp, pCmd, -1, cArgs >= 3);
684	DBGC_CMDHLP_ASSERT_PARSER_RET(pCmdHlp, pCmd, 0, pArgs[0].enmType == DBGCVAR_TYPE_STRING);
685	DBGC_CMDHLP_ASSERT_PARSER_RET(pCmdHlp, pCmd, 1, pArgs[1].enmType == DBGCVAR_TYPE_STRING);
686	DBGC_CMDHLP_ASSERT_PARSER_RET(pCmdHlp, pCmd, 2, pArgs[2].enmType == DBGCVAR_TYPE_NUMBER);
687
688	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile;
689	pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pUVM->pdm.s.apAsyncCompletionEndpointClass[PDMASYNCCOMPLETIONEPCLASSTYPE_FILE];
690
691	/* Syntax is "read\|write\|flush\|any <filename> <delay> [reqs]" */
692	PDMACFILEREQTYPEDELAY enmDelayType = PDMACFILEREQTYPEDELAY_ANY;
693	if (!RTStrCmp(pArgs[0].u.pszString, "read"))
694	enmDelayType = PDMACFILEREQTYPEDELAY_READ;
695	else if (!RTStrCmp(pArgs[0].u.pszString, "write"))
696	enmDelayType = PDMACFILEREQTYPEDELAY_WRITE;
697	else if (!RTStrCmp(pArgs[0].u.pszString, "flush"))
698	enmDelayType = PDMACFILEREQTYPEDELAY_FLUSH;
699	else if (!RTStrCmp(pArgs[0].u.pszString, "any"))
700	enmDelayType = PDMACFILEREQTYPEDELAY_ANY;
701	else
702	return DBGCCmdHlpFail(pCmdHlp, pCmd, "invalid transfer direction '%s'", pArgs[0].u.pszString);
703
704	uint32_t msDelay = (uint32_t)pArgs[2].u.u64Number;
705	if ((uint64_t)msDelay != pArgs[2].u.u64Number)
706	return DBGCCmdHlpFail(pCmdHlp, pCmd, "The delay '%lld' is out of range", pArgs[0].u.u64Number);
707
708	uint32_t cReqsDelay = 1;
709	uint32_t msJitter = 0;
710	if (cArgs >= 4)
711	msJitter = (uint32_t)pArgs[3].u.u64Number;
712	if (cArgs == 5)
713	cReqsDelay = (uint32_t)pArgs[4].u.u64Number;
714
715	/*
716	* Search for the matching endpoint.
717	*/
718	RTCritSectEnter(&pEpClassFile->Core.CritSect);
719
720	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEpClassFile->Core.pEndpointsHead;
721	while (pEpFile)
722	{
723	if (!RTStrCmp(pArgs[1].u.pszString, RTPathFilename(pEpFile->Core.pszUri)))
724	break;
725	pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEpFile->Core.pNext;
726	}
727
728	if (pEpFile)
729	{
730	ASMAtomicWriteSize(&pEpFile->enmTypeDelay, enmDelayType);
731	ASMAtomicWriteU32(&pEpFile->msDelay, msDelay);
732	ASMAtomicWriteU32(&pEpFile->msJitter, msJitter);
733	ASMAtomicWriteU32(&pEpFile->cReqsDelay, cReqsDelay);
734
735	DBGCCmdHlpPrintf(pCmdHlp, "Injected delay for the next %u requests of %u ms into '%s' for %s\n",
736	cReqsDelay, msDelay, pArgs[1].u.pszString, pArgs[0].u.pszString);
737	}
738
739	RTCritSectLeave(&pEpClassFile->Core.CritSect);
740
741	if (!pEpFile)
742	return DBGCCmdHlpFail(pCmdHlp, pCmd, "No file with name '%s' found", pArgs[1].u.pszString);
743	return VINF_SUCCESS;
744	}
745
746	/**
747	* @callback_method_impl{FNTMTIMERINT, }
748	*/
749	static DECLCALLBACK(void) pdmacR3TimerCallback(PVM pVM, TMTIMERHANDLE hTimer, void *pvUser)
750	{
751	Assert(hTimer == pEpClassFile->hTimer);
752	uint64_t tsCur = RTTimeProgramMilliTS();
753	uint64_t cMilliesNext = UINT64_MAX;
754	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pvUser;
755
756	ASMAtomicWriteU64(&pEpClassFile->cMilliesNext, UINT64_MAX);
757
758	/* Go through all endpoints and check for expired requests. */
759	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEpClassFile->Core.pEndpointsHead;
760
761	while (pEpFile)
762	{
763	/* Check for an expired delay. */
764	if (pEpFile->pDelayedHead != NULL)
765	{
766	PPDMASYNCCOMPLETIONTASKFILE pTaskFile = ASMAtomicXchgPtrT(&pEpFile->pDelayedHead, NULL, PPDMASYNCCOMPLETIONTASKFILE);
767
768	while (pTaskFile)
769	{
770	PPDMASYNCCOMPLETIONTASKFILE pTmp = pTaskFile;
771	pTaskFile = pTaskFile->pDelayedNext;
772
773	if (tsCur >= pTmp->tsDelayEnd)
774	{
775	LogRel(("AIOMgr: Delayed request %#p completed\n", pTmp));
776	pdmR3AsyncCompletionCompleteTask(&pTmp->Core, pTmp->rc, true);
777	}
778	else
779	{
780	/* Prepend to the delayed list again. */
781	PPDMASYNCCOMPLETIONTASKFILE pHead = NULL;
782
783	if (pTmp->tsDelayEnd - tsCur < cMilliesNext)
784	cMilliesNext = pTmp->tsDelayEnd - tsCur;
785
786	do
787	{
788	pHead = ASMAtomicReadPtrT(&pEpFile->pDelayedHead, PPDMASYNCCOMPLETIONTASKFILE);
789	pTmp->pDelayedNext = pHead;
790	} while (!ASMAtomicCmpXchgPtr(&pEpFile->pDelayedHead, pTmp, pHead));
791	}
792	}
793	}
794
795	pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEpFile->Core.pNext;
796	}
797
798	if (cMilliesNext < pEpClassFile->cMilliesNext)
799	{
800	ASMAtomicWriteU64(&pEpClassFile->cMilliesNext, cMilliesNext);
801	TMTimerSetMillies(pVM, hTimer, cMilliesNext);
802	}
803	}
804
805	# endif /* PDM_ASYNC_COMPLETION_FILE_WITH_DELAY */
806
807	#endif /* VBOX_WITH_DEBUGGER */
808
809	static DECLCALLBACK(int) pdmacFileInitialize(PPDMASYNCCOMPLETIONEPCLASS pClassGlobals, PCFGMNODE pCfgNode)
810	{
811	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pClassGlobals;
812	RTFILEAIOLIMITS AioLimits; /** < Async I/O limitations. */
813
814	int rc = RTFileAioGetLimits(&AioLimits);
815	#ifdef DEBUG
816	if (RT_SUCCESS(rc) && RTEnvExist("VBOX_ASYNC_IO_FAILBACK"))
817	rc = VERR_ENV_VAR_NOT_FOUND;
818	#endif
819	if (RT_FAILURE(rc))
820	{
821	LogRel(("AIO: Async I/O manager not supported (rc=%Rrc). Falling back to simple manager\n", rc));
822	pEpClassFile->enmMgrTypeOverride = PDMACEPFILEMGRTYPE_SIMPLE;
823	pEpClassFile->enmEpBackendDefault = PDMACFILEEPBACKEND_BUFFERED;
824	}
825	else
826	{
827	pEpClassFile->uBitmaskAlignment = AioLimits.cbBufferAlignment ? ~((RTR3UINTPTR)AioLimits.cbBufferAlignment - 1) : RTR3UINTPTR_MAX;
828	pEpClassFile->cReqsOutstandingMax = AioLimits.cReqsOutstandingMax;
829
830	if (pCfgNode)
831	{
832	/* Query the default manager type */
833	char *pszVal = NULL;
834	rc = CFGMR3QueryStringAllocDef(pCfgNode, "IoMgr", &pszVal, "Async");
835	AssertLogRelRCReturn(rc, rc);
836
837	rc = pdmacFileMgrTypeFromName(pszVal, &pEpClassFile->enmMgrTypeOverride);
838	MMR3HeapFree(pszVal);
839	if (RT_FAILURE(rc))
840	return rc;
841
842	LogRel(("AIOMgr: Default manager type is '%s'\n", pdmacFileMgrTypeToName(pEpClassFile->enmMgrTypeOverride)));
843
844	/* Query default backend type */
845	rc = CFGMR3QueryStringAllocDef(pCfgNode, "FileBackend", &pszVal, "NonBuffered");
846	AssertLogRelRCReturn(rc, rc);
847
848	rc = pdmacFileBackendTypeFromName(pszVal, &pEpClassFile->enmEpBackendDefault);
849	MMR3HeapFree(pszVal);
850	if (RT_FAILURE(rc))
851	return rc;
852
853	LogRel(("AIOMgr: Default file backend is '%s'\n", pdmacFileBackendTypeToName(pEpClassFile->enmEpBackendDefault)));
854
855	#ifdef RT_OS_LINUX
856	if ( pEpClassFile->enmMgrTypeOverride == PDMACEPFILEMGRTYPE_ASYNC
857	&& pEpClassFile->enmEpBackendDefault == PDMACFILEEPBACKEND_BUFFERED)
858	{
859	LogRel(("AIOMgr: Linux does not support buffered async I/O, changing to non buffered\n"));
860	pEpClassFile->enmEpBackendDefault = PDMACFILEEPBACKEND_NON_BUFFERED;
861	}
862	#endif
863	}
864	else
865	{
866	/* No configuration supplied, set defaults */
867	pEpClassFile->enmEpBackendDefault = PDMACFILEEPBACKEND_NON_BUFFERED;
868	pEpClassFile->enmMgrTypeOverride = PDMACEPFILEMGRTYPE_ASYNC;
869	}
870	}
871
872	/* Init critical section. */
873	rc = RTCritSectInit(&pEpClassFile->CritSect);
874
875	#ifdef VBOX_WITH_DEBUGGER
876	/* Install the error injection handler. */
877	if (RT_SUCCESS(rc))
878	{
879	rc = DBGCRegisterCommands(&g_aCmds[0], RT_ELEMENTS(g_aCmds));
880	AssertRC(rc);
881	}
882
883	# ifdef PDM_ASYNC_COMPLETION_FILE_WITH_DELAY
884	rc = TMR3TimerCreate(pEpClassFile->Core.pVM, TMCLOCK_REAL, pdmacR3TimerCallback, pEpClassFile,
885	TMTIMER_FLAGS_NO_RING0, "AC Delay", &pEpClassFile->hTimer);
886	AssertRC(rc);
887	pEpClassFile->cMilliesNext = UINT64_MAX;
888	# endif
889	#endif
890
891	return rc;
892	}
893
894	static DECLCALLBACK(void) pdmacFileTerminate(PPDMASYNCCOMPLETIONEPCLASS pClassGlobals)
895	{
896	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pClassGlobals;
897
898	/* All endpoints should be closed at this point. */
899	AssertMsg(!pEpClassFile->Core.pEndpointsHead, ("There are still endpoints left\n"));
900
901	/* Destroy all left async I/O managers. */
902	while (pEpClassFile->pAioMgrHead)
903	pdmacFileAioMgrDestroy(pEpClassFile, pEpClassFile->pAioMgrHead);
904
905	RTCritSectDelete(&pEpClassFile->CritSect);
906	}
907
908	static DECLCALLBACK(int) pdmacFileEpInitialize(PPDMASYNCCOMPLETIONENDPOINT pEndpoint,
909	const char *pszUri, uint32_t fFlags)
910	{
911	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEndpoint;
912	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->pEpClass;
913	PDMACEPFILEMGRTYPE enmMgrType = pEpClassFile->enmMgrTypeOverride;
914	PDMACFILEEPBACKEND enmEpBackend = pEpClassFile->enmEpBackendDefault;
915
916	AssertMsgReturn((fFlags & ~(PDMACEP_FILE_FLAGS_READ_ONLY \| PDMACEP_FILE_FLAGS_DONT_LOCK \| PDMACEP_FILE_FLAGS_HOST_CACHE_ENABLED)) == 0,
917	("PDMAsyncCompletion: Invalid flag specified\n"), VERR_INVALID_PARAMETER);
918
919	unsigned fFileFlags = RTFILE_O_OPEN;
920
921	/*
922	* Revert to the simple manager and the buffered backend if
923	* the host cache should be enabled.
924	*/
925	if (fFlags & PDMACEP_FILE_FLAGS_HOST_CACHE_ENABLED)
926	{
927	enmMgrType = PDMACEPFILEMGRTYPE_SIMPLE;
928	enmEpBackend = PDMACFILEEPBACKEND_BUFFERED;
929	}
930
931	if (fFlags & PDMACEP_FILE_FLAGS_READ_ONLY)
932	fFileFlags \|= RTFILE_O_READ \| RTFILE_O_DENY_NONE;
933	else
934	{
935	fFileFlags \|= RTFILE_O_READWRITE;
936
937	/*
938	* Opened in read/write mode. Check whether the caller wants to
939	* avoid the lock. Return an error in case caching is enabled
940	* because this can lead to data corruption.
941	*/
942	if (fFlags & PDMACEP_FILE_FLAGS_DONT_LOCK)
943	fFileFlags \|= RTFILE_O_DENY_NONE;
944	else
945	fFileFlags \|= RTFILE_O_DENY_WRITE;
946	}
947
948	if (enmMgrType == PDMACEPFILEMGRTYPE_ASYNC)
949	fFileFlags \|= RTFILE_O_ASYNC_IO;
950
951	int rc;
952	if (enmEpBackend == PDMACFILEEPBACKEND_NON_BUFFERED)
953	{
954	/*
955	* We only disable the cache if the size of the file is a multiple of 512.
956	* Certain hosts like Windows, Linux and Solaris require that transfer sizes
957	* are aligned to the volume sector size.
958	* If not we just make sure that the data is written to disk with RTFILE_O_WRITE_THROUGH
959	* which will trash the host cache but ensures that the host cache will not
960	* contain dirty buffers.
961	*/
962	RTFILE hFile;
963	rc = RTFileOpen(&hFile, pszUri, RTFILE_O_READ \| RTFILE_O_OPEN \| RTFILE_O_DENY_NONE);
964	if (RT_SUCCESS(rc))
965	{
966	uint64_t cbSize;
967
968	rc = RTFileQuerySize(hFile, &cbSize);
969
970	if (RT_SUCCESS(rc) && ((cbSize % 512) == 0))
971	fFileFlags \|= RTFILE_O_NO_CACHE;
972	else
973	{
974	/* Downgrade to the buffered backend */
975	enmEpBackend = PDMACFILEEPBACKEND_BUFFERED;
976
977	#ifdef RT_OS_LINUX
978	fFileFlags &= ~RTFILE_O_ASYNC_IO;
979	enmMgrType = PDMACEPFILEMGRTYPE_SIMPLE;
980	#endif
981	}
982	RTFileClose(hFile);
983	}
984	}
985
986	/* Open with final flags. */
987	rc = RTFileOpen(&pEpFile->hFile, pszUri, fFileFlags);
988	if ( rc == VERR_INVALID_FUNCTION
989	\|\| rc == VERR_INVALID_PARAMETER)
990	{
991	LogRel(("AIOMgr: pdmacFileEpInitialize: RTFileOpen %s / %08x failed with %Rrc\n",
992	pszUri, fFileFlags, rc));
993	/*
994	* Solaris doesn't support directio on ZFS so far. :-\
995	* Trying to enable it returns VERR_INVALID_FUNCTION
996	* (ENOTTY). Remove it and hope for the best.
997	* ZFS supports write throttling in case applications
998	* write more data than can be synced to the disk
999	* without blocking the whole application.
1000	*
1001	* On Linux we have the same problem with cifs.
1002	* Have to disable async I/O here too because it requires O_DIRECT.
1003	*/
1004	fFileFlags &= ~RTFILE_O_NO_CACHE;
1005	enmEpBackend = PDMACFILEEPBACKEND_BUFFERED;
1006
1007	#ifdef RT_OS_LINUX
1008	fFileFlags &= ~RTFILE_O_ASYNC_IO;
1009	enmMgrType = PDMACEPFILEMGRTYPE_SIMPLE;
1010	#endif
1011
1012	/* Open again. */
1013	rc = RTFileOpen(&pEpFile->hFile, pszUri, fFileFlags);
1014
1015	if (RT_FAILURE(rc))
1016	{
1017	LogRel(("AIOMgr: pdmacFileEpInitialize: RTFileOpen %s / %08x failed AGAIN(!) with %Rrc\n",
1018	pszUri, fFileFlags, rc));
1019	}
1020	}
1021
1022	if (RT_SUCCESS(rc))
1023	{
1024	pEpFile->fFlags = fFileFlags;
1025
1026	rc = RTFileQuerySize(pEpFile->hFile, (uint64_t *)&pEpFile->cbFile);
1027	if (RT_SUCCESS(rc))
1028	{
1029	/* Initialize the segment cache */
1030	rc = MMR3HeapAllocZEx(pEpClassFile->Core.pVM, MM_TAG_PDM_ASYNC_COMPLETION,
1031	sizeof(PDMACTASKFILE),
1032	(void **)&pEpFile->pTasksFreeHead);
1033	if (RT_SUCCESS(rc))
1034	{
1035	PPDMACEPFILEMGR pAioMgr = NULL;
1036
1037	pEpFile->pTasksFreeTail = pEpFile->pTasksFreeHead;
1038	pEpFile->cTasksCached = 0;
1039	pEpFile->enmBackendType = enmEpBackend;
1040	/*
1041	* Disable async flushes on Solaris for now.
1042	* They cause weird hangs which needs more investigations.
1043	*/
1044	#ifndef RT_OS_SOLARIS
1045	pEpFile->fAsyncFlushSupported = true;
1046	#else
1047	pEpFile->fAsyncFlushSupported = false;
1048	#endif
1049
1050	if (enmMgrType == PDMACEPFILEMGRTYPE_SIMPLE)
1051	{
1052	/* Simple mode. Every file has its own async I/O manager. */
1053	rc = pdmacFileAioMgrCreate(pEpClassFile, &pAioMgr, PDMACEPFILEMGRTYPE_SIMPLE);
1054	}
1055	else
1056	{
1057	pAioMgr = pEpClassFile->pAioMgrHead;
1058
1059	/* Check for an idling manager of the same type */
1060	while (pAioMgr)
1061	{
1062	if (pAioMgr->enmMgrType == enmMgrType)
1063	break;
1064	pAioMgr = pAioMgr->pNext;
1065	}
1066
1067	if (!pAioMgr)
1068	rc = pdmacFileAioMgrCreate(pEpClassFile, &pAioMgr, enmMgrType);
1069	}
1070
1071	if (RT_SUCCESS(rc))
1072	{
1073	pEpFile->AioMgr.pTreeRangesLocked = (PAVLRFOFFTREE)RTMemAllocZ(sizeof(AVLRFOFFTREE));
1074	if (!pEpFile->AioMgr.pTreeRangesLocked)
1075	rc = VERR_NO_MEMORY;
1076	else
1077	{
1078	pEpFile->enmState = PDMASYNCCOMPLETIONENDPOINTFILESTATE_ACTIVE;
1079
1080	/* Assign the endpoint to the thread. */
1081	rc = pdmacFileAioMgrAddEndpoint(pAioMgr, pEpFile);
1082	if (RT_FAILURE(rc))
1083	{
1084	RTMemFree(pEpFile->AioMgr.pTreeRangesLocked);
1085	MMR3HeapFree(pEpFile->pTasksFreeHead);
1086	}
1087	}
1088	}
1089	else if (rc == VERR_FILE_AIO_INSUFFICIENT_EVENTS)
1090	{
1091	PUVM pUVM = VMR3GetUVM(pEpClassFile->Core.pVM);
1092	#if defined(RT_OS_LINUX)
1093	rc = VMR3SetError(pUVM, rc, RT_SRC_POS,
1094	N_("Failed to create I/O manager for VM due to insufficient resources on the host. "
1095	"Either increase the amount of allowed events in /proc/sys/fs/aio-max-nr or enable "
1096	"the host I/O cache"));
1097	#else
1098	rc = VMR3SetError(pUVM, rc, RT_SRC_POS,
1099	N_("Failed to create I/O manager for VM due to insufficient resources on the host. "
1100	"Enable the host I/O cache"));
1101	#endif
1102	}
1103	else
1104	{
1105	PUVM pUVM = VMR3GetUVM(pEpClassFile->Core.pVM);
1106	rc = VMR3SetError(pUVM, rc, RT_SRC_POS,
1107	N_("Failed to create I/O manager for VM due to an unknown error"));
1108	}
1109	}
1110	}
1111
1112	if (RT_FAILURE(rc))
1113	RTFileClose(pEpFile->hFile);
1114	}
1115
1116	#ifdef VBOX_WITH_STATISTICS
1117	if (RT_SUCCESS(rc))
1118	{
1119	STAMR3RegisterF(pEpClassFile->Core.pVM, &pEpFile->StatRead,
1120	STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1121	STAMUNIT_TICKS_PER_CALL, "Time taken to read from the endpoint",
1122	"/PDM/AsyncCompletion/File/%s/%d/Read", RTPathFilename(pEpFile->Core.pszUri), pEpFile->Core.iStatId);
1123
1124	STAMR3RegisterF(pEpClassFile->Core.pVM, &pEpFile->StatWrite,
1125	STAMTYPE_PROFILE_ADV, STAMVISIBILITY_ALWAYS,
1126	STAMUNIT_TICKS_PER_CALL, "Time taken to write to the endpoint",
1127	"/PDM/AsyncCompletion/File/%s/%d/Write", RTPathFilename(pEpFile->Core.pszUri), pEpFile->Core.iStatId);
1128	}
1129	#endif
1130
1131	if (RT_SUCCESS(rc))
1132	LogRel(("AIOMgr: Endpoint for file '%s' (flags %08x) created successfully\n", pszUri, pEpFile->fFlags));
1133
1134	return rc;
1135	}
1136
1137	static DECLCALLBACK(int) pdmacFileEpRangesLockedDestroy(PAVLRFOFFNODECORE pNode, void *pvUser)
1138	{
1139	NOREF(pNode); NOREF(pvUser);
1140	AssertMsgFailed(("The locked ranges tree should be empty at that point\n"));
1141	return VINF_SUCCESS;
1142	}
1143
1144	static DECLCALLBACK(int) pdmacFileEpClose(PPDMASYNCCOMPLETIONENDPOINT pEndpoint)
1145	{
1146	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEndpoint;
1147	PPDMASYNCCOMPLETIONEPCLASSFILE pEpClassFile = (PPDMASYNCCOMPLETIONEPCLASSFILE)pEndpoint->pEpClass;
1148
1149	/* Make sure that all tasks finished for this endpoint. */
1150	int rc = pdmacFileAioMgrCloseEndpoint(pEpFile->pAioMgr, pEpFile);
1151	AssertRC(rc);
1152
1153	/*
1154	* If the async I/O manager is in failsafe mode this is the only endpoint
1155	* he processes and thus can be destroyed now.
1156	*/
1157	if (pEpFile->pAioMgr->enmMgrType == PDMACEPFILEMGRTYPE_SIMPLE)
1158	pdmacFileAioMgrDestroy(pEpClassFile, pEpFile->pAioMgr);
1159
1160	/* Free cached tasks. */
1161	PPDMACTASKFILE pTask = pEpFile->pTasksFreeHead;
1162
1163	while (pTask)
1164	{
1165	PPDMACTASKFILE pTaskFree = pTask;
1166	pTask = pTask->pNext;
1167	MMR3HeapFree(pTaskFree);
1168	}
1169
1170	/* Destroy the locked ranges tree now. */
1171	RTAvlrFileOffsetDestroy(pEpFile->AioMgr.pTreeRangesLocked, pdmacFileEpRangesLockedDestroy, NULL);
1172	RTMemFree(pEpFile->AioMgr.pTreeRangesLocked);
1173	pEpFile->AioMgr.pTreeRangesLocked = NULL;
1174
1175	RTFileClose(pEpFile->hFile);
1176
1177	#ifdef VBOX_WITH_STATISTICS
1178	/* Not sure if this might be unnecessary because of similar statement in pdmR3AsyncCompletionStatisticsDeregister? */
1179	STAMR3DeregisterF(pEpClassFile->Core.pVM->pUVM, "/PDM/AsyncCompletion/File/%s/*", RTPathFilename(pEpFile->Core.pszUri));
1180	#endif
1181
1182	return VINF_SUCCESS;
1183	}
1184
1185	static DECLCALLBACK(int) pdmacFileEpRead(PPDMASYNCCOMPLETIONTASK pTask,
1186	PPDMASYNCCOMPLETIONENDPOINT pEndpoint, RTFOFF off,
1187	PCRTSGSEG paSegments, size_t cSegments,
1188	size_t cbRead)
1189	{
1190	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEndpoint;
1191
1192	LogFlowFunc(("pTask=%#p pEndpoint=%#p off=%RTfoff paSegments=%#p cSegments=%zu cbRead=%zu\n",
1193	pTask, pEndpoint, off, paSegments, cSegments, cbRead));
1194
1195	if (RT_UNLIKELY((uint64_t)off + cbRead > pEpFile->cbFile))
1196	return VERR_EOF;
1197
1198	STAM_PROFILE_ADV_START(&pEpFile->StatRead, Read);
1199	pdmacFileEpTaskInit(pTask, cbRead);
1200	int rc = pdmacFileEpTaskInitiate(pTask, pEndpoint, off, paSegments, cSegments, cbRead,
1201	PDMACTASKFILETRANSFER_READ);
1202	STAM_PROFILE_ADV_STOP(&pEpFile->StatRead, Read);
1203
1204	return rc;
1205	}
1206
1207	static DECLCALLBACK(int) pdmacFileEpWrite(PPDMASYNCCOMPLETIONTASK pTask,
1208	PPDMASYNCCOMPLETIONENDPOINT pEndpoint, RTFOFF off,
1209	PCRTSGSEG paSegments, size_t cSegments,
1210	size_t cbWrite)
1211	{
1212	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEndpoint;
1213
1214	if (RT_UNLIKELY(pEpFile->fReadonly))
1215	return VERR_NOT_SUPPORTED;
1216
1217	STAM_PROFILE_ADV_START(&pEpFile->StatWrite, Write);
1218
1219	pdmacFileEpTaskInit(pTask, cbWrite);
1220
1221	int rc = pdmacFileEpTaskInitiate(pTask, pEndpoint, off, paSegments, cSegments, cbWrite,
1222	PDMACTASKFILETRANSFER_WRITE);
1223
1224	STAM_PROFILE_ADV_STOP(&pEpFile->StatWrite, Write);
1225
1226	return rc;
1227	}
1228
1229	static DECLCALLBACK(int) pdmacFileEpFlush(PPDMASYNCCOMPLETIONTASK pTask,
1230	PPDMASYNCCOMPLETIONENDPOINT pEndpoint)
1231	{
1232	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEndpoint;
1233	PPDMASYNCCOMPLETIONTASKFILE pTaskFile = (PPDMASYNCCOMPLETIONTASKFILE)pTask;
1234
1235	if (RT_UNLIKELY(pEpFile->fReadonly))
1236	return VERR_NOT_SUPPORTED;
1237
1238	pdmacFileEpTaskInit(pTask, 0);
1239
1240	PPDMACTASKFILE pIoTask = pdmacFileTaskAlloc(pEpFile);
1241	if (RT_UNLIKELY(!pIoTask))
1242	return VERR_NO_MEMORY;
1243
1244	pIoTask->pEndpoint = pEpFile;
1245	pIoTask->enmTransferType = PDMACTASKFILETRANSFER_FLUSH;
1246	pIoTask->pvUser = pTaskFile;
1247	pIoTask->pfnCompleted = pdmacFileEpTaskCompleted;
1248	pdmacFileEpAddTask(pEpFile, pIoTask);
1249
1250	return VINF_AIO_TASK_PENDING;
1251	}
1252
1253	static DECLCALLBACK(int) pdmacFileEpGetSize(PPDMASYNCCOMPLETIONENDPOINT pEndpoint, uint64_t *pcbSize)
1254	{
1255	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEndpoint;
1256
1257	*pcbSize = ASMAtomicReadU64(&pEpFile->cbFile);
1258
1259	return VINF_SUCCESS;
1260	}
1261
1262	static DECLCALLBACK(int) pdmacFileEpSetSize(PPDMASYNCCOMPLETIONENDPOINT pEndpoint, uint64_t cbSize)
1263	{
1264	int rc;
1265	PPDMASYNCCOMPLETIONENDPOINTFILE pEpFile = (PPDMASYNCCOMPLETIONENDPOINTFILE)pEndpoint;
1266
1267	rc = RTFileSetSize(pEpFile->hFile, cbSize);
1268	if (RT_SUCCESS(rc))
1269	ASMAtomicWriteU64(&pEpFile->cbFile, cbSize);
1270
1271	return rc;
1272	}
1273
1274	const PDMASYNCCOMPLETIONEPCLASSOPS g_PDMAsyncCompletionEndpointClassFile =
1275	{
1276	/* u32Version */
1277	PDMAC_EPCLASS_OPS_VERSION,
1278	/* pcszName */
1279	"File",
1280	/* enmClassType */
1281	PDMASYNCCOMPLETIONEPCLASSTYPE_FILE,
1282	/* cbEndpointClassGlobal */
1283	sizeof(PDMASYNCCOMPLETIONEPCLASSFILE),
1284	/* cbEndpoint */
1285	sizeof(PDMASYNCCOMPLETIONENDPOINTFILE),
1286	/* cbTask */
1287	sizeof(PDMASYNCCOMPLETIONTASKFILE),
1288	/* pfnInitialize */
1289	pdmacFileInitialize,
1290	/* pfnTerminate */
1291	pdmacFileTerminate,
1292	/* pfnEpInitialize. */
1293	pdmacFileEpInitialize,
1294	/* pfnEpClose */
1295	pdmacFileEpClose,
1296	/* pfnEpRead */
1297	pdmacFileEpRead,
1298	/* pfnEpWrite */
1299	pdmacFileEpWrite,
1300	/* pfnEpFlush */
1301	pdmacFileEpFlush,
1302	/* pfnEpGetSize */
1303	pdmacFileEpGetSize,
1304	/* pfnEpSetSize */
1305	pdmacFileEpSetSize,
1306	/* u32VersionEnd */
1307	PDMAC_EPCLASS_OPS_VERSION
1308	};
1309

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

source: vbox/trunk/src/VBox/VMM/VMMR3/PDMAsyncCompletionFile.cpp@ 106952

Download in other formats: