fileaio-posix.cpp@ 23396

Last change on this file since 23396 was 23396, checked in by vboxsync, 15 years ago
Runteim/Aio-posix: Respect AIO_LISTIO_MAX for aio_suspend too. Should fix another darwin issue
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 33.0 KB

Line
1	/* $Id: fileaio-posix.cpp 23396 2009-09-28 17:25:54Z vboxsync $ */
2	/** @file
3	* IPRT - File async I/O, native implementation for POSIX compliant host platforms.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 Sun Microsystems, Inc.
8	*
9	* This file is part of VirtualBox Open Source Edition (OSE), as
10	* available from http://www.virtualbox.org. This file is free software;
11	* you can redistribute it and/or modify it under the terms of the GNU
12	* General Public License (GPL) as published by the Free Software
13	* Foundation, in version 2 as it comes in the "COPYING" file of the
14	* VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15	* hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16	*
17	* The contents of this file may alternatively be used under the terms
18	* of the Common Development and Distribution License Version 1.0
19	* (CDDL) only, as it comes in the "COPYING.CDDL" file of the
20	* VirtualBox OSE distribution, in which case the provisions of the
21	* CDDL are applicable instead of those of the GPL.
22	*
23	* You may elect to license modified versions of this file under the
24	* terms and conditions of either the GPL or the CDDL or both.
25	*
26	* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
27	* Clara, CA 95054 USA or visit http://www.sun.com if you need
28	* additional information or have any questions.
29	*/
30
31
32	/*******************************************************************************
33	* Header Files *
34	*******************************************************************************/
35	#define LOG_GROUP RTLOGGROUP_DIR
36	#include <iprt/asm.h>
37	#include <iprt/file.h>
38	#include <iprt/mem.h>
39	#include <iprt/assert.h>
40	#include <iprt/string.h>
41	#include <iprt/err.h>
42	#include <iprt/log.h>
43	#include <iprt/thread.h>
44	#include <iprt/semaphore.h>
45	#include "internal/fileaio.h"
46
47	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_FREEBSD)
48	# include <sys/types.h>
49	# include <sys/sysctl.h> /* for sysctlbyname */
50	#endif
51	#if defined(RT_OS_FREEBSD)
52	# include <fcntl.h> /* O_SYNC */
53	#endif
54	#include <aio.h>
55	#include <errno.h>
56	#include <time.h>
57
58	/*
59	* Linux does not define this value.
60	* Just define it with really big
61	* value.
62	*/
63	#ifndef AIO_LISTIO_MAX
64	# define AIO_LISTIO_MAX UINT32_MAX
65	#endif
66
67	#if 0 /* Only used for debugging */
68	# undef AIO_LISTIO_MAX
69	# define AIO_LISTIO_MAX 16
70	#endif
71
72	/** Invalid entry in the waiting array. */
73	#define RTFILEAIOCTX_WAIT_ENTRY_INVALID (~0U)
74
75	/*******************************************************************************
76	* Structures and Typedefs *
77	*******************************************************************************/
78	/**
79	* Async I/O request state.
80	*/
81	typedef struct RTFILEAIOREQINTERNAL
82	{
83	/** The aio control block. FIRST ELEMENT! */
84	struct aiocb AioCB;
85	/** Next element in the chain. */
86	struct RTFILEAIOREQINTERNAL *pNext;
87	/** Previous element in the chain. */
88	struct RTFILEAIOREQINTERNAL *pPrev;
89	/** Current state the request is in. */
90	RTFILEAIOREQSTATE enmState;
91	/** Flag whether this is a flush request. */
92	bool fFlush;
93	/** Flag indicating if the request was canceled. */
94	volatile bool fCanceled;
95	/** Opaque user data. */
96	void *pvUser;
97	/** Number of bytes actually transfered. */
98	size_t cbTransfered;
99	/** Status code. */
100	int Rc;
101	/** Completion context we are assigned to. */
102	struct RTFILEAIOCTXINTERNAL *pCtxInt;
103	/** Entry in the waiting list the request is in. */
104	unsigned iWaitingList;
105	/** Magic value (RTFILEAIOREQ_MAGIC). */
106	uint32_t u32Magic;
107	} RTFILEAIOREQINTERNAL, *PRTFILEAIOREQINTERNAL;
108
109	/**
110	* Async I/O completion context state.
111	*/
112	typedef struct RTFILEAIOCTXINTERNAL
113	{
114	/** Current number of requests active on this context. */
115	volatile int32_t cRequests;
116	/** Maximum number of requests this context can handle. */
117	uint32_t cMaxRequests;
118	/** The ID of the thread which is currently waiting for requests. */
119	volatile RTTHREAD hThreadWait;
120	/** Flag whether the thread was woken up. */
121	volatile bool fWokenUp;
122	/** Flag whether the thread is currently waiting in the syscall. */
123	volatile bool fWaiting;
124	/** Magic value (RTFILEAIOCTX_MAGIC). */
125	uint32_t u32Magic;
126	/** Flag whether the thread was woken up due to a internal event. */
127	volatile bool fWokenUpInternal;
128	/** List of new requests which needs to be inserted into apReqs by the
129	* waiting thread. */
130	volatile PRTFILEAIOREQINTERNAL apReqsNewHead[5];
131	/** Special entry for requests which are canceled. Because only one
132	* request can be canceled at a time and the thread canceling the request
133	* has to wait we need only one entry. */
134	volatile PRTFILEAIOREQINTERNAL pReqToCancel;
135	/** Event semaphore the canceling thread is waiting for completion of
136	* the operation. */
137	RTSEMEVENT SemEventCancel;
138	/** Head of submitted elements waiting to get into the array. */
139	PRTFILEAIOREQINTERNAL pReqsWaitHead;
140	/** Tail of submitted elements waiting to get into the array. */
141	PRTFILEAIOREQINTERNAL pReqsWaitTail;
142	/** Maximum number of elements in the waiting array. */
143	unsigned cReqsWaitMax;
144	/** First free slot in the waiting list. */
145	unsigned iFirstFree;
146	/** List of requests we are currently waiting on.
147	* Size depends on cMaxRequests and AIO_LISTIO_MAX. */
148	volatile PRTFILEAIOREQINTERNAL apReqs[1];
149	} RTFILEAIOCTXINTERNAL, *PRTFILEAIOCTXINTERNAL;
150
151	/**
152	* Internal worker for waking up the waiting thread.
153	*/
154	static void rtFileAioCtxWakeup(PRTFILEAIOCTXINTERNAL pCtxInt)
155	{
156	/*
157	* Read the thread handle before the status flag.
158	* If we read the handle after the flag we might
159	* end up with an invalid handle because the thread
160	* waiting in RTFileAioCtxWakeup() might get scheduled
161	* before we read the flag and returns.
162	* We can ensure that the handle is valid if fWaiting is true
163	* when reading the handle before the status flag.
164	*/
165	RTTHREAD hThread;
166	ASMAtomicReadHandle(&pCtxInt->hThreadWait, &hThread);
167	bool fWaiting = ASMAtomicReadBool(&pCtxInt->fWaiting);
168	if (fWaiting)
169	{
170	/*
171	* If a thread waits the handle must be valid.
172	* It is possible that the thread returns from
173	* aio_suspend() before the signal is send.
174	* This is no problem because we already set fWokenUp
175	* to true which will let the thread return VERR_INTERRUPTED
176	* and the next call to RTFileAioCtxWait() will not
177	* return VERR_INTERRUPTED because signals are not saved
178	* and will simply vanish if the destination thread can't
179	* receive it.
180	*/
181	Assert(hThread != NIL_RTTHREAD);
182	RTThreadPoke(hThread);
183	}
184	}
185
186	/**
187	* Internal worker processing events and inserting new requests into the waiting list.
188	*/
189	static int rtFileAioCtxProcessEvents(PRTFILEAIOCTXINTERNAL pCtxInt)
190	{
191	int rc = VINF_SUCCESS;
192
193	/* Process new requests first. */
194	bool fWokenUp = ASMAtomicXchgBool(&pCtxInt->fWokenUpInternal, false);
195	if (fWokenUp)
196	{
197	for (unsigned iSlot = 0; iSlot < RT_ELEMENTS(pCtxInt->apReqsNewHead); iSlot++)
198	{
199	PRTFILEAIOREQINTERNAL pReqHead = (PRTFILEAIOREQINTERNAL)ASMAtomicXchgPtr((void* volatile*)&pCtxInt->apReqsNewHead[iSlot],
200	NULL);
201
202	while ( (pCtxInt->iFirstFree < pCtxInt->cReqsWaitMax)
203	&& pReqHead)
204	{
205	pCtxInt->apReqs[pCtxInt->iFirstFree] = pReqHead;
206	pReqHead->iWaitingList = pCtxInt->iFirstFree;
207	pReqHead = pReqHead->pNext;
208
209	/* Clear pointer to next and previous element just for safety. */
210	pCtxInt->apReqs[pCtxInt->iFirstFree]->pNext = NULL;
211	pCtxInt->apReqs[pCtxInt->iFirstFree]->pPrev = NULL;
212	pCtxInt->iFirstFree++;
213
214	Assert( (pCtxInt->iFirstFree <= pCtxInt->cMaxRequests)
215	&& (pCtxInt->iFirstFree <= pCtxInt->cReqsWaitMax));
216	}
217
218	/* Append the rest to the wait list. */
219	if (pReqHead)
220	{
221	if (!pCtxInt->pReqsWaitHead)
222	{
223	Assert(!pCtxInt->pReqsWaitTail);
224	pCtxInt->pReqsWaitHead = pReqHead;
225	pReqHead->pPrev = NULL;
226	}
227	else
228	{
229	AssertPtr(pCtxInt->pReqsWaitTail);
230
231	pCtxInt->pReqsWaitTail->pNext = pReqHead;
232	pReqHead->pPrev = pCtxInt->pReqsWaitTail;
233	}
234
235	/* Update tail. */
236	while (pReqHead->pNext)
237	pReqHead = pReqHead->pNext;
238
239	pCtxInt->pReqsWaitTail = pReqHead;
240	pCtxInt->pReqsWaitTail->pNext = NULL;
241	}
242	}
243
244	/* Check if a request needs to be canceled. */
245	PRTFILEAIOREQINTERNAL pReqToCancel = (PRTFILEAIOREQINTERNAL)ASMAtomicReadPtr((void* volatile*)&pCtxInt->pReqToCancel);
246	if (pReqToCancel)
247	{
248	/* The request can be in the array waiting for completion or still in the list because it is full. */
249	if (pReqToCancel->iWaitingList != RTFILEAIOCTX_WAIT_ENTRY_INVALID)
250	{
251	/* Put it out of the waiting list. */
252	pCtxInt->apReqs[pReqToCancel->iWaitingList] = pCtxInt->apReqs[--pCtxInt->iFirstFree];
253	pCtxInt->apReqs[pReqToCancel->iWaitingList]->iWaitingList = pReqToCancel->iWaitingList;
254	}
255	else
256	{
257	/* Unlink from the waiting list. */
258	PRTFILEAIOREQINTERNAL pPrev = pReqToCancel->pPrev;
259	PRTFILEAIOREQINTERNAL pNext = pReqToCancel->pNext;
260
261	if (pNext)
262	pNext->pPrev = pPrev;
263	else
264	{
265	/* We canceled the tail. */
266	pCtxInt->pReqsWaitTail = pPrev;
267	}
268
269	if (pPrev)
270	pPrev->pNext = pNext;
271	else
272	{
273	/* We canceled the head. */
274	pCtxInt->pReqsWaitHead = pNext;
275	}
276	}
277
278	ASMAtomicDecS32(&pCtxInt->cRequests);
279	RTSemEventSignal(pCtxInt->SemEventCancel);
280	}
281	}
282	else
283	{
284	if (ASMAtomicXchgBool(&pCtxInt->fWokenUp, false))
285	rc = VERR_INTERRUPTED;
286	}
287
288	return rc;
289	}
290
291	RTR3DECL(int) RTFileAioGetLimits(PRTFILEAIOLIMITS pAioLimits)
292	{
293	int rcBSD = 0;
294	AssertPtrReturn(pAioLimits, VERR_INVALID_POINTER);
295
296	#if defined(RT_OS_DARWIN)
297	int cReqsOutstandingMax = 0;
298	size_t cbParameter = sizeof(int);
299
300	rcBSD = sysctlbyname("kern.aioprocmax", /* name */
301	&cReqsOutstandingMax, /* Where to store the old value. */
302	&cbParameter, /* Size of the memory pointed to. */
303	NULL, /* Where the new value is located. */
304	NULL); /* Where the size of the new value is stored. */
305	if (rcBSD == -1)
306	return RTErrConvertFromErrno(errno);
307
308	pAioLimits->cReqsOutstandingMax = cReqsOutstandingMax;
309	pAioLimits->cbBufferAlignment = 0;
310	#elif defined(RT_OS_FREEBSD)
311	/*
312	* The AIO API is implemented in a kernel module which is not
313	* loaded by default.
314	* If it is loaded there are additional sysctl parameters.
315	*/
316	int cReqsOutstandingMax = 0;
317	size_t cbParameter = sizeof(int);
318
319	rcBSD = sysctlbyname("vfs.aio.max_aio_per_proc", /* name */
320	&cReqsOutstandingMax, /* Where to store the old value. */
321	&cbParameter, /* Size of the memory pointed to. */
322	NULL, /* Where the new value is located. */
323	NULL); /* Where the size of the new value is stored. */
324	if (rcBSD == -1)
325	{
326	/* ENOENT means the value is unknown thus the module is not loaded. */
327	if (errno == ENOENT)
328	return VERR_NOT_SUPPORTED;
329	else
330	return RTErrConvertFromErrno(errno);
331	}
332
333	pAioLimits->cReqsOutstandingMax = cReqsOutstandingMax;
334	pAioLimits->cbBufferAlignment = 0;
335	#else
336	pAioLimits->cReqsOutstandingMax = RTFILEAIO_UNLIMITED_REQS;
337	pAioLimits->cbBufferAlignment = 0;
338	#endif
339
340	return VINF_SUCCESS;
341	}
342
343	RTR3DECL(int) RTFileAioReqCreate(PRTFILEAIOREQ phReq)
344	{
345	AssertPtrReturn(phReq, VERR_INVALID_POINTER);
346
347	PRTFILEAIOREQINTERNAL pReqInt = (PRTFILEAIOREQINTERNAL)RTMemAllocZ(sizeof(RTFILEAIOREQINTERNAL));
348	if (RT_UNLIKELY(!pReqInt))
349	return VERR_NO_MEMORY;
350
351	pReqInt->pCtxInt = NULL;
352	pReqInt->u32Magic = RTFILEAIOREQ_MAGIC;
353	pReqInt->iWaitingList = RTFILEAIOCTX_WAIT_ENTRY_INVALID;
354	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
355
356	*phReq = (RTFILEAIOREQ)pReqInt;
357
358	return VINF_SUCCESS;
359	}
360
361
362	RTDECL(int) RTFileAioReqDestroy(RTFILEAIOREQ hReq)
363	{
364	/*
365	* Validate the handle and ignore nil.
366	*/
367	if (hReq == NIL_RTFILEAIOREQ)
368	return VINF_SUCCESS;
369	PRTFILEAIOREQINTERNAL pReqInt = hReq;
370	RTFILEAIOREQ_VALID_RETURN(pReqInt);
371	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
372
373	/*
374	* Trash the magic and free it.
375	*/
376	ASMAtomicUoWriteU32(&pReqInt->u32Magic, ~RTFILEAIOREQ_MAGIC);
377	RTMemFree(pReqInt);
378	return VINF_SUCCESS;
379	}
380
381	/**
382	* Worker setting up the request.
383	*/
384	DECLINLINE(int) rtFileAioReqPrepareTransfer(RTFILEAIOREQ hReq, RTFILE hFile,
385	unsigned uTransferDirection,
386	RTFOFF off, void *pvBuf, size_t cbTransfer,
387	void *pvUser)
388	{
389	/*
390	* Validate the input.
391	*/
392	PRTFILEAIOREQINTERNAL pReqInt = hReq;
393	RTFILEAIOREQ_VALID_RETURN(pReqInt);
394	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
395	Assert(hFile != NIL_RTFILE);
396	AssertPtr(pvBuf);
397	Assert(off >= 0);
398	Assert(cbTransfer > 0);
399
400	memset(&pReqInt->AioCB, 0, sizeof(struct aiocb));
401	pReqInt->AioCB.aio_lio_opcode = uTransferDirection;
402	pReqInt->AioCB.aio_fildes = (int)hFile;
403	pReqInt->AioCB.aio_offset = off;
404	pReqInt->AioCB.aio_nbytes = cbTransfer;
405	pReqInt->AioCB.aio_buf = pvBuf;
406	pReqInt->pvUser = pvUser;
407	pReqInt->pCtxInt = NULL;
408	pReqInt->Rc = VERR_FILE_AIO_IN_PROGRESS;
409	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
410
411	return VINF_SUCCESS;
412	}
413
414
415	RTDECL(int) RTFileAioReqPrepareRead(RTFILEAIOREQ hReq, RTFILE hFile, RTFOFF off,
416	void pvBuf, size_t cbRead, void pvUser)
417	{
418	return rtFileAioReqPrepareTransfer(hReq, hFile, LIO_READ,
419	off, pvBuf, cbRead, pvUser);
420	}
421
422
423	RTDECL(int) RTFileAioReqPrepareWrite(RTFILEAIOREQ hReq, RTFILE hFile, RTFOFF off,
424	void pvBuf, size_t cbWrite, void pvUser)
425	{
426	return rtFileAioReqPrepareTransfer(hReq, hFile, LIO_WRITE,
427	off, pvBuf, cbWrite, pvUser);
428	}
429
430
431	RTDECL(int) RTFileAioReqPrepareFlush(RTFILEAIOREQ hReq, RTFILE hFile, void *pvUser)
432	{
433	PRTFILEAIOREQINTERNAL pReqInt = (PRTFILEAIOREQINTERNAL)hReq;
434
435	RTFILEAIOREQ_VALID_RETURN(pReqInt);
436	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
437	Assert(hFile != NIL_RTFILE);
438
439	pReqInt->fFlush = true;
440	pReqInt->AioCB.aio_fildes = (int)hFile;
441	pReqInt->pvUser = pvUser;
442	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
443
444	return VINF_SUCCESS;
445	}
446
447
448	RTDECL(void *) RTFileAioReqGetUser(RTFILEAIOREQ hReq)
449	{
450	PRTFILEAIOREQINTERNAL pReqInt = hReq;
451	RTFILEAIOREQ_VALID_RETURN_RC(pReqInt, NULL);
452
453	return pReqInt->pvUser;
454	}
455
456
457	RTDECL(int) RTFileAioReqCancel(RTFILEAIOREQ hReq)
458	{
459	PRTFILEAIOREQINTERNAL pReqInt = hReq;
460	RTFILEAIOREQ_VALID_RETURN(pReqInt);
461	RTFILEAIOREQ_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_NOT_SUBMITTED);
462
463	ASMAtomicXchgBool(&pReqInt->fCanceled, true);
464
465	int rcPosix = aio_cancel(pReqInt->AioCB.aio_fildes, &pReqInt->AioCB);
466
467	if (rcPosix == AIO_CANCELED)
468	{
469	PRTFILEAIOCTXINTERNAL pCtxInt = pReqInt->pCtxInt;
470	/*
471	* Notify the waiting thread that the request was canceled.
472	*/
473	AssertMsg(VALID_PTR(pCtxInt),
474	("Invalid state. Request was canceled but wasn't submitted\n"));
475
476	Assert(!pCtxInt->pReqToCancel);
477	ASMAtomicWritePtr((void* volatile*)&pCtxInt->pReqToCancel, pReqInt);
478	rtFileAioCtxWakeup(pCtxInt);
479
480	/* Wait for acknowledge. */
481	int rc = RTSemEventWait(pCtxInt->SemEventCancel, RT_INDEFINITE_WAIT);
482	AssertRC(rc);
483
484	ASMAtomicWritePtr((void* volatile*)&pCtxInt->pReqToCancel, NULL);
485	pReqInt->Rc = VERR_FILE_AIO_CANCELED;
486	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
487	return VINF_SUCCESS;
488	}
489	else if (rcPosix == AIO_ALLDONE)
490	return VERR_FILE_AIO_COMPLETED;
491	else if (rcPosix == AIO_NOTCANCELED)
492	return VERR_FILE_AIO_IN_PROGRESS;
493	else
494	return RTErrConvertFromErrno(errno);
495	}
496
497
498	RTDECL(int) RTFileAioReqGetRC(RTFILEAIOREQ hReq, size_t *pcbTransfered)
499	{
500	PRTFILEAIOREQINTERNAL pReqInt = hReq;
501	RTFILEAIOREQ_VALID_RETURN(pReqInt);
502	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
503	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, PREPARED, VERR_FILE_AIO_NOT_SUBMITTED);
504	AssertPtrNull(pcbTransfered);
505
506	if ( (RT_SUCCESS(pReqInt->Rc))
507	&& (pcbTransfered))
508	*pcbTransfered = pReqInt->cbTransfered;
509
510	return pReqInt->Rc;
511	}
512
513
514	RTDECL(int) RTFileAioCtxCreate(PRTFILEAIOCTX phAioCtx, uint32_t cAioReqsMax)
515	{
516	PRTFILEAIOCTXINTERNAL pCtxInt;
517	unsigned cReqsWaitMax;
518
519	AssertPtrReturn(phAioCtx, VERR_INVALID_POINTER);
520
521	if (cAioReqsMax == RTFILEAIO_UNLIMITED_REQS)
522	return VERR_OUT_OF_RANGE;
523
524	cReqsWaitMax = RT_MIN(cAioReqsMax, AIO_LISTIO_MAX);
525
526	pCtxInt = (PRTFILEAIOCTXINTERNAL)RTMemAllocZ( sizeof(RTFILEAIOCTXINTERNAL)
527	+ cReqsWaitMax * sizeof(PRTFILEAIOREQINTERNAL));
528	if (RT_UNLIKELY(!pCtxInt))
529	return VERR_NO_MEMORY;
530
531	/* Create event semaphore. */
532	int rc = RTSemEventCreate(&pCtxInt->SemEventCancel);
533	if (RT_FAILURE(rc))
534	{
535	RTMemFree(pCtxInt);
536	return rc;
537	}
538
539	pCtxInt->u32Magic = RTFILEAIOCTX_MAGIC;
540	pCtxInt->cMaxRequests = cAioReqsMax;
541	pCtxInt->cReqsWaitMax = cReqsWaitMax;
542	*phAioCtx = (RTFILEAIOCTX)pCtxInt;
543
544	return VINF_SUCCESS;
545	}
546
547
548	RTDECL(int) RTFileAioCtxDestroy(RTFILEAIOCTX hAioCtx)
549	{
550	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
551
552	AssertPtrReturn(pCtxInt, VERR_INVALID_HANDLE);
553
554	if (RT_UNLIKELY(pCtxInt->cRequests))
555	return VERR_FILE_AIO_BUSY;
556
557	RTSemEventDestroy(pCtxInt->SemEventCancel);
558	RTMemFree(pCtxInt);
559
560	return VINF_SUCCESS;
561	}
562
563
564	RTDECL(uint32_t) RTFileAioCtxGetMaxReqCount(RTFILEAIOCTX hAioCtx)
565	{
566	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
567
568	if (hAioCtx == NIL_RTFILEAIOCTX)
569	return RTFILEAIO_UNLIMITED_REQS;
570	else
571	return pCtxInt->cMaxRequests;
572	}
573
574	RTDECL(int) RTFileAioCtxAssociateWithFile(RTFILEAIOCTX hAioCtx, RTFILE hFile)
575	{
576	return VINF_SUCCESS;
577	}
578
579	RTDECL(int) RTFileAioCtxSubmit(RTFILEAIOCTX hAioCtx, PRTFILEAIOREQ pahReqs, size_t cReqs)
580	{
581	int rc = VINF_SUCCESS;
582	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
583
584	/* Parameter checks */
585	AssertPtrReturn(pCtxInt, VERR_INVALID_HANDLE);
586	AssertReturn(cReqs != 0, VERR_INVALID_POINTER);
587	AssertPtrReturn(pahReqs, VERR_INVALID_PARAMETER);
588
589	/* Check that we don't exceed the limit */
590	if (ASMAtomicUoReadS32(&pCtxInt->cRequests) + cReqs > pCtxInt->cMaxRequests)
591	return VERR_FILE_AIO_LIMIT_EXCEEDED;
592
593	PRTFILEAIOREQINTERNAL pHead = NULL;
594
595	do
596	{
597	int rcPosix = 0;
598	size_t cReqsSubmit = 0;
599	size_t i = 0;
600	PRTFILEAIOREQINTERNAL pReqInt;
601
602	while ( (i < cReqs)
603	&& (i < AIO_LISTIO_MAX))
604	{
605	pReqInt = pahReqs[i];
606	if (RTFILEAIOREQ_IS_NOT_VALID(pReqInt))
607	{
608	/* Undo everything and stop submitting. */
609	for (size_t iUndo = 0; iUndo < i; iUndo++)
610	{
611	pReqInt = pahReqs[iUndo];
612	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
613	pReqInt->pCtxInt = NULL;
614
615	/* Unlink from the list again. */
616	PRTFILEAIOREQINTERNAL pNext, pPrev;
617	pNext = pReqInt->pNext;
618	pPrev = pReqInt->pPrev;
619	if (pNext)
620	pNext->pPrev = pPrev;
621	if (pPrev)
622	pPrev->pNext = pNext;
623	else
624	pHead = pNext;
625	}
626	rc = VERR_INVALID_HANDLE;
627	break;
628	}
629
630	pReqInt->pCtxInt = pCtxInt;
631
632	/* Link them together. */
633	pReqInt->pNext = pHead;
634	if (pHead)
635	pHead->pPrev = pReqInt;
636	pReqInt->pPrev = NULL;
637	pHead = pReqInt;
638	RTFILEAIOREQ_SET_STATE(pReqInt, SUBMITTED);
639
640	if (pReqInt->fFlush)
641	break;
642
643	cReqsSubmit++;
644	i++;
645	}
646
647	if (cReqsSubmit)
648	{
649	rcPosix = lio_listio(LIO_NOWAIT, (struct aiocb **)pahReqs, cReqsSubmit, NULL);
650	if (RT_UNLIKELY(rcPosix < 0))
651	{
652	if (errno == EAGAIN)
653	rc = VERR_FILE_AIO_INSUFFICIENT_RESSOURCES;
654	else
655	rc = RTErrConvertFromErrno(errno);
656
657	/* Check which ones were not submitted. */
658	for (i = 0; i < cReqs; i++)
659	{
660	pReqInt = pahReqs[i];
661
662	rcPosix = aio_error(&pReqInt->AioCB);
663	Assert(rcPosix != 0);
664
665	if (rcPosix != EINPROGRESS)
666	{
667	cReqsSubmit--;
668
669	if (rcPosix == EINVAL)
670	{
671	/* Was not submitted. */
672	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
673	}
674	else
675	{
676	/* An error occurred. */
677	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
678
679	/*
680	* Looks like Apple and glibc interpret the standard in different ways.
681	* glibc returns the error code which would be in errno but Apple returns
682	* -1 and sets errno to the appropriate value
683	*/
684	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_FREEBSD)
685	Assert(rcPosix == -1);
686	pReqInt->Rc = RTErrConvertFromErrno(errno);
687	#elif defined(RT_OS_LINUX)
688	pReqInt->Rc = RTErrConvertFromErrno(rcPosix);
689	#endif
690	pReqInt->cbTransfered = 0;
691	}
692	/* Unlink from the list. */
693	PRTFILEAIOREQINTERNAL pNext, pPrev;
694	pNext = pReqInt->pNext;
695	pPrev = pReqInt->pPrev;
696	if (pNext)
697	pNext->pPrev = pPrev;
698	if (pPrev)
699	pPrev->pNext = pNext;
700	else
701	pHead = pNext;
702	}
703	}
704
705	break;
706	}
707
708	ASMAtomicAddS32(&pCtxInt->cRequests, cReqsSubmit);
709	cReqs -= cReqsSubmit;
710	pahReqs += cReqsSubmit;
711	}
712
713	/*
714	* Check if we have a flush request now.
715	* If not we hit the AIO_LISTIO_MAX limit
716	* and will continue submitting requests
717	* above.
718	*/
719	if (cReqs)
720	{
721	pReqInt = pahReqs[0];
722	RTFILEAIOREQ_VALID_RETURN(pReqInt);
723
724
725	if (pReqInt->fFlush)
726	{
727	/*
728	* lio_listio does not work with flush requests so
729	* we have to use aio_fsync directly.
730	*/
731	rcPosix = aio_fsync(O_SYNC, &pReqInt->AioCB);
732	if (RT_UNLIKELY(rcPosix < 0))
733	{
734	rc = RTErrConvertFromErrno(errno);
735	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
736	pReqInt->Rc = rc;
737	pReqInt->cbTransfered = 0;
738
739	/* Unlink from the list. */
740	PRTFILEAIOREQINTERNAL pNext, pPrev;
741	pNext = pReqInt->pNext;
742	pPrev = pReqInt->pPrev;
743	if (pNext)
744	pNext->pPrev = pPrev;
745	if (pPrev)
746	pPrev->pNext = pNext;
747	else
748	pHead = pNext;
749	break;
750	}
751
752	ASMAtomicIncS32(&pCtxInt->cRequests);
753	cReqs--;
754	pahReqs++;
755	}
756	}
757	} while (cReqs);
758
759	if (pHead)
760	{
761	/*
762	* Forward successfully submitted requests to the thread waiting for requests.
763	* We search for a free slot first and if we don't find one
764	* we will grab the first one and append our list to the existing entries.
765	*/
766	unsigned iSlot = 0;
767	while ( (iSlot < RT_ELEMENTS(pCtxInt->apReqsNewHead))
768	&& !ASMAtomicCmpXchgPtr((void * volatile *)&pCtxInt->apReqsNewHead[iSlot], pHead, NULL))
769	iSlot++;
770
771	if (iSlot == RT_ELEMENTS(pCtxInt->apReqsNewHead))
772	{
773	/* Nothing found. */
774	PRTFILEAIOREQINTERNAL pOldHead = (PRTFILEAIOREQINTERNAL)ASMAtomicXchgPtr((void * volatile *)&pCtxInt->apReqsNewHead[0],
775	NULL);
776
777	/* Find the end of the current head and link the old list to the current. */
778	PRTFILEAIOREQINTERNAL pTail = pHead;
779	while (pTail->pNext)
780	pTail = pTail->pNext;
781
782	pTail->pNext = pOldHead;
783
784	ASMAtomicXchgPtr((void * volatile *)&pCtxInt->apReqsNewHead[0], pHead);
785	}
786
787	/* Set the internal wakeup flag and wakeup the thread if possible. */
788	bool fWokenUp = ASMAtomicXchgBool(&pCtxInt->fWokenUpInternal, true);
789	if (!fWokenUp)
790	rtFileAioCtxWakeup(pCtxInt);
791	}
792
793	return rc;
794	}
795
796
797	RTDECL(int) RTFileAioCtxWait(RTFILEAIOCTX hAioCtx, size_t cMinReqs, unsigned cMillisTimeout,
798	PRTFILEAIOREQ pahReqs, size_t cReqs, uint32_t *pcReqs)
799	{
800	int rc = VINF_SUCCESS;
801	int cRequestsCompleted = 0;
802	PRTFILEAIOCTXINTERNAL pCtxInt = (PRTFILEAIOCTXINTERNAL)hAioCtx;
803	struct timespec Timeout;
804	struct timespec *pTimeout = NULL;
805	uint64_t StartNanoTS = 0;
806
807	/* Check parameters. */
808	AssertPtrReturn(pCtxInt, VERR_INVALID_HANDLE);
809	AssertPtrReturn(pcReqs, VERR_INVALID_POINTER);
810	AssertPtrReturn(pahReqs, VERR_INVALID_POINTER);
811	AssertReturn(cReqs != 0, VERR_INVALID_PARAMETER);
812	AssertReturn(cReqs >= cMinReqs, VERR_OUT_OF_RANGE);
813
814	if (RT_UNLIKELY(ASMAtomicReadS32(&pCtxInt->cRequests) == 0))
815	return VERR_FILE_AIO_NO_REQUEST;
816
817	if (cMillisTimeout != RT_INDEFINITE_WAIT)
818	{
819	Timeout.tv_sec = cMillisTimeout / 1000;
820	Timeout.tv_nsec = (cMillisTimeout % 1000) * 1000000;
821	pTimeout = &Timeout;
822	StartNanoTS = RTTimeNanoTS();
823	}
824
825	/* Wait for at least one. */
826	if (!cMinReqs)
827	cMinReqs = 1;
828
829	/* For the wakeup call. */
830	Assert(pCtxInt->hThreadWait == NIL_RTTHREAD);
831	ASMAtomicWriteHandle(&pCtxInt->hThreadWait, RTThreadSelf());
832
833	/* Update the waiting list once before we enter the loop. */
834	rc = rtFileAioCtxProcessEvents(pCtxInt);
835
836	while ( cMinReqs
837	&& RT_SUCCESS_NP(rc))
838	{
839	ASMAtomicXchgBool(&pCtxInt->fWaiting, true);
840	int rcPosix = aio_suspend((const struct aiocb * const *)pCtxInt->apReqs,
841	pCtxInt->iFirstFree, pTimeout);
842	ASMAtomicXchgBool(&pCtxInt->fWaiting, false);
843	if (rcPosix < 0)
844	{
845	/* Check that this is an external wakeup event. */
846	if (errno == EINTR)
847	rc = rtFileAioCtxProcessEvents(pCtxInt);
848	else
849	rc = RTErrConvertFromErrno(errno);
850	}
851	else
852	{
853	/* Requests finished. */
854	unsigned iReqCurr = 0;
855	unsigned cDone = 0;
856
857	/* Remove completed requests from the waiting list. */
858	while ( (iReqCurr < pCtxInt->iFirstFree)
859	&& (cDone < cReqs))
860	{
861	PRTFILEAIOREQINTERNAL pReq = pCtxInt->apReqs[iReqCurr];
862	int rcReq = aio_error(&pReq->AioCB);
863
864	if (rcReq != EINPROGRESS)
865	{
866	/* Completed store the return code. */
867	if (rcReq == 0)
868	{
869	pReq->Rc = VINF_SUCCESS;
870	/* Call aio_return() to free ressources. */
871	pReq->cbTransfered = aio_return(&pReq->AioCB);
872	}
873	else
874	{
875	#if defined(RT_OS_DARWIN) \|\| defined(RT_OS_FREEBSD)
876	pReq->Rc = RTErrConvertFromErrno(errno);
877	#else
878	pReq->Rc = RTErrConvertFromErrno(rcReq);
879	#endif
880	}
881
882	/* Mark the request as finished. */
883	RTFILEAIOREQ_SET_STATE(pReq, COMPLETED);
884	cDone++;
885
886	/* If there are other entries waiting put the head into the now free entry. */
887	if (pCtxInt->pReqsWaitHead)
888	{
889	PRTFILEAIOREQINTERNAL pReqInsert = pCtxInt->pReqsWaitHead;
890
891	pCtxInt->pReqsWaitHead = pReqInsert->pNext;
892	if (!pCtxInt->pReqsWaitHead)
893	{
894	/* List is empty now. Clear tail too. */
895	pCtxInt->pReqsWaitTail = NULL;
896	}
897
898	pReqInsert->iWaitingList = pReq->iWaitingList;
899	pCtxInt->apReqs[pReqInsert->iWaitingList] = pReqInsert;
900	iReqCurr++;
901	}
902	else
903	{
904	/*
905	* Move the last entry into the current position to avoid holes
906	* but only if it is not the last element already.
907	*/
908	if (pReq->iWaitingList < pCtxInt->iFirstFree - 1)
909	{
910	pCtxInt->apReqs[pReq->iWaitingList] = pCtxInt->apReqs[--pCtxInt->iFirstFree];
911	pCtxInt->apReqs[pReq->iWaitingList]->iWaitingList = pReq->iWaitingList;
912	pCtxInt->apReqs[pCtxInt->iFirstFree] = NULL;
913	}
914	else
915	pCtxInt->iFirstFree--;
916	}
917
918	/* Put the request into the completed list. */
919	pahReqs[cRequestsCompleted++] = pReq;
920	pReq->iWaitingList = RTFILEAIOCTX_WAIT_ENTRY_INVALID;
921	}
922	else
923	iReqCurr++;
924	}
925
926	AssertMsg( (cDone <= cMinReqs)
927	&& (cDone <= cReqs), ("Overflow cReqs=%u cMinReqs=%u cDone=%u\n",
928	cReqs, cMinReqs, cDone));
929	cReqs -= cDone;
930	cMinReqs -= cDone;
931	ASMAtomicSubS32(&pCtxInt->cRequests, cDone);
932
933	if ((cMillisTimeout != RT_INDEFINITE_WAIT) && (cMinReqs > 0))
934	{
935	uint64_t TimeDiff;
936
937	/* Recalculate the timeout. */
938	TimeDiff = RTTimeSystemNanoTS() - StartNanoTS;
939	Timeout.tv_sec = Timeout.tv_sec - (TimeDiff / 1000000);
940	Timeout.tv_nsec = Timeout.tv_nsec - (TimeDiff % 1000000);
941	}
942
943	/* Check for new elements. */
944	rc = rtFileAioCtxProcessEvents(pCtxInt);
945	}
946	}
947
948	*pcReqs = cRequestsCompleted;
949	Assert(pCtxInt->hThreadWait == RTThreadSelf());
950	ASMAtomicWriteHandle(&pCtxInt->hThreadWait, NIL_RTTHREAD);
951
952	return rc;
953	}
954
955
956	RTDECL(int) RTFileAioCtxWakeup(RTFILEAIOCTX hAioCtx)
957	{
958	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
959	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
960
961	/** @todo r=bird: Define the protocol for how to resume work after calling
962	* this function. */
963
964	bool fWokenUp = ASMAtomicXchgBool(&pCtxInt->fWokenUp, true);
965	if (!fWokenUp)
966	rtFileAioCtxWakeup(pCtxInt);
967
968	return VINF_SUCCESS;
969	}
970

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

source: vbox/trunk/src/VBox/Runtime/r3/posix/fileaio-posix.cpp@ 23396

Download in other formats: