fileaio-linux.cpp@ 19167

Last change on this file since 19167 was 19126, checked in by vboxsync, 16 years ago

Runtime/Aio: * Move the validation macros to a new header because they are used in

the other implmenetations as well.

Fix two bugs in the linux implementation:
1. If a request is canceled the request count is not decremented making it impossible to destroy a context.
2. Fix a race in the wakeup method which could lead to a nil thread handle.

Property svn:eol-style set to native
Property svn:keywords set to Author Date Id Revision

File size: 22.8 KB

Line
1	/* $Id: fileaio-linux.cpp 19126 2009-04-22 22:52:02Z 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 <linux/aio_abi.h>
66	#include <unistd.h>
67	#include <sys/syscall.h>
68	#include <errno.h>
69
70	#include <iprt/file.h>
71
72
73	/*******************************************************************************
74	* Structures and Typedefs *
75	*******************************************************************************/
76	/**
77	* The iocb structure of a request which is passed to the kernel.
78	*
79	* We redefined this here because the version in the header lacks padding
80	* for 32bit.
81	*/
82	typedef struct LNXKAIOIOCB
83	{
84	/** Opaque pointer to data which is returned on an I/O event. */
85	void *pvUser;
86	#ifdef RT_ARCH_X86
87	uint32_t u32Padding0;
88	#endif
89	/** Contains the request number and is set by the kernel. */
90	uint32_t u32Key;
91	/** Reserved. */
92	uint32_t u32Reserved0;
93	/** The I/O opcode. */
94	uint16_t u16IoOpCode;
95	/** Request priority. */
96	int16_t i16Priority;
97	/** The file descriptor. */
98	uint32_t File;
99	/** The userspace pointer to the buffer containing/receiving the data. */
100	void *pvBuf;
101	#ifdef RT_ARCH_X86
102	uint32_t u32Padding1;
103	#endif
104	/** How many bytes to transfer. */
105	#ifdef RT_ARCH_X86
106	uint32_t cbTransfer;
107	uint32_t u32Padding2;
108	#elif defined(RT_ARCH_AMD64)
109	uint64_t cbTransfer;
110	#else
111	# error "Unknown architecture"
112	#endif
113	/** At which offset to start the transfer. */
114	int64_t off;
115	/** Reserved. */
116	uint64_t u64Reserved1;
117	/** Flags */
118	uint32_t fFlags;
119	/** Readyness signal file descriptor. */
120	uint32_t u32ResFd;
121	} LNXKAIOIOCB, *PLNXKAIOIOCB;
122
123	/**
124	* I/O event structure to notify about completed requests.
125	* Redefined here too because of the padding.
126	*/
127	typedef struct LNXKAIOIOEVENT
128	{
129	/** The pvUser field from the iocb. */
130	void *pvUser;
131	#ifdef RT_ARCH_X86
132	uint32_t u32Padding0;
133	#endif
134	/** The LNXKAIOIOCB object this event is for. */
135	PLNXKAIOIOCB *pIoCB;
136	#ifdef RT_ARCH_X86
137	uint32_t u32Padding1;
138	#endif
139	/** The result code of the operation .*/
140	#ifdef RT_ARCH_X86
141	int32_t rc;
142	uint32_t u32Padding2;
143	#elif defined(RT_ARCH_AMD64)
144	int64_t rc;
145	#else
146	# error "Unknown architecture"
147	#endif
148	/** Secondary result code. */
149	#ifdef RT_ARCH_X86
150	int32_t rc2;
151	uint32_t u32Padding3;
152	#elif defined(RT_ARCH_AMD64)
153	int64_t rc2;
154	#else
155	# error "Unknown architecture"
156	#endif
157	} LNXKAIOIOEVENT, *PLNXKAIOIOEVENT;
158
159
160	/**
161	* Async I/O completion context state.
162	*/
163	typedef struct RTFILEAIOCTXINTERNAL
164	{
165	/** Handle to the async I/O context. */
166	aio_context_t AioContext;
167	/** Maximum number of requests this context can handle. */
168	int cRequestsMax;
169	/** Current number of requests active on this context. */
170	volatile int32_t cRequests;
171	/** The ID of the thread which is currently waiting for requests. */
172	volatile RTTHREAD hThreadWait;
173	/** Flag whether the thread was woken up. */
174	volatile bool fWokenUp;
175	/** Flag whether the thread is currently waiting in the syscall. */
176	volatile bool fWaiting;
177	/** Magic value (RTFILEAIOCTX_MAGIC). */
178	uint32_t u32Magic;
179	} RTFILEAIOCTXINTERNAL;
180	/** Pointer to an internal context structure. */
181	typedef RTFILEAIOCTXINTERNAL *PRTFILEAIOCTXINTERNAL;
182
183	/**
184	* Async I/O request state.
185	*/
186	typedef struct RTFILEAIOREQINTERNAL
187	{
188	/** The aio control block. This must be the FIRST elment in
189	* the structure! (see notes below) */
190	LNXKAIOIOCB AioCB;
191	/** The I/O context this request is associated with. */
192	aio_context_t AioContext;
193	/** Return code the request completed with. */
194	int Rc;
195	/** Flag whether the request is in process or not. */
196	bool fFinished;
197	/** Number of bytes actually trasnfered. */
198	size_t cbTransfered;
199	/** Completion context we are assigned to. */
200	PRTFILEAIOCTXINTERNAL pCtxInt;
201	/** Magic value (RTFILEAIOREQ_MAGIC). */
202	uint32_t u32Magic;
203	} RTFILEAIOREQINTERNAL;
204	/** Pointer to an internal request structure. */
205	typedef RTFILEAIOREQINTERNAL *PRTFILEAIOREQINTERNAL;
206
207
208	/*******************************************************************************
209	* Defined Constants And Macros *
210	*******************************************************************************/
211	/** The max number of events to get in one call. */
212	#define AIO_MAXIMUM_REQUESTS_PER_CONTEXT 64
213
214
215	/**
216	* Creates a new async I/O context.
217	*/
218	DECLINLINE(int) rtFileAsyncIoLinuxCreate(unsigned cEvents, aio_context_t *pAioContext)
219	{
220	int rc = syscall(__NR_io_setup, cEvents, pAioContext);
221	if (RT_UNLIKELY(rc == -1))
222	return RTErrConvertFromErrno(errno);
223
224	return VINF_SUCCESS;
225	}
226
227	/**
228	* Destroys a async I/O context.
229	*/
230	DECLINLINE(int) rtFileAsyncIoLinuxDestroy(aio_context_t AioContext)
231	{
232	int rc = syscall(__NR_io_destroy, AioContext);
233	if (RT_UNLIKELY(rc == -1))
234	return RTErrConvertFromErrno(errno);
235
236	return VINF_SUCCESS;
237	}
238
239	/**
240	* Submits an array of I/O requests to the kernel.
241	*/
242	DECLINLINE(int) rtFileAsyncIoLinuxSubmit(aio_context_t AioContext, long cReqs, LNXKAIOIOCB **ppIoCB)
243	{
244	int rc = syscall(__NR_io_submit, AioContext, cReqs, ppIoCB);
245	if (RT_UNLIKELY(rc == -1))
246	return RTErrConvertFromErrno(errno);
247
248	return VINF_SUCCESS;
249	}
250
251	/**
252	* Cancels a I/O request.
253	*/
254	DECLINLINE(int) rtFileAsyncIoLinuxCancel(aio_context_t AioContext, PLNXKAIOIOCB pIoCB, PLNXKAIOIOEVENT pIoResult)
255	{
256	int rc = syscall(__NR_io_cancel, AioContext, pIoCB, pIoResult);
257	if (RT_UNLIKELY(rc == -1))
258	return RTErrConvertFromErrno(errno);
259
260	return VINF_SUCCESS;
261	}
262
263	/**
264	* Waits for I/O events.
265	* @returns Number of events (natural number w/ 0), IPRT error code (negative).
266	*/
267	DECLINLINE(int) rtFileAsyncIoLinuxGetEvents(aio_context_t AioContext, long cReqsMin, long cReqs,
268	PLNXKAIOIOEVENT paIoResults, struct timespec *pTimeout)
269	{
270	int rc = syscall(__NR_io_getevents, AioContext, cReqsMin, cReqs, paIoResults, pTimeout);
271	if (RT_UNLIKELY(rc == -1))
272	return RTErrConvertFromErrno(errno);
273
274	return rc;
275	}
276
277	RTR3DECL(int) RTFileAioReqCreate(PRTFILEAIOREQ phReq)
278	{
279	AssertPtrReturn(phReq, VERR_INVALID_POINTER);
280
281	/*
282	* Allocate a new request and initialize it.
283	*/
284	PRTFILEAIOREQINTERNAL pReqInt = (PRTFILEAIOREQINTERNAL)RTMemAllocZ(sizeof(*pReqInt));
285	if (RT_UNLIKELY(!pReqInt))
286	return VERR_NO_MEMORY;
287
288	pReqInt->fFinished = false;
289	pReqInt->pCtxInt = NULL;
290	pReqInt->u32Magic = RTFILEAIOREQ_MAGIC;
291
292	*phReq = (RTFILEAIOREQ)pReqInt;
293	return VINF_SUCCESS;
294	}
295
296
297	RTDECL(void) RTFileAioReqDestroy(RTFILEAIOREQ hReq)
298	{
299	/*
300	* Validate the handle and ignore nil.
301	*/
302	if (hReq == NIL_RTFILEAIOREQ)
303	return;
304	PRTFILEAIOREQINTERNAL pReqInt = hReq;
305	RTFILEAIOREQ_VALID_RETURN_VOID(pReqInt);
306
307	/*
308	* Trash the magic and free it.
309	*/
310	ASMAtomicUoWriteU32(&pReqInt->u32Magic, ~RTFILEAIOREQ_MAGIC);
311	RTMemFree(pReqInt);
312	}
313
314
315	/**
316	* Worker setting up the request.
317	*/
318	DECLINLINE(int) rtFileAioReqPrepareTransfer(RTFILEAIOREQ hReq, RTFILE hFile,
319	uint16_t uTransferDirection,
320	RTFOFF off, void *pvBuf, size_t cbTransfer,
321	void *pvUser)
322	{
323	/*
324	* Validate the input.
325	*/
326	PRTFILEAIOREQINTERNAL pReqInt = hReq;
327	RTFILEAIOREQ_VALID_RETURN(pReqInt);
328	Assert(hFile != NIL_RTFILE);
329	AssertPtr(pvBuf);
330	Assert(off >= 0);
331	Assert(cbTransfer > 0);
332
333	/*
334	* Setup the control block and clear the finished flag.
335	*/
336	pReqInt->AioCB.u16IoOpCode = uTransferDirection;
337	pReqInt->AioCB.File = (uint32_t)hFile;
338	pReqInt->AioCB.off = off;
339	pReqInt->AioCB.cbTransfer = cbTransfer;
340	pReqInt->AioCB.pvBuf = pvBuf;
341	pReqInt->AioCB.pvUser = pvUser;
342
343	pReqInt->fFinished = false;
344	pReqInt->pCtxInt = NULL;
345
346	return VINF_SUCCESS;
347	}
348
349
350	RTDECL(int) RTFileAioReqPrepareRead(RTFILEAIOREQ hReq, RTFILE hFile, RTFOFF off,
351	void pvBuf, size_t cbRead, void pvUser)
352	{
353	return rtFileAioReqPrepareTransfer(hReq, hFile, IOCB_CMD_PREAD,
354	off, pvBuf, cbRead, pvUser);
355	}
356
357
358	RTDECL(int) RTFileAioReqPrepareWrite(RTFILEAIOREQ hReq, RTFILE hFile, RTFOFF off,
359	void pvBuf, size_t cbWrite, void pvUser)
360	{
361	return rtFileAioReqPrepareTransfer(hReq, hFile, IOCB_CMD_PWRITE,
362	off, pvBuf, cbWrite, pvUser);
363	}
364
365
366	RTDECL(int) RTFileAioReqPrepareFlush(RTFILEAIOREQ hReq, RTFILE hFile, void *pvUser)
367	{
368	PRTFILEAIOREQINTERNAL pReqInt = hReq;
369	RTFILEAIOREQ_VALID_RETURN(pReqInt);
370	AssertReturn(hFile != NIL_RTFILE, VERR_INVALID_HANDLE);
371
372	/** @todo: Flushing is not neccessary on Linux because O_DIRECT is mandatory
373	* which disables caching.
374	* We could setup a fake request which isn't really executed
375	* to avoid platform dependent code in the caller.
376	*/
377	#if 0
378	return rtFileAsyncPrepareTransfer(pRequest, File, TRANSFERDIRECTION_FLUSH,
379	0, NULL, 0, pvUser);
380	#endif
381	return VERR_NOT_IMPLEMENTED;
382	}
383
384
385	RTDECL(void *) RTFileAioReqGetUser(RTFILEAIOREQ hReq)
386	{
387	PRTFILEAIOREQINTERNAL pReqInt = hReq;
388	RTFILEAIOREQ_VALID_RETURN_RC(pReqInt, NULL);
389
390	return pReqInt->AioCB.pvUser;
391	}
392
393
394	RTDECL(int) RTFileAioReqCancel(RTFILEAIOREQ hReq)
395	{
396	PRTFILEAIOREQINTERNAL pReqInt = hReq;
397	RTFILEAIOREQ_VALID_RETURN(pReqInt);
398
399	LNXKAIOIOEVENT AioEvent;
400	int rc = rtFileAsyncIoLinuxCancel(pReqInt->AioContext, &pReqInt->AioCB, &AioEvent);
401	if (RT_SUCCESS(rc))
402	{
403	/*
404	* Decrement request count because the request will never arrive at the
405	* completion port.
406	*/
407	AssertMsg(VALID_PTR(pReqInt->pCtxInt),
408	("Invalid state. Request was canceled but wasn't submitted\n"));
409
410	ASMAtomicDecS32(&pReqInt->pCtxInt->cRequests);
411	return VINF_SUCCESS;
412	}
413	if (rc == VERR_TRY_AGAIN)
414	return VERR_FILE_AIO_IN_PROGRESS;
415	return rc;
416	}
417
418
419	RTDECL(int) RTFileAioReqGetRC(RTFILEAIOREQ hReq, size_t *pcbTransfered)
420	{
421	PRTFILEAIOREQINTERNAL pReqInt = hReq;
422	RTFILEAIOREQ_VALID_RETURN(pReqInt);
423	AssertPtrNull(pcbTransfered);
424
425	if (!pReqInt->fFinished)
426	return VERR_FILE_AIO_IN_PROGRESS;
427
428	if ( pcbTransfered
429	&& RT_SUCCESS(pReqInt->Rc))
430	*pcbTransfered = pReqInt->cbTransfered;
431
432	return pReqInt->Rc;
433	}
434
435
436	RTDECL(int) RTFileAioCtxCreate(PRTFILEAIOCTX phAioCtx, uint32_t cAioReqsMax)
437	{
438	PRTFILEAIOCTXINTERNAL pCtxInt;
439	AssertPtrReturn(phAioCtx, VERR_INVALID_POINTER);
440
441	/* The kernel interface needs a maximum. */
442	if (cAioReqsMax == RTFILEAIO_UNLIMITED_REQS)
443	return VERR_OUT_OF_RANGE;
444
445	pCtxInt = (PRTFILEAIOCTXINTERNAL)RTMemAllocZ(sizeof(RTFILEAIOCTXINTERNAL));
446	if (RT_UNLIKELY(!pCtxInt))
447	return VERR_NO_MEMORY;
448
449	/* Init the event handle. */
450	int rc = rtFileAsyncIoLinuxCreate(cAioReqsMax, &pCtxInt->AioContext);
451	if (RT_SUCCESS(rc))
452	{
453	pCtxInt->fWokenUp = false;
454	pCtxInt->fWaiting = false;
455	pCtxInt->hThreadWait = NIL_RTTHREAD;
456	pCtxInt->cRequestsMax = cAioReqsMax;
457	pCtxInt->u32Magic = RTFILEAIOCTX_MAGIC;
458	*phAioCtx = (RTFILEAIOCTX)pCtxInt;
459	}
460
461	return rc;
462	}
463
464
465	RTDECL(int) RTFileAioCtxDestroy(RTFILEAIOCTX hAioCtx)
466	{
467	/* Validate the handle and ignore nil. */
468	if (hAioCtx == NIL_RTFILEAIOCTX)
469	return VINF_SUCCESS;
470	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
471	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
472
473	/* Cannot destroy a busy context. */
474	if (RT_UNLIKELY(pCtxInt->cRequests))
475	return VERR_FILE_AIO_BUSY;
476
477	/* The native bit first, then mark it as dead and free it. */
478	int rc = rtFileAsyncIoLinuxDestroy(pCtxInt->AioContext);
479	if (RT_FAILURE(rc))
480	return rc;
481	ASMAtomicUoWriteU32(&pCtxInt->u32Magic, RTFILEAIOCTX_MAGIC_DEAD);
482	RTMemFree(pCtxInt);
483
484	return VINF_SUCCESS;
485	}
486
487
488	RTDECL(uint32_t) RTFileAioCtxGetMaxReqCount(RTFILEAIOCTX hAioCtx)
489	{
490	/* Nil means global here. */
491	if (hAioCtx == NIL_RTFILEAIOCTX)
492	return RTFILEAIO_UNLIMITED_REQS; /** @todo r=bird: I'm a bit puzzled by this return value since it
493	* is completely useless in RTFileAioCtxCreate. */
494
495	/* Return 0 if the handle is invalid, it's better than garbage I think... */
496	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
497	RTFILEAIOCTX_VALID_RETURN_RC(pCtxInt, 0);
498
499	return pCtxInt->cRequestsMax;
500	}
501
502
503	RTDECL(int) RTFileAioCtxSubmit(RTFILEAIOCTX hAioCtx, PRTFILEAIOREQ pahReqs, size_t cReqs)
504	{
505	/*
506	* Parameter validation.
507	*/
508	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
509	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
510	AssertReturn(cReqs > 0, VERR_INVALID_PARAMETER);
511	AssertPtrReturn(pahReqs, VERR_INVALID_POINTER);
512	uint32_t i = cReqs;
513
514	/*
515	* Vaildate requests and associate with the context.
516	*/
517	while (i-- > 0)
518	{
519	PRTFILEAIOREQINTERNAL pReqInt = pahReqs[i];
520	RTFILEAIOREQ_VALID_RETURN(pReqInt);
521
522	pReqInt->AioContext = pCtxInt->AioContext;
523	pReqInt->pCtxInt = pCtxInt;
524	}
525
526	/*
527	* Add the submitted requests to the counter
528	* to prevent destroying the context while
529	* it is still used.
530	*/
531	ASMAtomicAddS32(&pCtxInt->cRequests, cReqs);
532
533	/*
534	* We cast phReqs to the Linux iocb structure to avoid copying the requests
535	* into a temporary array. This is possible because the iocb structure is
536	* the first element in the request structure (see PRTFILEAIOCTXINTERNAL).
537	*/
538	int rc = rtFileAsyncIoLinuxSubmit(pCtxInt->AioContext, cReqs, (PLNXKAIOIOCB *)pahReqs);
539	if (RT_FAILURE(rc))
540	ASMAtomicSubS32(&pCtxInt->cRequests, cReqs);
541
542	return rc;
543	}
544
545
546	RTDECL(int) RTFileAioCtxWait(RTFILEAIOCTX hAioCtx, size_t cMinReqs, unsigned cMillisTimeout,
547	PRTFILEAIOREQ pahReqs, size_t cReqs, uint32_t *pcReqs)
548	{
549	/*
550	* Validate the parameters, making sure to always set pcReqs.
551	*/
552	AssertPtrReturn(pcReqs, VERR_INVALID_POINTER);
553	pcReqs = 0; / always set */
554	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
555	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
556	AssertPtrReturn(pahReqs, VERR_INVALID_POINTER);
557	AssertReturn(cReqs != 0, VERR_INVALID_PARAMETER);
558	AssertReturn(cReqs >= cMinReqs, VERR_OUT_OF_RANGE);
559
560	/*
561	* Can't wait if there are not requests around.
562	*/
563	if (RT_UNLIKELY(ASMAtomicUoReadS32(&pCtxInt->cRequests) == 0))
564	return VERR_FILE_AIO_NO_REQUEST;
565
566	/*
567	* Convert the timeout if specified.
568	*/
569	struct timespec *pTimeout = NULL;
570	struct timespec Timeout = {0,0};
571	uint64_t StartNanoTS = 0;
572	if (cMillisTimeout != RT_INDEFINITE_WAIT)
573	{
574	Timeout.tv_sec = cMillisTimeout / 1000;
575	Timeout.tv_nsec = cMillisTimeout % 1000 * 1000000;
576	pTimeout = &Timeout;
577	StartNanoTS = RTTimeNanoTS();
578	}
579
580	/* Wait for at least one. */
581	if (!cMinReqs)
582	cMinReqs = 1;
583
584	/* For the wakeup call. */
585	Assert(pCtxInt->hThreadWait == NIL_RTTHREAD);
586	ASMAtomicWriteHandle(&pCtxInt->hThreadWait, RTThreadSelf());
587
588	/*
589	* Loop until we're woken up, hit an error (incl timeout), or
590	* have collected the desired number of requests.
591	*/
592	int rc = VINF_SUCCESS;
593	int cRequestsCompleted = 0;
594	while (!pCtxInt->fWokenUp)
595	{
596	LNXKAIOIOEVENT aPortEvents[AIO_MAXIMUM_REQUESTS_PER_CONTEXT];
597	int cRequestsToWait = RT_MIN(cReqs, AIO_MAXIMUM_REQUESTS_PER_CONTEXT);
598	ASMAtomicXchgBool(&pCtxInt->fWaiting, true);
599	rc = rtFileAsyncIoLinuxGetEvents(pCtxInt->AioContext, cMinReqs, cRequestsToWait, &aPortEvents[0], pTimeout);
600	ASMAtomicXchgBool(&pCtxInt->fWaiting, false);
601	if (RT_FAILURE(rc))
602	break;
603	uint32_t const cDone = rc;
604	rc = VINF_SUCCESS;
605
606	/*
607	* Process received events / requests.
608	*/
609	for (uint32_t i = 0; i < cDone; i++)
610	{
611	/*
612	* The iocb is the first element in our request structure.
613	* So we can safely cast it directly to the handle (see above)
614	*/
615	PRTFILEAIOREQINTERNAL pReqInt = (PRTFILEAIOREQINTERNAL)aPortEvents[i].pIoCB;
616	AssertPtr(pReqInt);
617	Assert(pReqInt->u32Magic == RTFILEAIOREQ_MAGIC);
618
619	/** @todo aeichner: The rc field contains the result code
620	* like you can find in errno for the normal read/write ops.
621	* But there is a second field called rc2. I don't know the
622	* purpose for it yet.
623	*/
624	if (RT_UNLIKELY(aPortEvents[i].rc < 0))
625	pReqInt->Rc = RTErrConvertFromErrno(aPortEvents[i].rc);
626	else
627	{
628	pReqInt->Rc = VINF_SUCCESS;
629	pReqInt->cbTransfered = aPortEvents[i].rc;
630	}
631
632	/* Mark the request as finished. */
633	pReqInt->fFinished = true;
634
635	pahReqs[cRequestsCompleted++] = (RTFILEAIOREQ)pReqInt;
636	}
637
638	/*
639	* Done Yet? If not advance and try again.
640	*/
641	if (cDone >= cMinReqs)
642	break;
643	cMinReqs -= cDone;
644	cReqs -= cDone;
645
646	if (cMillisTimeout != RT_INDEFINITE_WAIT)
647	{
648	/* The API doesn't return ETIMEDOUT, so we have to fix that ourselves. */
649	uint64_t NanoTS = RTTimeNanoTS();
650	uint64_t cMilliesElapsed = (NanoTS - StartNanoTS) / 1000000;
651	if (cMilliesElapsed >= cMillisTimeout)
652	{
653	rc = VERR_TIMEOUT;
654	break;
655	}
656
657	/* The syscall supposedly updates it, but we're paranoid. :-) */
658	Timeout.tv_sec = (cMillisTimeout - (unsigned)cMilliesElapsed) / 1000;
659	Timeout.tv_nsec = (cMillisTimeout - (unsigned)cMilliesElapsed) % 1000 * 1000000;
660	}
661	}
662
663	/*
664	* Update the context state and set the return value.
665	*/
666	*pcReqs = cRequestsCompleted;
667	ASMAtomicSubS32(&pCtxInt->cRequests, cRequestsCompleted);
668	Assert(pCtxInt->hThreadWait == RTThreadSelf());
669	ASMAtomicWriteHandle(&pCtxInt->hThreadWait, NIL_RTTHREAD);
670
671	/*
672	* Clear the wakeup flag and set rc.
673	*/
674	if ( pCtxInt->fWokenUp
675	&& RT_SUCCESS(rc))
676	{
677	ASMAtomicXchgBool(&pCtxInt->fWokenUp, false);
678	rc = VERR_INTERRUPTED;
679	}
680
681	return rc;
682	}
683
684
685	RTDECL(int) RTFileAioCtxWakeup(RTFILEAIOCTX hAioCtx)
686	{
687	PRTFILEAIOCTXINTERNAL pCtxInt = hAioCtx;
688	RTFILEAIOCTX_VALID_RETURN(pCtxInt);
689
690	/** @todo r=bird: Define the protocol for how to resume work after calling
691	* this function. */
692
693	bool fWokenUp = ASMAtomicXchgBool(&pCtxInt->fWokenUp, true);
694
695	/*
696	* Read the thread handle before the status flag.
697	* If we read the handle after the flag we might
698	* end up with an invalid handle because the thread
699	* waiting in RTFileAioCtxWakeup() might get scheduled
700	* before we read the flag and returns.
701	* We can ensure that the handle is valid if fWaiting is true
702	* when reading the handle before the status flag.
703	*/
704	RTTHREAD hThread;
705	ASMAtomicReadHandle(&pCtxInt->hThreadWait, &hThread);
706	bool fWaiting = ASMAtomicReadBool(&pCtxInt->fWaiting);
707	if ( !fWokenUp
708	&& fWaiting)
709	{
710	/*
711	* If a thread waits the handle must be valid.
712	* It is possible that the thread returns from
713	* rtFileAsyncIoLinuxGetEvents() before the signal
714	* is send.
715	* This is no problem because we already set fWokenUp
716	* to true which will let the thread return VERR_INTERRUPTED
717	* and the next call to RTFileAioCtxWait() will not
718	* return VERR_INTERRUPTED because signals are not saved
719	* and will simply vanish if the destination thread can't
720	* receive it.
721	*/
722	Assert(hThread != NIL_RTTHREAD);
723	RTThreadPoke(hThread);
724	}
725
726	return VINF_SUCCESS;
727	}
728

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

source: vbox/trunk/src/VBox/Runtime/r3/linux/fileaio-linux.cpp@ 19167

Download in other formats: