fileaio-linux.cpp@ 34244

Last change on this file since 34244 was 33540, checked in by vboxsync, 14 years ago
*: spelling fixes, thanks Timeless!
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 25.7 KB

Line
1	/* $Id: fileaio-linux.cpp 33540 2010-10-28 09:27:05Z vboxsync $ */
2	/** @file
3	* IPRT - File async I/O, native implementation for the Linux host platform.
4	*/
5
6	/*
7	* Copyright (C) 2006-2007 Oracle Corporation
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
27	/** @page pg_rtfileaio_linux RTFile Async I/O - Linux Implementation Notes
28	* @internal
29	*
30	* Linux implements the kernel async I/O API through the io_* syscalls. They are
31	* not exposed in the glibc (the aio_* API uses userspace threads and blocking
32	* I/O operations to simulate async behavior). There is an external library
33	* called libaio which implements these syscalls but because we don't want to
34	* have another dependency and this library is not installed by default and the
35	* interface is really simple we use the kernel interface directly using wrapper
36	* functions.
37	*
38	* The interface has some limitations. The first one is that the file must be
39	* opened with O_DIRECT. This disables caching done by the kernel which can be
40	* compensated if the user of this API implements caching itself. The next
41	* limitation is that data buffers must be aligned at a 512 byte boundary or the
42	* request will fail.
43	*/
44	/** @todo r=bird: What's this about "must be opened with O_DIRECT"? An
45	* explanation would be nice, esp. seeing what Linus is quoted saying
46	* about it in the open man page... */
47
48	/*******************************************************************************
49	* Header Files *
50	*******************************************************************************/
51	#define LOG_GROUP RTLOGGROUP_FILE
52	#include <iprt/asm.h>
53	#include <iprt/mem.h>
54	#include <iprt/assert.h>
55	#include <iprt/string.h>
56	#include <iprt/err.h>
57	#include <iprt/log.h>
58	#include <iprt/thread.h>
59	#include "internal/fileaio.h"
60
61	#include <unistd.h>
62	#include <sys/syscall.h>
63	#include <errno.h>
64
65	#include <iprt/file.h>
66
67
68	/*******************************************************************************
69	* Structures and Typedefs *
70	*******************************************************************************/
71	/** The async I/O context handle */
72	typedef unsigned long LNXKAIOCONTEXT;
73
74	/**
75	* Supported commands for the iocbs
76	*/
77	enum
78	{
79	LNXKAIO_IOCB_CMD_READ = 0,
80	LNXKAIO_IOCB_CMD_WRITE = 1,
81	LNXKAIO_IOCB_CMD_FSYNC = 2,
82	LNXKAIO_IOCB_CMD_FDSYNC = 3
83	};
84
85	/**
86	* The iocb structure of a request which is passed to the kernel.
87	*
88	* We redefined this here because the version in the header lacks padding
89	* for 32bit.
90	*/
91	typedef struct LNXKAIOIOCB
92	{
93	/** Opaque pointer to data which is returned on an I/O event. */
94	void *pvUser;
95	#ifdef RT_ARCH_X86
96	uint32_t u32Padding0;
97	#endif
98	/** Contains the request number and is set by the kernel. */
99	uint32_t u32Key;
100	/** Reserved. */
101	uint32_t u32Reserved0;
102	/** The I/O opcode. */
103	uint16_t u16IoOpCode;
104	/** Request priority. */
105	int16_t i16Priority;
106	/** The file descriptor. */
107	uint32_t File;
108	/** The userspace pointer to the buffer containing/receiving the data. */
109	void *pvBuf;
110	#ifdef RT_ARCH_X86
111	uint32_t u32Padding1;
112	#endif
113	/** How many bytes to transfer. */
114	#ifdef RT_ARCH_X86
115	uint32_t cbTransfer;
116	uint32_t u32Padding2;
117	#elif defined(RT_ARCH_AMD64)
118	uint64_t cbTransfer;
119	#else
120	# error "Unknown architecture"
121	#endif
122	/** At which offset to start the transfer. */
123	int64_t off;
124	/** Reserved. */
125	uint64_t u64Reserved1;
126	/** Flags */
127	uint32_t fFlags;
128	/** Readyness signal file descriptor. */
129	uint32_t u32ResFd;
130	} LNXKAIOIOCB, *PLNXKAIOIOCB;
131
132	/**
133	* I/O event structure to notify about completed requests.
134	* Redefined here too because of the padding.
135	*/
136	typedef struct LNXKAIOIOEVENT
137	{
138	/** The pvUser field from the iocb. */
139	void *pvUser;
140	#ifdef RT_ARCH_X86
141	uint32_t u32Padding0;
142	#endif
143	/** The LNXKAIOIOCB object this event is for. */
144	PLNXKAIOIOCB *pIoCB;
145	#ifdef RT_ARCH_X86
146	uint32_t u32Padding1;
147	#endif
148	/** The result code of the operation .*/
149	#ifdef RT_ARCH_X86
150	int32_t rc;
151	uint32_t u32Padding2;
152	#elif defined(RT_ARCH_AMD64)
153	int64_t rc;
154	#else
155	# error "Unknown architecture"
156	#endif
157	/** Secondary result code. */
158	#ifdef RT_ARCH_X86
159	int32_t rc2;
160	uint32_t u32Padding3;
161	#elif defined(RT_ARCH_AMD64)
162	int64_t rc2;
163	#else
164	# error "Unknown architecture"
165	#endif
166	} LNXKAIOIOEVENT, *PLNXKAIOIOEVENT;
167
168
169	/**
170	* Async I/O completion context state.
171	*/
172	typedef struct RTFILEAIOCTXINTERNAL
173	{
174	/** Handle to the async I/O context. */
175	LNXKAIOCONTEXT AioContext;
176	/** Maximum number of requests this context can handle. */
177	int cRequestsMax;
178	/** Current number of requests active on this context. */
179	volatile int32_t cRequests;
180	/** The ID of the thread which is currently waiting for requests. */
181	volatile RTTHREAD hThreadWait;
182	/** Flag whether the thread was woken up. */
183	volatile bool fWokenUp;
184	/** Flag whether the thread is currently waiting in the syscall. */
185	volatile bool fWaiting;
186	/** Magic value (RTFILEAIOCTX_MAGIC). */
187	uint32_t u32Magic;
188	} RTFILEAIOCTXINTERNAL;
189	/** Pointer to an internal context structure. */
190	typedef RTFILEAIOCTXINTERNAL *PRTFILEAIOCTXINTERNAL;
191
192	/**
193	* Async I/O request state.
194	*/
195	typedef struct RTFILEAIOREQINTERNAL
196	{
197	/** The aio control block. This must be the FIRST elment in
198	* the structure! (see notes below) */
199	LNXKAIOIOCB AioCB;
200	/** Current state the request is in. */
201	RTFILEAIOREQSTATE enmState;
202	/** The I/O context this request is associated with. */
203	LNXKAIOCONTEXT AioContext;
204	/** Return code the request completed with. */
205	int Rc;
206	/** Number of bytes actually transferred. */
207	size_t cbTransfered;
208	/** Completion context we are assigned to. */
209	PRTFILEAIOCTXINTERNAL pCtxInt;
210	/** Magic value (RTFILEAIOREQ_MAGIC). */
211	uint32_t u32Magic;
212	} RTFILEAIOREQINTERNAL;
213	/** Pointer to an internal request structure. */
214	typedef RTFILEAIOREQINTERNAL *PRTFILEAIOREQINTERNAL;
215
216
217	/*******************************************************************************
218	* Defined Constants And Macros *
219	*******************************************************************************/
220	/** The max number of events to get in one call. */
221	#define AIO_MAXIMUM_REQUESTS_PER_CONTEXT 64
222
223
224	/**
225	* Creates a new async I/O context.
226	*/
227	DECLINLINE(int) rtFileAsyncIoLinuxCreate(unsigned cEvents, LNXKAIOCONTEXT *pAioContext)
228	{
229	int rc = syscall(__NR_io_setup, cEvents, pAioContext);
230	if (RT_UNLIKELY(rc == -1))
231	return RTErrConvertFromErrno(errno);
232
233	return VINF_SUCCESS;
234	}
235
236	/**
237	* Destroys a async I/O context.
238	*/
239	DECLINLINE(int) rtFileAsyncIoLinuxDestroy(LNXKAIOCONTEXT AioContext)
240	{
241	int rc = syscall(__NR_io_destroy, AioContext);
242	if (RT_UNLIKELY(rc == -1))
243	return RTErrConvertFromErrno(errno);
244
245	return VINF_SUCCESS;
246	}
247
248	/**
249	* Submits an array of I/O requests to the kernel.
250	*/
251	DECLINLINE(int) rtFileAsyncIoLinuxSubmit(LNXKAIOCONTEXT AioContext, long cReqs, LNXKAIOIOCB *ppIoCB, int pcSubmitted)
252	{
253	int rc = syscall(__NR_io_submit, AioContext, cReqs, ppIoCB);
254	if (RT_UNLIKELY(rc == -1))
255	return RTErrConvertFromErrno(errno);
256
257	*pcSubmitted = rc;
258
259	return VINF_SUCCESS;
260	}
261
262	/**
263	* Cancels a I/O request.
264	*/
265	DECLINLINE(int) rtFileAsyncIoLinuxCancel(LNXKAIOCONTEXT AioContext, PLNXKAIOIOCB pIoCB, PLNXKAIOIOEVENT pIoResult)
266	{
267	int rc = syscall(__NR_io_cancel, AioContext, pIoCB, pIoResult);
268	if (RT_UNLIKELY(rc == -1))
269	return RTErrConvertFromErrno(errno);
270
271	return VINF_SUCCESS;
272	}
273
274	/**
275	* Waits for I/O events.
276	* @returns Number of events (natural number w/ 0), IPRT error code (negative).
277	*/
278	DECLINLINE(int) rtFileAsyncIoLinuxGetEvents(LNXKAIOCONTEXT AioContext, long cReqsMin, long cReqs,
279	PLNXKAIOIOEVENT paIoResults, struct timespec *pTimeout)
280	{
281	int rc = syscall(__NR_io_getevents, AioContext, cReqsMin, cReqs, paIoResults, pTimeout);
282	if (RT_UNLIKELY(rc == -1))
283	return RTErrConvertFromErrno(errno);
284
285	return rc;
286	}
287
288	RTR3DECL(int) RTFileAioGetLimits(PRTFILEAIOLIMITS pAioLimits)
289	{
290	int rc = VINF_SUCCESS;
291	AssertPtrReturn(pAioLimits, VERR_INVALID_POINTER);
292
293	/*
294	* Check if the API is implemented by creating a
295	* completion port.
296	*/
297	LNXKAIOCONTEXT AioContext = 0;
298	rc = rtFileAsyncIoLinuxCreate(1, &AioContext);
299	if (RT_FAILURE(rc))
300	return rc;
301
302	rc = rtFileAsyncIoLinuxDestroy(AioContext);
303	if (RT_FAILURE(rc))
304	return rc;
305
306	/* Supported - fill in the limits. The alignment is the only restriction. */
307	pAioLimits->cReqsOutstandingMax = RTFILEAIO_UNLIMITED_REQS;
308	pAioLimits->cbBufferAlignment = 512;
309
310	return VINF_SUCCESS;
311	}
312
313
314	RTR3DECL(int) RTFileAioReqCreate(PRTFILEAIOREQ phReq)
315	{
316	AssertPtrReturn(phReq, VERR_INVALID_POINTER);
317
318	/*
319	* Allocate a new request and initialize it.
320	*/
321	PRTFILEAIOREQINTERNAL pReqInt = (PRTFILEAIOREQINTERNAL)RTMemAllocZ(sizeof(*pReqInt));
322	if (RT_UNLIKELY(!pReqInt))
323	return VERR_NO_MEMORY;
324
325	pReqInt->pCtxInt = NULL;
326	pReqInt->u32Magic = RTFILEAIOREQ_MAGIC;
327	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
328
329	*phReq = (RTFILEAIOREQ)pReqInt;
330	return VINF_SUCCESS;
331	}
332
333
334	RTDECL(int) RTFileAioReqDestroy(RTFILEAIOREQ hReq)
335	{
336	/*
337	* Validate the handle and ignore nil.
338	*/
339	if (hReq == NIL_RTFILEAIOREQ)
340	return VINF_SUCCESS;
341	PRTFILEAIOREQINTERNAL pReqInt = hReq;
342	RTFILEAIOREQ_VALID_RETURN(pReqInt);
343	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
344
345	/*
346	* Trash the magic and free it.
347	*/
348	ASMAtomicUoWriteU32(&pReqInt->u32Magic, ~RTFILEAIOREQ_MAGIC);
349	RTMemFree(pReqInt);
350	return VINF_SUCCESS;
351	}
352
353
354	/**
355	* Worker setting up the request.
356	*/
357	DECLINLINE(int) rtFileAioReqPrepareTransfer(RTFILEAIOREQ hReq, RTFILE hFile,
358	uint16_t uTransferDirection,
359	RTFOFF off, void *pvBuf, size_t cbTransfer,
360	void *pvUser)
361	{
362	/*
363	* Validate the input.
364	*/
365	PRTFILEAIOREQINTERNAL pReqInt = hReq;
366	RTFILEAIOREQ_VALID_RETURN(pReqInt);
367	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
368	Assert(hFile != NIL_RTFILE);
369
370	if (uTransferDirection != LNXKAIO_IOCB_CMD_FSYNC)
371	{
372	AssertPtr(pvBuf);
373	Assert(off >= 0);
374	Assert(cbTransfer > 0);
375	}
376
377	/*
378	* Setup the control block and clear the finished flag.
379	*/
380	pReqInt->AioCB.u16IoOpCode = uTransferDirection;
381	pReqInt->AioCB.File = (uint32_t)hFile;
382	pReqInt->AioCB.off = off;
383	pReqInt->AioCB.cbTransfer = cbTransfer;
384	pReqInt->AioCB.pvBuf = pvBuf;
385	pReqInt->AioCB.pvUser = pvUser;
386
387	pReqInt->pCtxInt = NULL;
388	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
389
390	return VINF_SUCCESS;
391	}
392
393
394	RTDECL(int) RTFileAioReqPrepareRead(RTFILEAIOREQ hReq, RTFILE hFile, RTFOFF off,
395	void pvBuf, size_t cbRead, void pvUser)
396	{
397	return rtFileAioReqPrepareTransfer(hReq, hFile, LNXKAIO_IOCB_CMD_READ,
398	off, pvBuf, cbRead, pvUser);
399	}
400
401
402	RTDECL(int) RTFileAioReqPrepareWrite(RTFILEAIOREQ hReq, RTFILE hFile, RTFOFF off,
403	void const pvBuf, size_t cbWrite, void pvUser)
404	{
405	return rtFileAioReqPrepareTransfer(hReq, hFile, LNXKAIO_IOCB_CMD_WRITE,
406	off, (void *)pvBuf, cbWrite, pvUser);
407	}
408
409
410	RTDECL(int) RTFileAioReqPrepareFlush(RTFILEAIOREQ hReq, RTFILE hFile, void *pvUser)
411	{
412	PRTFILEAIOREQINTERNAL pReqInt = hReq;
413	RTFILEAIOREQ_VALID_RETURN(pReqInt);
414	AssertReturn(hFile != NIL_RTFILE, VERR_INVALID_HANDLE);
415	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
416
417	return rtFileAioReqPrepareTransfer(pReqInt, hFile, LNXKAIO_IOCB_CMD_FSYNC,
418	0, NULL, 0, pvUser);
419	}
420
421
422	RTDECL(void *) RTFileAioReqGetUser(RTFILEAIOREQ hReq)
423	{
424	PRTFILEAIOREQINTERNAL pReqInt = hReq;
425	RTFILEAIOREQ_VALID_RETURN_RC(pReqInt, NULL);
426
427	return pReqInt->AioCB.pvUser;
428	}
429
430
431	RTDECL(int) RTFileAioReqCancel(RTFILEAIOREQ hReq)
432	{
433	PRTFILEAIOREQINTERNAL pReqInt = hReq;
434	RTFILEAIOREQ_VALID_RETURN(pReqInt);
435	RTFILEAIOREQ_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_NOT_SUBMITTED);
436
437	LNXKAIOIOEVENT AioEvent;
438	int rc = rtFileAsyncIoLinuxCancel(pReqInt->AioContext, &pReqInt->AioCB, &AioEvent);
439	if (RT_SUCCESS(rc))
440	{
441	/*
442	* Decrement request count because the request will never arrive at the
443	* completion port.
444	*/
445	AssertMsg(VALID_PTR(pReqInt->pCtxInt),
446	("Invalid state. Request was canceled but wasn't submitted\n"));
447
448	ASMAtomicDecS32(&pReqInt->pCtxInt->cRequests);
449	pReqInt->Rc = VERR_FILE_AIO_CANCELED;
450	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
451	return VINF_SUCCESS;
452	}
453	if (rc == VERR_TRY_AGAIN)
454	return VERR_FILE_AIO_IN_PROGRESS;
455	return rc;
456	}
457
458
459	RTDECL(int) RTFileAioReqGetRC(RTFILEAIOREQ hReq, size_t *pcbTransfered)
460	{
461	PRTFILEAIOREQINTERNAL pReqInt = hReq;
462	RTFILEAIOREQ_VALID_RETURN(pReqInt);
463	AssertPtrNull(pcbTransfered);
464	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
465	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, PREPARED, VERR_FILE_AIO_NOT_SUBMITTED);
466
467	if ( pcbTransfered
468	&& RT_SUCCESS(pReqInt->Rc))
469	*pcbTransfered = pReqInt->cbTransfered;
470
471	return pReqInt->Rc;
472	}
473
474
475	RTDECL(int) RTFileAioCtxCreate(PRTFILEAIOCTX phAioCtx, uint32_t cAioReqsMax)
476	{
477	PRTFILEAIOCTXINTERNAL pCtxInt;
478	AssertPtrReturn(phAioCtx, VERR_INVALID_POINTER);
479
480	/* The kernel interface needs a maximum. */
481	if (cAioReqsMax == RTFILEAIO_UNLIMITED_REQS)
482	return VERR_OUT_OF_RANGE;
483
484	pCtxInt = (PRTFILEAIOCTXINTERNAL)RTMemAllocZ(sizeof(RTFILEAIOCTXINTERNAL));
485	if (RT_UNLIKELY(!pCtxInt))
486	return VERR_NO_MEMORY;
487
488	/* Init the event handle. */
489	int rc = rtFileAsyncIoLinuxCreate(cAioReqsMax, &pCtxInt->AioContext);
490	if (RT_SUCCESS(rc))
491	{
492	pCtxInt->fWokenUp = false;
493	pCtxInt->fWaiting = false;
494	pCtxInt->hThreadWait = NIL_RTTHREAD;
495	pCtxInt->cRequestsMax = cAioReqsMax;
496	pCtxInt->u32Magic = RTFILEAIOCTX_MAGIC;
497	*phAioCtx = (RTFILEAIOCTX)pCtxInt;
498	}
499	else
500	RTMemFree(pCtxInt);
501
502	return rc;
503	}
504
505
506	RTDECL(int) RTFileAioCtxDestroy(RTFILEAIOCTX hAioCtx)
507	{
508	/* Validate the handle and ignore nil. */
509	if (hAioCtx == NIL_RTFILEAIOCTX)
510	return VINF_SUCCESS;
511	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
512	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
513
514	/* Cannot destroy a busy context. */
515	if (RT_UNLIKELY(pCtxInt->cRequests))
516	return VERR_FILE_AIO_BUSY;
517
518	/* The native bit first, then mark it as dead and free it. */
519	int rc = rtFileAsyncIoLinuxDestroy(pCtxInt->AioContext);
520	if (RT_FAILURE(rc))
521	return rc;
522	ASMAtomicUoWriteU32(&pCtxInt->u32Magic, RTFILEAIOCTX_MAGIC_DEAD);
523	RTMemFree(pCtxInt);
524
525	return VINF_SUCCESS;
526	}
527
528
529	RTDECL(uint32_t) RTFileAioCtxGetMaxReqCount(RTFILEAIOCTX hAioCtx)
530	{
531	/* Nil means global here. */
532	if (hAioCtx == NIL_RTFILEAIOCTX)
533	return RTFILEAIO_UNLIMITED_REQS; /** @todo r=bird: I'm a bit puzzled by this return value since it
534	* is completely useless in RTFileAioCtxCreate. */
535
536	/* Return 0 if the handle is invalid, it's better than garbage I think... */
537	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
538	RTFILEAIOCTX_VALID_RETURN_RC(pCtxInt, 0);
539
540	return pCtxInt->cRequestsMax;
541	}
542
543	RTDECL(int) RTFileAioCtxAssociateWithFile(RTFILEAIOCTX hAioCtx, RTFILE hFile)
544	{
545	/* Nothing to do. */
546	return VINF_SUCCESS;
547	}
548
549	RTDECL(int) RTFileAioCtxSubmit(RTFILEAIOCTX hAioCtx, PRTFILEAIOREQ pahReqs, size_t cReqs)
550	{
551	int rc = VINF_SUCCESS;
552
553	/*
554	* Parameter validation.
555	*/
556	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
557	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
558	AssertReturn(cReqs > 0, VERR_INVALID_PARAMETER);
559	AssertPtrReturn(pahReqs, VERR_INVALID_POINTER);
560	uint32_t i = cReqs;
561	PRTFILEAIOREQINTERNAL pReqInt = NULL;
562
563	/*
564	* Validate requests and associate with the context.
565	*/
566	while (i-- > 0)
567	{
568	pReqInt = pahReqs[i];
569	if (RTFILEAIOREQ_IS_NOT_VALID(pReqInt))
570	{
571	/* Undo everything and stop submitting. */
572	size_t iUndo = cReqs;
573	while (iUndo-- > i)
574	{
575	pReqInt = pahReqs[iUndo];
576	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
577	pReqInt->pCtxInt = NULL;
578	}
579	return VERR_INVALID_HANDLE;
580	}
581
582	pReqInt->AioContext = pCtxInt->AioContext;
583	pReqInt->pCtxInt = pCtxInt;
584	RTFILEAIOREQ_SET_STATE(pReqInt, SUBMITTED);
585	}
586
587	do
588	{
589	/*
590	* We cast pahReqs to the Linux iocb structure to avoid copying the requests
591	* into a temporary array. This is possible because the iocb structure is
592	* the first element in the request structure (see PRTFILEAIOCTXINTERNAL).
593	*/
594	int cReqsSubmitted = 0;
595	rc = rtFileAsyncIoLinuxSubmit(pCtxInt->AioContext, cReqs,
596	(PLNXKAIOIOCB *)pahReqs,
597	&cReqsSubmitted);
598	if (RT_FAILURE(rc))
599	{
600	/*
601	* We encountered an error.
602	* This means that the first IoCB
603	* is not correctly initialized
604	* (invalid buffer alignment or bad file descriptor).
605	* Revert every request into the prepared state except
606	* the first one which will switch to completed.
607	* Another reason could be insufficient resources.
608	*/
609	i = cReqs;
610	while (i-- > 0)
611	{
612	/* Already validated. */
613	pReqInt = pahReqs[i];
614	pReqInt->pCtxInt = NULL;
615	pReqInt->AioContext = 0;
616	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
617	}
618
619	if (rc == VERR_TRY_AGAIN)
620	return VERR_FILE_AIO_INSUFFICIENT_RESSOURCES;
621	else
622	{
623	/* The first request failed. */
624	pReqInt = pahReqs[0];
625	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
626	pReqInt->Rc = rc;
627	pReqInt->cbTransfered = 0;
628	return rc;
629	}
630	}
631
632	/* Advance. */
633	cReqs -= cReqsSubmitted;
634	pahReqs += cReqsSubmitted;
635	ASMAtomicAddS32(&pCtxInt->cRequests, cReqsSubmitted);
636
637	} while (cReqs);
638
639	return rc;
640	}
641
642
643	RTDECL(int) RTFileAioCtxWait(RTFILEAIOCTX hAioCtx, size_t cMinReqs, RTMSINTERVAL cMillies,
644	PRTFILEAIOREQ pahReqs, size_t cReqs, uint32_t *pcReqs)
645	{
646	/*
647	* Validate the parameters, making sure to always set pcReqs.
648	*/
649	AssertPtrReturn(pcReqs, VERR_INVALID_POINTER);
650	pcReqs = 0; / always set */
651	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
652	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
653	AssertPtrReturn(pahReqs, VERR_INVALID_POINTER);
654	AssertReturn(cReqs != 0, VERR_INVALID_PARAMETER);
655	AssertReturn(cReqs >= cMinReqs, VERR_OUT_OF_RANGE);
656
657	/*
658	* Can't wait if there are not requests around.
659	*/
660	if (RT_UNLIKELY(ASMAtomicUoReadS32(&pCtxInt->cRequests) == 0))
661	return VERR_FILE_AIO_NO_REQUEST;
662
663	/*
664	* Convert the timeout if specified.
665	*/
666	struct timespec *pTimeout = NULL;
667	struct timespec Timeout = {0,0};
668	uint64_t StartNanoTS = 0;
669	if (cMillies != RT_INDEFINITE_WAIT)
670	{
671	Timeout.tv_sec = cMillies / 1000;
672	Timeout.tv_nsec = cMillies % 1000 * 1000000;
673	pTimeout = &Timeout;
674	StartNanoTS = RTTimeNanoTS();
675	}
676
677	/* Wait for at least one. */
678	if (!cMinReqs)
679	cMinReqs = 1;
680
681	/* For the wakeup call. */
682	Assert(pCtxInt->hThreadWait == NIL_RTTHREAD);
683	ASMAtomicWriteHandle(&pCtxInt->hThreadWait, RTThreadSelf());
684
685	/*
686	* Loop until we're woken up, hit an error (incl timeout), or
687	* have collected the desired number of requests.
688	*/
689	int rc = VINF_SUCCESS;
690	int cRequestsCompleted = 0;
691	while (!pCtxInt->fWokenUp)
692	{
693	LNXKAIOIOEVENT aPortEvents[AIO_MAXIMUM_REQUESTS_PER_CONTEXT];
694	int cRequestsToWait = RT_MIN(cReqs, AIO_MAXIMUM_REQUESTS_PER_CONTEXT);
695	ASMAtomicXchgBool(&pCtxInt->fWaiting, true);
696	rc = rtFileAsyncIoLinuxGetEvents(pCtxInt->AioContext, cMinReqs, cRequestsToWait, &aPortEvents[0], pTimeout);
697	ASMAtomicXchgBool(&pCtxInt->fWaiting, false);
698	if (RT_FAILURE(rc))
699	break;
700	uint32_t const cDone = rc;
701	rc = VINF_SUCCESS;
702
703	/*
704	* Process received events / requests.
705	*/
706	for (uint32_t i = 0; i < cDone; i++)
707	{
708	/*
709	* The iocb is the first element in our request structure.
710	* So we can safely cast it directly to the handle (see above)
711	*/
712	PRTFILEAIOREQINTERNAL pReqInt = (PRTFILEAIOREQINTERNAL)aPortEvents[i].pIoCB;
713	AssertPtr(pReqInt);
714	Assert(pReqInt->u32Magic == RTFILEAIOREQ_MAGIC);
715
716	/** @todo aeichner: The rc field contains the result code
717	* like you can find in errno for the normal read/write ops.
718	* But there is a second field called rc2. I don't know the
719	* purpose for it yet.
720	*/
721	if (RT_UNLIKELY(aPortEvents[i].rc < 0))
722	pReqInt->Rc = RTErrConvertFromErrno(-aPortEvents[i].rc); /* Convert to positive value. */
723	else
724	{
725	pReqInt->Rc = VINF_SUCCESS;
726	pReqInt->cbTransfered = aPortEvents[i].rc;
727	}
728
729	/* Mark the request as finished. */
730	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
731
732	pahReqs[cRequestsCompleted++] = (RTFILEAIOREQ)pReqInt;
733	}
734
735	/*
736	* Done Yet? If not advance and try again.
737	*/
738	if (cDone >= cMinReqs)
739	break;
740	cMinReqs -= cDone;
741	cReqs -= cDone;
742
743	if (cMillies != RT_INDEFINITE_WAIT)
744	{
745	/* The API doesn't return ETIMEDOUT, so we have to fix that ourselves. */
746	uint64_t NanoTS = RTTimeNanoTS();
747	uint64_t cMilliesElapsed = (NanoTS - StartNanoTS) / 1000000;
748	if (cMilliesElapsed >= cMillies)
749	{
750	rc = VERR_TIMEOUT;
751	break;
752	}
753
754	/* The syscall supposedly updates it, but we're paranoid. :-) */
755	Timeout.tv_sec = (cMillies - (RTMSINTERVAL)cMilliesElapsed) / 1000;
756	Timeout.tv_nsec = (cMillies - (RTMSINTERVAL)cMilliesElapsed) % 1000 * 1000000;
757	}
758	}
759
760	/*
761	* Update the context state and set the return value.
762	*/
763	*pcReqs = cRequestsCompleted;
764	ASMAtomicSubS32(&pCtxInt->cRequests, cRequestsCompleted);
765	Assert(pCtxInt->hThreadWait == RTThreadSelf());
766	ASMAtomicWriteHandle(&pCtxInt->hThreadWait, NIL_RTTHREAD);
767
768	/*
769	* Clear the wakeup flag and set rc.
770	*/
771	if ( pCtxInt->fWokenUp
772	&& RT_SUCCESS(rc))
773	{
774	ASMAtomicXchgBool(&pCtxInt->fWokenUp, false);
775	rc = VERR_INTERRUPTED;
776	}
777
778	return rc;
779	}
780
781
782	RTDECL(int) RTFileAioCtxWakeup(RTFILEAIOCTX hAioCtx)
783	{
784	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
785	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
786
787	/** @todo r=bird: Define the protocol for how to resume work after calling
788	* this function. */
789
790	bool fWokenUp = ASMAtomicXchgBool(&pCtxInt->fWokenUp, true);
791
792	/*
793	* Read the thread handle before the status flag.
794	* If we read the handle after the flag we might
795	* end up with an invalid handle because the thread
796	* waiting in RTFileAioCtxWakeup() might get scheduled
797	* before we read the flag and returns.
798	* We can ensure that the handle is valid if fWaiting is true
799	* when reading the handle before the status flag.
800	*/
801	RTTHREAD hThread;
802	ASMAtomicReadHandle(&pCtxInt->hThreadWait, &hThread);
803	bool fWaiting = ASMAtomicReadBool(&pCtxInt->fWaiting);
804	if ( !fWokenUp
805	&& fWaiting)
806	{
807	/*
808	* If a thread waits the handle must be valid.
809	* It is possible that the thread returns from
810	* rtFileAsyncIoLinuxGetEvents() before the signal
811	* is send.
812	* This is no problem because we already set fWokenUp
813	* to true which will let the thread return VERR_INTERRUPTED
814	* and the next call to RTFileAioCtxWait() will not
815	* return VERR_INTERRUPTED because signals are not saved
816	* and will simply vanish if the destination thread can't
817	* receive it.
818	*/
819	Assert(hThread != NIL_RTTHREAD);
820	RTThreadPoke(hThread);
821	}
822
823	return VINF_SUCCESS;
824	}
825

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

source: vbox/trunk/src/VBox/Runtime/r3/linux/fileaio-linux.cpp@ 34244

Download in other formats: