fileaio-linux.cpp@ 26618

Last change on this file since 26618 was 25724, checked in by vboxsync, 15 years ago
iprt: Use RTMSINTERVAL for timeouts. Fixed missing timeout underflow checks in two RTFileAioCtxWait implementations.
Property svn:eol-style set to `native` Property svn:keywords set to `Author Date Id Revision`
File size: 26.1 KB

Line
1	/* $Id: fileaio-linux.cpp 25724 2010-01-11 14:45:34Z vboxsync $ */
2	/** @file
3	* IPRT - File async I/O, native implementation for the Linux host platform.
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	/** @page pg_rtfileaio_linux RTFile Async I/O - Linux Implementation Notes
32	* @internal
33	*
34	* Linux implements the kernel async I/O API through the io_* syscalls. They are
35	* not exposed in the glibc (the aio_* API uses userspace threads and blocking
36	* I/O operations to simulate async behavior). There is an external library
37	* called libaio which implements these syscalls but because we don't want to
38	* have another dependency and this library is not installed by default and the
39	* interface is really simple we use the kernel interface directly using wrapper
40	* functions.
41	*
42	* The interface has some limitations. The first one is that the file must be
43	* opened with O_DIRECT. This disables caching done by the kernel which can be
44	* compensated if the user of this API implements caching itself. The next
45	* limitation is that data buffers must be aligned at a 512 byte boundary or the
46	* request will fail.
47	*/
48	/** @todo r=bird: What's this about "must be opened with O_DIRECT"? An
49	* explanation would be nice, esp. seeing what Linus is quoted saying
50	* about it in the open man page... */
51
52	/*******************************************************************************
53	* Header Files *
54	*******************************************************************************/
55	#define LOG_GROUP RTLOGGROUP_FILE
56	#include <iprt/asm.h>
57	#include <iprt/mem.h>
58	#include <iprt/assert.h>
59	#include <iprt/string.h>
60	#include <iprt/err.h>
61	#include <iprt/log.h>
62	#include <iprt/thread.h>
63	#include "internal/fileaio.h"
64
65	#include <unistd.h>
66	#include <sys/syscall.h>
67	#include <errno.h>
68
69	#include <iprt/file.h>
70
71
72	/*******************************************************************************
73	* Structures and Typedefs *
74	*******************************************************************************/
75	/** The async I/O context handle */
76	typedef unsigned long LNXKAIOCONTEXT;
77
78	/**
79	* Supported commands for the iocbs
80	*/
81	enum
82	{
83	LNXKAIO_IOCB_CMD_READ = 0,
84	LNXKAIO_IOCB_CMD_WRITE
85	};
86
87	/**
88	* The iocb structure of a request which is passed to the kernel.
89	*
90	* We redefined this here because the version in the header lacks padding
91	* for 32bit.
92	*/
93	typedef struct LNXKAIOIOCB
94	{
95	/** Opaque pointer to data which is returned on an I/O event. */
96	void *pvUser;
97	#ifdef RT_ARCH_X86
98	uint32_t u32Padding0;
99	#endif
100	/** Contains the request number and is set by the kernel. */
101	uint32_t u32Key;
102	/** Reserved. */
103	uint32_t u32Reserved0;
104	/** The I/O opcode. */
105	uint16_t u16IoOpCode;
106	/** Request priority. */
107	int16_t i16Priority;
108	/** The file descriptor. */
109	uint32_t File;
110	/** The userspace pointer to the buffer containing/receiving the data. */
111	void *pvBuf;
112	#ifdef RT_ARCH_X86
113	uint32_t u32Padding1;
114	#endif
115	/** How many bytes to transfer. */
116	#ifdef RT_ARCH_X86
117	uint32_t cbTransfer;
118	uint32_t u32Padding2;
119	#elif defined(RT_ARCH_AMD64)
120	uint64_t cbTransfer;
121	#else
122	# error "Unknown architecture"
123	#endif
124	/** At which offset to start the transfer. */
125	int64_t off;
126	/** Reserved. */
127	uint64_t u64Reserved1;
128	/** Flags */
129	uint32_t fFlags;
130	/** Readyness signal file descriptor. */
131	uint32_t u32ResFd;
132	} LNXKAIOIOCB, *PLNXKAIOIOCB;
133
134	/**
135	* I/O event structure to notify about completed requests.
136	* Redefined here too because of the padding.
137	*/
138	typedef struct LNXKAIOIOEVENT
139	{
140	/** The pvUser field from the iocb. */
141	void *pvUser;
142	#ifdef RT_ARCH_X86
143	uint32_t u32Padding0;
144	#endif
145	/** The LNXKAIOIOCB object this event is for. */
146	PLNXKAIOIOCB *pIoCB;
147	#ifdef RT_ARCH_X86
148	uint32_t u32Padding1;
149	#endif
150	/** The result code of the operation .*/
151	#ifdef RT_ARCH_X86
152	int32_t rc;
153	uint32_t u32Padding2;
154	#elif defined(RT_ARCH_AMD64)
155	int64_t rc;
156	#else
157	# error "Unknown architecture"
158	#endif
159	/** Secondary result code. */
160	#ifdef RT_ARCH_X86
161	int32_t rc2;
162	uint32_t u32Padding3;
163	#elif defined(RT_ARCH_AMD64)
164	int64_t rc2;
165	#else
166	# error "Unknown architecture"
167	#endif
168	} LNXKAIOIOEVENT, *PLNXKAIOIOEVENT;
169
170
171	/**
172	* Async I/O completion context state.
173	*/
174	typedef struct RTFILEAIOCTXINTERNAL
175	{
176	/** Handle to the async I/O context. */
177	LNXKAIOCONTEXT AioContext;
178	/** Maximum number of requests this context can handle. */
179	int cRequestsMax;
180	/** Current number of requests active on this context. */
181	volatile int32_t cRequests;
182	/** The ID of the thread which is currently waiting for requests. */
183	volatile RTTHREAD hThreadWait;
184	/** Flag whether the thread was woken up. */
185	volatile bool fWokenUp;
186	/** Flag whether the thread is currently waiting in the syscall. */
187	volatile bool fWaiting;
188	/** Magic value (RTFILEAIOCTX_MAGIC). */
189	uint32_t u32Magic;
190	} RTFILEAIOCTXINTERNAL;
191	/** Pointer to an internal context structure. */
192	typedef RTFILEAIOCTXINTERNAL *PRTFILEAIOCTXINTERNAL;
193
194	/**
195	* Async I/O request state.
196	*/
197	typedef struct RTFILEAIOREQINTERNAL
198	{
199	/** The aio control block. This must be the FIRST elment in
200	* the structure! (see notes below) */
201	LNXKAIOIOCB AioCB;
202	/** Current state the request is in. */
203	RTFILEAIOREQSTATE enmState;
204	/** The I/O context this request is associated with. */
205	LNXKAIOCONTEXT AioContext;
206	/** Return code the request completed with. */
207	int Rc;
208	/** Number of bytes actually trasnfered. */
209	size_t cbTransfered;
210	/** Completion context we are assigned to. */
211	PRTFILEAIOCTXINTERNAL pCtxInt;
212	/** Magic value (RTFILEAIOREQ_MAGIC). */
213	uint32_t u32Magic;
214	} RTFILEAIOREQINTERNAL;
215	/** Pointer to an internal request structure. */
216	typedef RTFILEAIOREQINTERNAL *PRTFILEAIOREQINTERNAL;
217
218
219	/*******************************************************************************
220	* Defined Constants And Macros *
221	*******************************************************************************/
222	/** The max number of events to get in one call. */
223	#define AIO_MAXIMUM_REQUESTS_PER_CONTEXT 64
224
225
226	/**
227	* Creates a new async I/O context.
228	*/
229	DECLINLINE(int) rtFileAsyncIoLinuxCreate(unsigned cEvents, LNXKAIOCONTEXT *pAioContext)
230	{
231	int rc = syscall(__NR_io_setup, cEvents, pAioContext);
232	if (RT_UNLIKELY(rc == -1))
233	return RTErrConvertFromErrno(errno);
234
235	return VINF_SUCCESS;
236	}
237
238	/**
239	* Destroys a async I/O context.
240	*/
241	DECLINLINE(int) rtFileAsyncIoLinuxDestroy(LNXKAIOCONTEXT AioContext)
242	{
243	int rc = syscall(__NR_io_destroy, AioContext);
244	if (RT_UNLIKELY(rc == -1))
245	return RTErrConvertFromErrno(errno);
246
247	return VINF_SUCCESS;
248	}
249
250	/**
251	* Submits an array of I/O requests to the kernel.
252	*/
253	DECLINLINE(int) rtFileAsyncIoLinuxSubmit(LNXKAIOCONTEXT AioContext, long cReqs, LNXKAIOIOCB *ppIoCB, int pcSubmitted)
254	{
255	int rc = syscall(__NR_io_submit, AioContext, cReqs, ppIoCB);
256	if (RT_UNLIKELY(rc == -1))
257	return RTErrConvertFromErrno(errno);
258
259	*pcSubmitted = rc;
260
261	return VINF_SUCCESS;
262	}
263
264	/**
265	* Cancels a I/O request.
266	*/
267	DECLINLINE(int) rtFileAsyncIoLinuxCancel(LNXKAIOCONTEXT AioContext, PLNXKAIOIOCB pIoCB, PLNXKAIOIOEVENT pIoResult)
268	{
269	int rc = syscall(__NR_io_cancel, AioContext, pIoCB, pIoResult);
270	if (RT_UNLIKELY(rc == -1))
271	return RTErrConvertFromErrno(errno);
272
273	return VINF_SUCCESS;
274	}
275
276	/**
277	* Waits for I/O events.
278	* @returns Number of events (natural number w/ 0), IPRT error code (negative).
279	*/
280	DECLINLINE(int) rtFileAsyncIoLinuxGetEvents(LNXKAIOCONTEXT AioContext, long cReqsMin, long cReqs,
281	PLNXKAIOIOEVENT paIoResults, struct timespec *pTimeout)
282	{
283	int rc = syscall(__NR_io_getevents, AioContext, cReqsMin, cReqs, paIoResults, pTimeout);
284	if (RT_UNLIKELY(rc == -1))
285	return RTErrConvertFromErrno(errno);
286
287	return rc;
288	}
289
290	RTR3DECL(int) RTFileAioGetLimits(PRTFILEAIOLIMITS pAioLimits)
291	{
292	int rc = VINF_SUCCESS;
293	AssertPtrReturn(pAioLimits, VERR_INVALID_POINTER);
294
295	/*
296	* Check if the API is implemented by creating a
297	* completion port.
298	*/
299	LNXKAIOCONTEXT AioContext = 0;
300	rc = rtFileAsyncIoLinuxCreate(1, &AioContext);
301	if (RT_FAILURE(rc))
302	return rc;
303
304	rc = rtFileAsyncIoLinuxDestroy(AioContext);
305	if (RT_FAILURE(rc))
306	return rc;
307
308	/* Supported - fill in the limits. The alignment is the only restriction. */
309	pAioLimits->cReqsOutstandingMax = RTFILEAIO_UNLIMITED_REQS;
310	pAioLimits->cbBufferAlignment = 512;
311
312	return VINF_SUCCESS;
313	}
314
315
316	RTR3DECL(int) RTFileAioReqCreate(PRTFILEAIOREQ phReq)
317	{
318	AssertPtrReturn(phReq, VERR_INVALID_POINTER);
319
320	/*
321	* Allocate a new request and initialize it.
322	*/
323	PRTFILEAIOREQINTERNAL pReqInt = (PRTFILEAIOREQINTERNAL)RTMemAllocZ(sizeof(*pReqInt));
324	if (RT_UNLIKELY(!pReqInt))
325	return VERR_NO_MEMORY;
326
327	pReqInt->pCtxInt = NULL;
328	pReqInt->u32Magic = RTFILEAIOREQ_MAGIC;
329	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
330
331	*phReq = (RTFILEAIOREQ)pReqInt;
332	return VINF_SUCCESS;
333	}
334
335
336	RTDECL(int) RTFileAioReqDestroy(RTFILEAIOREQ hReq)
337	{
338	/*
339	* Validate the handle and ignore nil.
340	*/
341	if (hReq == NIL_RTFILEAIOREQ)
342	return VINF_SUCCESS;
343	PRTFILEAIOREQINTERNAL pReqInt = hReq;
344	RTFILEAIOREQ_VALID_RETURN(pReqInt);
345	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
346
347	/*
348	* Trash the magic and free it.
349	*/
350	ASMAtomicUoWriteU32(&pReqInt->u32Magic, ~RTFILEAIOREQ_MAGIC);
351	RTMemFree(pReqInt);
352	return VINF_SUCCESS;
353	}
354
355
356	/**
357	* Worker setting up the request.
358	*/
359	DECLINLINE(int) rtFileAioReqPrepareTransfer(RTFILEAIOREQ hReq, RTFILE hFile,
360	uint16_t uTransferDirection,
361	RTFOFF off, void *pvBuf, size_t cbTransfer,
362	void *pvUser)
363	{
364	/*
365	* Validate the input.
366	*/
367	PRTFILEAIOREQINTERNAL pReqInt = hReq;
368	RTFILEAIOREQ_VALID_RETURN(pReqInt);
369	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
370	Assert(hFile != NIL_RTFILE);
371	AssertPtr(pvBuf);
372	Assert(off >= 0);
373	Assert(cbTransfer > 0);
374
375	/*
376	* Setup the control block and clear the finished flag.
377	*/
378	pReqInt->AioCB.u16IoOpCode = uTransferDirection;
379	pReqInt->AioCB.File = (uint32_t)hFile;
380	pReqInt->AioCB.off = off;
381	pReqInt->AioCB.cbTransfer = cbTransfer;
382	pReqInt->AioCB.pvBuf = pvBuf;
383	pReqInt->AioCB.pvUser = pvUser;
384
385	pReqInt->pCtxInt = NULL;
386	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
387
388	return VINF_SUCCESS;
389	}
390
391
392	RTDECL(int) RTFileAioReqPrepareRead(RTFILEAIOREQ hReq, RTFILE hFile, RTFOFF off,
393	void pvBuf, size_t cbRead, void pvUser)
394	{
395	return rtFileAioReqPrepareTransfer(hReq, hFile, LNXKAIO_IOCB_CMD_READ,
396	off, pvBuf, cbRead, pvUser);
397	}
398
399
400	RTDECL(int) RTFileAioReqPrepareWrite(RTFILEAIOREQ hReq, RTFILE hFile, RTFOFF off,
401	void const pvBuf, size_t cbWrite, void pvUser)
402	{
403	return rtFileAioReqPrepareTransfer(hReq, hFile, LNXKAIO_IOCB_CMD_WRITE,
404	off, (void *)pvBuf, cbWrite, pvUser);
405	}
406
407
408	RTDECL(int) RTFileAioReqPrepareFlush(RTFILEAIOREQ hReq, RTFILE hFile, void *pvUser)
409	{
410	PRTFILEAIOREQINTERNAL pReqInt = hReq;
411	RTFILEAIOREQ_VALID_RETURN(pReqInt);
412	AssertReturn(hFile != NIL_RTFILE, VERR_INVALID_HANDLE);
413	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
414
415	/** @todo: Flushing is not neccessary on Linux because O_DIRECT is mandatory
416	* which disables caching.
417	* We could setup a fake request which isn't really executed
418	* to avoid platform dependent code in the caller.
419	*/
420	#if 0
421	return rtFileAsyncPrepareTransfer(pRequest, File, TRANSFERDIRECTION_FLUSH,
422	0, NULL, 0, pvUser);
423	#endif
424	return VERR_NOT_IMPLEMENTED;
425	}
426
427
428	RTDECL(void *) RTFileAioReqGetUser(RTFILEAIOREQ hReq)
429	{
430	PRTFILEAIOREQINTERNAL pReqInt = hReq;
431	RTFILEAIOREQ_VALID_RETURN_RC(pReqInt, NULL);
432
433	return pReqInt->AioCB.pvUser;
434	}
435
436
437	RTDECL(int) RTFileAioReqCancel(RTFILEAIOREQ hReq)
438	{
439	PRTFILEAIOREQINTERNAL pReqInt = hReq;
440	RTFILEAIOREQ_VALID_RETURN(pReqInt);
441	RTFILEAIOREQ_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_NOT_SUBMITTED);
442
443	LNXKAIOIOEVENT AioEvent;
444	int rc = rtFileAsyncIoLinuxCancel(pReqInt->AioContext, &pReqInt->AioCB, &AioEvent);
445	if (RT_SUCCESS(rc))
446	{
447	/*
448	* Decrement request count because the request will never arrive at the
449	* completion port.
450	*/
451	AssertMsg(VALID_PTR(pReqInt->pCtxInt),
452	("Invalid state. Request was canceled but wasn't submitted\n"));
453
454	ASMAtomicDecS32(&pReqInt->pCtxInt->cRequests);
455	pReqInt->Rc = VERR_FILE_AIO_CANCELED;
456	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
457	return VINF_SUCCESS;
458	}
459	if (rc == VERR_TRY_AGAIN)
460	return VERR_FILE_AIO_IN_PROGRESS;
461	return rc;
462	}
463
464
465	RTDECL(int) RTFileAioReqGetRC(RTFILEAIOREQ hReq, size_t *pcbTransfered)
466	{
467	PRTFILEAIOREQINTERNAL pReqInt = hReq;
468	RTFILEAIOREQ_VALID_RETURN(pReqInt);
469	AssertPtrNull(pcbTransfered);
470	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, SUBMITTED, VERR_FILE_AIO_IN_PROGRESS);
471	RTFILEAIOREQ_NOT_STATE_RETURN_RC(pReqInt, PREPARED, VERR_FILE_AIO_NOT_SUBMITTED);
472
473	if ( pcbTransfered
474	&& RT_SUCCESS(pReqInt->Rc))
475	*pcbTransfered = pReqInt->cbTransfered;
476
477	return pReqInt->Rc;
478	}
479
480
481	RTDECL(int) RTFileAioCtxCreate(PRTFILEAIOCTX phAioCtx, uint32_t cAioReqsMax)
482	{
483	PRTFILEAIOCTXINTERNAL pCtxInt;
484	AssertPtrReturn(phAioCtx, VERR_INVALID_POINTER);
485
486	/* The kernel interface needs a maximum. */
487	if (cAioReqsMax == RTFILEAIO_UNLIMITED_REQS)
488	return VERR_OUT_OF_RANGE;
489
490	pCtxInt = (PRTFILEAIOCTXINTERNAL)RTMemAllocZ(sizeof(RTFILEAIOCTXINTERNAL));
491	if (RT_UNLIKELY(!pCtxInt))
492	return VERR_NO_MEMORY;
493
494	/* Init the event handle. */
495	int rc = rtFileAsyncIoLinuxCreate(cAioReqsMax, &pCtxInt->AioContext);
496	if (RT_SUCCESS(rc))
497	{
498	pCtxInt->fWokenUp = false;
499	pCtxInt->fWaiting = false;
500	pCtxInt->hThreadWait = NIL_RTTHREAD;
501	pCtxInt->cRequestsMax = cAioReqsMax;
502	pCtxInt->u32Magic = RTFILEAIOCTX_MAGIC;
503	*phAioCtx = (RTFILEAIOCTX)pCtxInt;
504	}
505	else
506	RTMemFree(pCtxInt);
507
508	return rc;
509	}
510
511
512	RTDECL(int) RTFileAioCtxDestroy(RTFILEAIOCTX hAioCtx)
513	{
514	/* Validate the handle and ignore nil. */
515	if (hAioCtx == NIL_RTFILEAIOCTX)
516	return VINF_SUCCESS;
517	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
518	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
519
520	/* Cannot destroy a busy context. */
521	if (RT_UNLIKELY(pCtxInt->cRequests))
522	return VERR_FILE_AIO_BUSY;
523
524	/* The native bit first, then mark it as dead and free it. */
525	int rc = rtFileAsyncIoLinuxDestroy(pCtxInt->AioContext);
526	if (RT_FAILURE(rc))
527	return rc;
528	ASMAtomicUoWriteU32(&pCtxInt->u32Magic, RTFILEAIOCTX_MAGIC_DEAD);
529	RTMemFree(pCtxInt);
530
531	return VINF_SUCCESS;
532	}
533
534
535	RTDECL(uint32_t) RTFileAioCtxGetMaxReqCount(RTFILEAIOCTX hAioCtx)
536	{
537	/* Nil means global here. */
538	if (hAioCtx == NIL_RTFILEAIOCTX)
539	return RTFILEAIO_UNLIMITED_REQS; /** @todo r=bird: I'm a bit puzzled by this return value since it
540	* is completely useless in RTFileAioCtxCreate. */
541
542	/* Return 0 if the handle is invalid, it's better than garbage I think... */
543	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
544	RTFILEAIOCTX_VALID_RETURN_RC(pCtxInt, 0);
545
546	return pCtxInt->cRequestsMax;
547	}
548
549	RTDECL(int) RTFileAioCtxAssociateWithFile(RTFILEAIOCTX hAioCtx, RTFILE hFile)
550	{
551	/* Nothing to do. */
552	return VINF_SUCCESS;
553	}
554
555	RTDECL(int) RTFileAioCtxSubmit(RTFILEAIOCTX hAioCtx, PRTFILEAIOREQ pahReqs, size_t cReqs)
556	{
557	int rc = VINF_SUCCESS;
558
559	/*
560	* Parameter validation.
561	*/
562	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
563	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
564	AssertReturn(cReqs > 0, VERR_INVALID_PARAMETER);
565	AssertPtrReturn(pahReqs, VERR_INVALID_POINTER);
566	uint32_t i = cReqs;
567	PRTFILEAIOREQINTERNAL pReqInt = NULL;
568
569	/*
570	* Vaildate requests and associate with the context.
571	*/
572	while (i-- > 0)
573	{
574	pReqInt = pahReqs[i];
575	if (RTFILEAIOREQ_IS_NOT_VALID(pReqInt))
576	{
577	/* Undo everything and stop submitting. */
578	size_t iUndo = cReqs;
579	while (iUndo-- > i)
580	{
581	pReqInt = pahReqs[iUndo];
582	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
583	pReqInt->pCtxInt = NULL;
584	}
585	return VERR_INVALID_HANDLE;
586	}
587
588	pReqInt->AioContext = pCtxInt->AioContext;
589	pReqInt->pCtxInt = pCtxInt;
590	RTFILEAIOREQ_SET_STATE(pReqInt, SUBMITTED);
591	}
592
593	do
594	{
595	/*
596	* We cast pahReqs to the Linux iocb structure to avoid copying the requests
597	* into a temporary array. This is possible because the iocb structure is
598	* the first element in the request structure (see PRTFILEAIOCTXINTERNAL).
599	*/
600	int cReqsSubmitted = 0;
601	rc = rtFileAsyncIoLinuxSubmit(pCtxInt->AioContext, cReqs,
602	(PLNXKAIOIOCB *)pahReqs,
603	&cReqsSubmitted);
604	if (RT_FAILURE(rc))
605	{
606	/*
607	* We encountered an error.
608	* This means that the first IoCB
609	* is not correctly initialized
610	* (invalid buffer alignment or bad file descriptor).
611	* Revert every request into the prepared state except
612	* the first one which will switch to completed.
613	* Another reason could be insuffidient ressources.
614	*/
615	i = cReqs;
616	while (i-- > 0)
617	{
618	/* Already validated. */
619	pReqInt = pahReqs[i];
620	pReqInt->pCtxInt = NULL;
621	pReqInt->AioContext = 0;
622	RTFILEAIOREQ_SET_STATE(pReqInt, PREPARED);
623	}
624
625	if (rc == VERR_TRY_AGAIN)
626	return VERR_FILE_AIO_INSUFFICIENT_RESSOURCES;
627	else
628	{
629	/* The first request failed. */
630	pReqInt = pahReqs[0];
631	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
632	pReqInt->Rc = rc;
633	pReqInt->cbTransfered = 0;
634	return rc;
635	}
636	}
637
638	/* Advance. */
639	cReqs -= cReqsSubmitted;
640	pahReqs += cReqsSubmitted;
641	ASMAtomicAddS32(&pCtxInt->cRequests, cReqsSubmitted);
642
643	} while (cReqs);
644
645	return rc;
646	}
647
648
649	RTDECL(int) RTFileAioCtxWait(RTFILEAIOCTX hAioCtx, size_t cMinReqs, RTMSINTERVAL cMillies,
650	PRTFILEAIOREQ pahReqs, size_t cReqs, uint32_t *pcReqs)
651	{
652	/*
653	* Validate the parameters, making sure to always set pcReqs.
654	*/
655	AssertPtrReturn(pcReqs, VERR_INVALID_POINTER);
656	pcReqs = 0; / always set */
657	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
658	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
659	AssertPtrReturn(pahReqs, VERR_INVALID_POINTER);
660	AssertReturn(cReqs != 0, VERR_INVALID_PARAMETER);
661	AssertReturn(cReqs >= cMinReqs, VERR_OUT_OF_RANGE);
662
663	/*
664	* Can't wait if there are not requests around.
665	*/
666	if (RT_UNLIKELY(ASMAtomicUoReadS32(&pCtxInt->cRequests) == 0))
667	return VERR_FILE_AIO_NO_REQUEST;
668
669	/*
670	* Convert the timeout if specified.
671	*/
672	struct timespec *pTimeout = NULL;
673	struct timespec Timeout = {0,0};
674	uint64_t StartNanoTS = 0;
675	if (cMillies != RT_INDEFINITE_WAIT)
676	{
677	Timeout.tv_sec = cMillies / 1000;
678	Timeout.tv_nsec = cMillies % 1000 * 1000000;
679	pTimeout = &Timeout;
680	StartNanoTS = RTTimeNanoTS();
681	}
682
683	/* Wait for at least one. */
684	if (!cMinReqs)
685	cMinReqs = 1;
686
687	/* For the wakeup call. */
688	Assert(pCtxInt->hThreadWait == NIL_RTTHREAD);
689	ASMAtomicWriteHandle(&pCtxInt->hThreadWait, RTThreadSelf());
690
691	/*
692	* Loop until we're woken up, hit an error (incl timeout), or
693	* have collected the desired number of requests.
694	*/
695	int rc = VINF_SUCCESS;
696	int cRequestsCompleted = 0;
697	while (!pCtxInt->fWokenUp)
698	{
699	LNXKAIOIOEVENT aPortEvents[AIO_MAXIMUM_REQUESTS_PER_CONTEXT];
700	int cRequestsToWait = RT_MIN(cReqs, AIO_MAXIMUM_REQUESTS_PER_CONTEXT);
701	ASMAtomicXchgBool(&pCtxInt->fWaiting, true);
702	rc = rtFileAsyncIoLinuxGetEvents(pCtxInt->AioContext, cMinReqs, cRequestsToWait, &aPortEvents[0], pTimeout);
703	ASMAtomicXchgBool(&pCtxInt->fWaiting, false);
704	if (RT_FAILURE(rc))
705	break;
706	uint32_t const cDone = rc;
707	rc = VINF_SUCCESS;
708
709	/*
710	* Process received events / requests.
711	*/
712	for (uint32_t i = 0; i < cDone; i++)
713	{
714	/*
715	* The iocb is the first element in our request structure.
716	* So we can safely cast it directly to the handle (see above)
717	*/
718	PRTFILEAIOREQINTERNAL pReqInt = (PRTFILEAIOREQINTERNAL)aPortEvents[i].pIoCB;
719	AssertPtr(pReqInt);
720	Assert(pReqInt->u32Magic == RTFILEAIOREQ_MAGIC);
721
722	/** @todo aeichner: The rc field contains the result code
723	* like you can find in errno for the normal read/write ops.
724	* But there is a second field called rc2. I don't know the
725	* purpose for it yet.
726	*/
727	if (RT_UNLIKELY(aPortEvents[i].rc < 0))
728	pReqInt->Rc = RTErrConvertFromErrno(-aPortEvents[i].rc); /* Convert to positive value. */
729	else
730	{
731	pReqInt->Rc = VINF_SUCCESS;
732	pReqInt->cbTransfered = aPortEvents[i].rc;
733	}
734
735	/* Mark the request as finished. */
736	RTFILEAIOREQ_SET_STATE(pReqInt, COMPLETED);
737
738	pahReqs[cRequestsCompleted++] = (RTFILEAIOREQ)pReqInt;
739	}
740
741	/*
742	* Done Yet? If not advance and try again.
743	*/
744	if (cDone >= cMinReqs)
745	break;
746	cMinReqs -= cDone;
747	cReqs -= cDone;
748
749	if (cMillies != RT_INDEFINITE_WAIT)
750	{
751	/* The API doesn't return ETIMEDOUT, so we have to fix that ourselves. */
752	uint64_t NanoTS = RTTimeNanoTS();
753	uint64_t cMilliesElapsed = (NanoTS - StartNanoTS) / 1000000;
754	if (cMilliesElapsed >= cMillies)
755	{
756	rc = VERR_TIMEOUT;
757	break;
758	}
759
760	/* The syscall supposedly updates it, but we're paranoid. :-) */
761	Timeout.tv_sec = (cMillies - (RTMSINTERVAL)cMilliesElapsed) / 1000;
762	Timeout.tv_nsec = (cMillies - (RTMSINTERVAL)cMilliesElapsed) % 1000 * 1000000;
763	}
764	}
765
766	/*
767	* Update the context state and set the return value.
768	*/
769	*pcReqs = cRequestsCompleted;
770	ASMAtomicSubS32(&pCtxInt->cRequests, cRequestsCompleted);
771	Assert(pCtxInt->hThreadWait == RTThreadSelf());
772	ASMAtomicWriteHandle(&pCtxInt->hThreadWait, NIL_RTTHREAD);
773
774	/*
775	* Clear the wakeup flag and set rc.
776	*/
777	if ( pCtxInt->fWokenUp
778	&& RT_SUCCESS(rc))
779	{
780	ASMAtomicXchgBool(&pCtxInt->fWokenUp, false);
781	rc = VERR_INTERRUPTED;
782	}
783
784	return rc;
785	}
786
787
788	RTDECL(int) RTFileAioCtxWakeup(RTFILEAIOCTX hAioCtx)
789	{
790	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
791	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
792
793	/** @todo r=bird: Define the protocol for how to resume work after calling
794	* this function. */
795
796	bool fWokenUp = ASMAtomicXchgBool(&pCtxInt->fWokenUp, true);
797
798	/*
799	* Read the thread handle before the status flag.
800	* If we read the handle after the flag we might
801	* end up with an invalid handle because the thread
802	* waiting in RTFileAioCtxWakeup() might get scheduled
803	* before we read the flag and returns.
804	* We can ensure that the handle is valid if fWaiting is true
805	* when reading the handle before the status flag.
806	*/
807	RTTHREAD hThread;
808	ASMAtomicReadHandle(&pCtxInt->hThreadWait, &hThread);
809	bool fWaiting = ASMAtomicReadBool(&pCtxInt->fWaiting);
810	if ( !fWokenUp
811	&& fWaiting)
812	{
813	/*
814	* If a thread waits the handle must be valid.
815	* It is possible that the thread returns from
816	* rtFileAsyncIoLinuxGetEvents() before the signal
817	* is send.
818	* This is no problem because we already set fWokenUp
819	* to true which will let the thread return VERR_INTERRUPTED
820	* and the next call to RTFileAioCtxWait() will not
821	* return VERR_INTERRUPTED because signals are not saved
822	* and will simply vanish if the destination thread can't
823	* receive it.
824	*/
825	Assert(hThread != NIL_RTTHREAD);
826	RTThreadPoke(hThread);
827	}
828
829	return VINF_SUCCESS;
830	}
831

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

source: vbox/trunk/src/VBox/Runtime/r3/linux/fileaio-linux.cpp@ 26618

Download in other formats: