[fio.git] / engines / windowsaio.c

/*
 * Native Windows async IO engine
 * Copyright (C) 2011 Bruce Cran <bruce@cran.org.uk>
 */

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <signal.h>
#include <errno.h>
#include <windows.h>

#include "../fio.h"

typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);

struct fio_overlapped {
	OVERLAPPED o;
	struct io_u *io_u;
	BOOL io_complete;
	BOOL io_free;
};

struct windowsaio_data {
	struct fio_overlapped *ovls;
	struct io_u **aio_events;
	HANDLE iothread;
	HANDLE iocomplete_event;
	CANCELIOEX pCancelIoEx;
	BOOL iothread_running;
};

struct thread_ctx {
	HANDLE iocp;
	struct windowsaio_data *wd;
};

static int fio_windowsaio_cancel(struct thread_data *td,
			       struct io_u *io_u);
static BOOL timeout_expired(DWORD start_count, DWORD end_count);
static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
				    unsigned int max, struct timespec *t);
static struct io_u *fio_windowsaio_event(struct thread_data *td, int event);
static int fio_windowsaio_queue(struct thread_data *td,
			      struct io_u *io_u);
static void fio_windowsaio_cleanup(struct thread_data *td);
static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
static int fio_windowsaio_init(struct thread_data *td);
static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f);
static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f);

int sync_file_range(int fd, off64_t offset, off64_t nbytes,
			   unsigned int flags)
{
	errno = ENOSYS;
	return -1;
}

static int fio_windowsaio_init(struct thread_data *td)
{
	struct windowsaio_data *wd;
	HANDLE hKernel32Dll;
	int rc = 0;
	int i;

	wd = malloc(sizeof(struct windowsaio_data));
	if (wd != NULL)
		ZeroMemory(wd, sizeof(struct windowsaio_data));
	else
		rc = 1;

	if (!rc) {
		wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
		if (wd->aio_events == NULL)
			rc = 1;
	}

	if (!rc) {
		wd->ovls = malloc(td->o.iodepth * sizeof(struct fio_overlapped));
		if (wd->ovls == NULL)
			rc = 1;
	}

	if (!rc) {
	    for (i = 0; i < td->o.iodepth; i++) {
	        wd->ovls[i].io_free = TRUE;
	        wd->ovls[i].io_complete = FALSE;

			wd->ovls[i].o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
			if (wd->ovls[i].o.hEvent == NULL) {
				rc = 1;
				break;
			}
	    }
	}

	if (!rc) {
		/* Create an auto-reset event */
		wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
		if (wd->iocomplete_event == NULL)
			rc = 1;
	}

	if (rc) {
		if (wd != NULL) {
			if (wd->ovls != NULL)
				free(wd->ovls);
			if (wd->aio_events != NULL)
				free(wd->aio_events);

			free(wd);
		}
	}

	hKernel32Dll = GetModuleHandle("kernel32.dll");
	wd->pCancelIoEx = GetProcAddress(hKernel32Dll, "CancelIoEx");

	td->io_ops->data = wd;
	return rc;
}

static void fio_windowsaio_cleanup(struct thread_data *td)
{
	int i;
	struct windowsaio_data *wd;

	wd = td->io_ops->data;

	if (wd != NULL) {
		wd->iothread_running = FALSE;
		WaitForSingleObject(wd->iothread, INFINITE);

		CloseHandle(wd->iothread);
		CloseHandle(wd->iocomplete_event);

		for (i = 0; i < td->o.iodepth; i++)
			CloseHandle(wd->ovls[i].o.hEvent);

		free(wd->aio_events);
		free(wd->ovls);
		free(wd);

		td->io_ops->data = NULL;
	}
}


static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
{
	int rc = 0;
	HANDLE hFile;
	DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED;
	DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE;
	DWORD openmode = OPEN_ALWAYS;
	DWORD access;

	dprint(FD_FILE, "fd open %s\n", f->file_name);

	if (f->filetype == FIO_TYPE_PIPE) {
		log_err("fio: windowsaio doesn't support pipes\n");
		return 1;
	}

	if (!strcmp(f->file_name, "-")) {
		log_err("fio: can't read/write to stdin/out\n");
		return 1;
	}

	if (td->o.odirect)
		flags |= FILE_FLAG_NO_BUFFERING;
	if (td->o.sync_io)
		flags |= FILE_FLAG_WRITE_THROUGH;

	if (td->o.td_ddir == TD_DDIR_READ  ||
		td->o.td_ddir == TD_DDIR_WRITE)
		flags |= FILE_FLAG_SEQUENTIAL_SCAN;
	else
		flags |= FILE_FLAG_RANDOM_ACCESS;

	if (!td_write(td) || read_only)
		access = GENERIC_READ;
	else
		access = (GENERIC_READ | GENERIC_WRITE);

	if (td->o.create_on_open > 0)
		openmode = OPEN_ALWAYS;
	else
		openmode = OPEN_EXISTING;

	f->hFile = CreateFile(f->file_name, access, sharemode,
		NULL, openmode, flags, NULL);

	if (f->hFile == INVALID_HANDLE_VALUE)
		rc = 1;

	/* Only set up the competion port and thread if we're not just
	 * querying the device size */
	if (!rc && td->io_ops->data != NULL) {
		struct thread_ctx *ctx;
		struct windowsaio_data *wd;

		hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);

		wd = td->io_ops->data;
		wd->iothread_running = TRUE;

		if (!rc) {
			ctx = malloc(sizeof(struct thread_ctx));
			ctx->iocp = hFile;
			ctx->wd = wd;

			wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
		}

		if (rc || wd->iothread == NULL)
			rc = 1;
	}

	return rc;
}

static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f)
{
	int rc = 0;

	dprint(FD_FILE, "fd close %s\n", f->file_name);

	if (f->hFile != INVALID_HANDLE_VALUE) {
		if (!CloseHandle(f->hFile))
			rc = 1;
	}

	f->hFile = INVALID_HANDLE_VALUE;
	return rc;
}

static BOOL timeout_expired(DWORD start_count, DWORD end_count)
{
	BOOL expired = FALSE;
	DWORD current_time;

	current_time = GetTickCount();

	if ((end_count > start_count) && current_time >= end_count)
		expired = TRUE;
	else if (current_time < start_count && current_time > end_count)
		expired = TRUE;

	return expired;
}

static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
{
	struct windowsaio_data *wd = td->io_ops->data;
	return wd->aio_events[event];
}

static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
				    unsigned int max, struct timespec *t)
{
	struct windowsaio_data *wd = td->io_ops->data;
	struct flist_head *entry;
	unsigned int dequeued = 0;
	struct io_u *io_u;
	struct fio_overlapped *fov;
	DWORD start_count = 0;
	DWORD end_count = 0;
	DWORD status;
	DWORD mswait = 250;

	if (t != NULL) {
		mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
		start_count = GetTickCount();
		end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	}

	do {
		flist_for_each(entry, &td->io_u_busylist) {
			io_u = flist_entry(entry, struct io_u, list);
			fov = (struct fio_overlapped*)io_u->engine_data;

			if (fov->io_complete) {
				fov->io_complete = FALSE;
				fov->io_free  = TRUE;
				wd->aio_events[dequeued] = io_u;
				dequeued++;
			}

			if (dequeued >= min)
				break;
		}

		if (dequeued < min) {
			status = WaitForSingleObject(wd->iocomplete_event, mswait);
			if (status != WAIT_OBJECT_0 && dequeued > 0)
			    break;
		}

		if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count)))
			break;
	} while (1);

	return dequeued;
}

static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u)
{
	LPOVERLAPPED lpOvl = NULL;
	struct windowsaio_data *wd;
	DWORD iobytes;
	BOOL success;
	int index;
	int rc = FIO_Q_COMPLETED;

	fio_ro_check(td, io_u);

	wd = td->io_ops->data;

	for (index = 0; index < td->o.iodepth; index++) {
		if (wd->ovls[index].io_free) {
			wd->ovls[index].io_free = FALSE;
			ResetEvent(wd->ovls[index].o.hEvent);
			break;
		}
	}

	assert(index < td->o.iodepth);

	lpOvl = &wd->ovls[index].o;
	wd->ovls[index].io_u = io_u;
	lpOvl->Internal = STATUS_PENDING;
	lpOvl->InternalHigh = 0;
	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	lpOvl->OffsetHigh = io_u->offset >> 32;
	io_u->engine_data = &wd->ovls[index];

	switch (io_u->ddir) {
	case DDIR_WRITE:
		success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
		break;
	case DDIR_READ:
		success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
		break;
	case DDIR_SYNC:
	case DDIR_DATASYNC:
	case DDIR_SYNC_FILE_RANGE:
		success = FlushFileBuffers(io_u->file->hFile);
		if (!success)
		    io_u->error = GetLastError();

		return FIO_Q_COMPLETED;
		break;
	case DDIR_TRIM:
		log_err("manual TRIM isn't supported on Windows");
		io_u->error = 1;
		io_u->resid = io_u->xfer_buflen;
		return FIO_Q_COMPLETED;
		break;
	default:
		assert(0);
	}

	if (success || GetLastError() == ERROR_IO_PENDING)
		rc = FIO_Q_QUEUED;
	else {
		io_u->error = GetLastError();
		io_u->resid = io_u->xfer_buflen;
	}

	return rc;
}

/* Runs as a thread and waits for queued IO to complete */
static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
{
	OVERLAPPED *ovl;
	struct fio_overlapped *fov;
	struct io_u *io_u;
	struct windowsaio_data *wd;
	struct thread_ctx *ctx;
	ULONG_PTR ulKey = 0;
	DWORD bytes;

	ctx = (struct thread_ctx*)lpParameter;
	wd = ctx->wd;

	do {
		if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250))
			continue;

		fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
		io_u = fov->io_u;

		if (ovl->Internal == ERROR_SUCCESS) {
			io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
			io_u->error = 0;
		} else {
			io_u->resid = io_u->xfer_buflen;
			io_u->error = ovl->Internal;
		}

		fov->io_complete = TRUE;
		SetEvent(wd->iocomplete_event);
	} while (ctx->wd->iothread_running);

	CloseHandle(ctx->iocp);
	free(ctx);
	return 0;
}

static int fio_windowsaio_cancel(struct thread_data *td,
			       struct io_u *io_u)
{
	int rc = 0;

	struct windowsaio_data *wd = td->io_ops->data;

	/* If we're running on Vista or newer, we can cancel individual IO requests */
	if (wd->pCancelIoEx != NULL) {
		struct fio_overlapped *ovl = io_u->engine_data;

		if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
			rc = 1;
	} else
		rc = 1;

	return rc;
}

static struct ioengine_ops ioengine = {
	.name		= "windowsaio",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_windowsaio_init,
	.queue		= fio_windowsaio_queue,
	.cancel		= fio_windowsaio_cancel,
	.getevents	= fio_windowsaio_getevents,
	.event		= fio_windowsaio_event,
	.cleanup	= fio_windowsaio_cleanup,
	.open_file	= fio_windowsaio_open_file,
	.close_file	= fio_windowsaio_close_file,
	.get_file_size	= generic_get_file_size
};

static void fio_init fio_posixaio_register(void)
{
	register_ioengine(&ioengine);
}

static void fio_exit fio_posixaio_unregister(void)
{
	unregister_ioengine(&ioengine);
}
Commit	Line	Data
ecc314ba BC	1	/*
ecc314ba BC	2	* Native Windows async IO engine
ea4500d8	3	* Copyright (C) 2011 Bruce Cran <bruce@cran.org.uk>
ecc314ba BC	4	*/
ecc314ba BC	5
ecc314ba BC	6	#include <stdio.h>
	7	#include <stdlib.h>
	8	#include <unistd.h>
	9	#include <signal.h>
	10	#include <errno.h>
	11	#include <windows.h>
	12
	13	#include "../fio.h"
	14
ea4500d8 BC	15	typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);
ea4500d8 BC	16
67897036 BC	17	struct fio_overlapped {
	18	OVERLAPPED o;
	19	struct io_u *io_u;
	20	BOOL io_complete;
40c5db35	21	BOOL io_free;
67897036	22	};
ecc314ba	23
9b836561	24	struct windowsaio_data {
67897036	25	struct fio_overlapped *ovls;
9b836561	26	struct io_u **aio_events;
67897036	27	HANDLE iothread;
9b836561	28	HANDLE iocomplete_event;
ea4500d8	29	CANCELIOEX pCancelIoEx;
67897036	30	BOOL iothread_running;
9b836561 BC	31	};
9b836561 BC	32
ecc314ba	33	struct thread_ctx {
9b836561	34	HANDLE iocp;
ecc314ba BC	35	struct windowsaio_data *wd;
	36	};
	37
ecc314ba BC	38	static int fio_windowsaio_cancel(struct thread_data *td,
ecc314ba BC	39	struct io_u *io_u);
67897036	40	static BOOL timeout_expired(DWORD start_count, DWORD end_count);
ecc314ba BC	41	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	42	unsigned int max, struct timespec *t);
	43	static struct io_u fio_windowsaio_event(struct thread_data td, int event);
	44	static int fio_windowsaio_queue(struct thread_data *td,
	45	struct io_u *io_u);
	46	static void fio_windowsaio_cleanup(struct thread_data *td);
	47	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
	48	static int fio_windowsaio_init(struct thread_data *td);
	49	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f);
	50	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f);
	51
9b836561 BC	52	int sync_file_range(int fd, off64_t offset, off64_t nbytes,
	53	unsigned int flags)
	54	{
	55	errno = ENOSYS;
	56	return -1;
	57	}
	58
ecc314ba BC	59	static int fio_windowsaio_init(struct thread_data *td)
ecc314ba BC	60	{
ecc314ba	61	struct windowsaio_data *wd;
ea4500d8	62	HANDLE hKernel32Dll;
9b836561	63	int rc = 0;
67897036	64	int i;
ecc314ba	65
ecc314ba	66	wd = malloc(sizeof(struct windowsaio_data));
9b836561 BC	67	if (wd != NULL)
	68	ZeroMemory(wd, sizeof(struct windowsaio_data));
	69	else
	70	rc = 1;
ecc314ba	71
9b836561 BC	72	if (!rc) {
	73	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	74	if (wd->aio_events == NULL)
	75	rc = 1;
e4db9fec BC	76	}
e4db9fec BC	77
9b836561	78	if (!rc) {
67897036 BC	79	wd->ovls = malloc(td->o.iodepth * sizeof(struct fio_overlapped));
67897036 BC	80	if (wd->ovls == NULL)
9b836561	81	rc = 1;
e4db9fec BC	82	}
e4db9fec BC	83
9b836561	84	if (!rc) {
67897036 BC	85	for (i = 0; i < td->o.iodepth; i++) {
	86	wd->ovls[i].io_free = TRUE;
	87	wd->ovls[i].io_complete = FALSE;
	88
	89	wd->ovls[i].o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
	90	if (wd->ovls[i].o.hEvent == NULL) {
	91	rc = 1;
	92	break;
	93	}
	94	}
9b836561	95	}
e4db9fec	96
9b836561 BC	97	if (!rc) {
	98	/* Create an auto-reset event */
	99	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	100	if (wd->iocomplete_event == NULL)
	101	rc = 1;
	102	}
	103
9b836561	104	if (rc) {
9b836561 BC	105	if (wd != NULL) {
	106	if (wd->ovls != NULL)
	107	free(wd->ovls);
9b836561 BC	108	if (wd->aio_events != NULL)
	109	free(wd->aio_events);
	110
	111	free(wd);
	112	}
	113	}
ecc314ba	114
ea4500d8 BC	115	hKernel32Dll = GetModuleHandle("kernel32.dll");
	116	wd->pCancelIoEx = GetProcAddress(hKernel32Dll, "CancelIoEx");
	117
ecc314ba	118	td->io_ops->data = wd;
4e79098f	119	return rc;
ecc314ba BC	120	}
ecc314ba BC	121
67897036 BC	122	static void fio_windowsaio_cleanup(struct thread_data *td)
	123	{
	124	int i;
	125	struct windowsaio_data *wd;
	126
	127	wd = td->io_ops->data;
	128
	129	if (wd != NULL) {
40c5db35 JA	130	wd->iothread_running = FALSE;
40c5db35 JA	131	WaitForSingleObject(wd->iothread, INFINITE);
67897036 BC	132
	133	CloseHandle(wd->iothread);
	134	CloseHandle(wd->iocomplete_event);
	135
40c5db35	136	for (i = 0; i < td->o.iodepth; i++)
67897036	137	CloseHandle(wd->ovls[i].o.hEvent);
67897036 BC	138
	139	free(wd->aio_events);
	140	free(wd->ovls);
	141	free(wd);
	142
	143	td->io_ops->data = NULL;
	144	}
	145	}
	146
	147
ecc314ba BC	148	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	149	{
	150	int rc = 0;
	151	HANDLE hFile;
4e79098f	152	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
ecc314ba BC	153	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	154	DWORD openmode = OPEN_ALWAYS;
	155	DWORD access;
	156
	157	dprint(FD_FILE, "fd open %s\n", f->file_name);
	158
ecc314ba BC	159	if (f->filetype == FIO_TYPE_PIPE) {
	160	log_err("fio: windowsaio doesn't support pipes\n");
	161	return 1;
	162	}
	163
	164	if (!strcmp(f->file_name, "-")) {
	165	log_err("fio: can't read/write to stdin/out\n");
	166	return 1;
	167	}
	168
	169	if (td->o.odirect)
	170	flags \|= FILE_FLAG_NO_BUFFERING;
	171	if (td->o.sync_io)
	172	flags \|= FILE_FLAG_WRITE_THROUGH;
	173
ecc314ba	174	if (td->o.td_ddir == TD_DDIR_READ \|\|
ea4500d8	175	td->o.td_ddir == TD_DDIR_WRITE)
ecc314ba	176	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
ecc314ba	177	else
ecc314ba	178	flags \|= FILE_FLAG_RANDOM_ACCESS;
ecc314ba	179
ea4500d8	180	if (!td_write(td) \|\| read_only)
ecc314ba BC	181	access = GENERIC_READ;
	182	else
	183	access = (GENERIC_READ \| GENERIC_WRITE);
	184
	185	if (td->o.create_on_open > 0)
	186	openmode = OPEN_ALWAYS;
	187	else
	188	openmode = OPEN_EXISTING;
	189
	190	f->hFile = CreateFile(f->file_name, access, sharemode,
	191	NULL, openmode, flags, NULL);
	192
4e79098f	193	if (f->hFile == INVALID_HANDLE_VALUE)
ecc314ba	194	rc = 1;
ecc314ba BC	195
	196	/* Only set up the competion port and thread if we're not just
	197	* querying the device size */
40c5db35	198	if (!rc && td->io_ops->data != NULL) {
ecc314ba	199	struct thread_ctx *ctx;
40c5db35	200	struct windowsaio_data *wd;
ecc314ba	201
40c5db35	202	hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
ea4500d8	203
40c5db35	204	wd = td->io_ops->data;
9b836561	205	wd->iothread_running = TRUE;
ecc314ba	206
9b836561 BC	207	if (!rc) {
	208	ctx = malloc(sizeof(struct thread_ctx));
	209	ctx->iocp = hFile;
	210	ctx->wd = wd;
	211
	212	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	213	}
ecc314ba	214
4e79098f	215	if (rc \|\| wd->iothread == NULL)
ecc314ba	216	rc = 1;
ecc314ba BC	217	}
ecc314ba BC	218
ecc314ba BC	219	return rc;
	220	}
	221
	222	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	223	{
4e79098f	224	int rc = 0;
1e42cc3d	225
9b836561 BC	226	dprint(FD_FILE, "fd close %s\n", f->file_name);
9b836561 BC	227
ecc314ba	228	if (f->hFile != INVALID_HANDLE_VALUE) {
9b836561	229	if (!CloseHandle(f->hFile))
4e79098f	230	rc = 1;
ecc314ba BC	231	}
	232
	233	f->hFile = INVALID_HANDLE_VALUE;
4e79098f	234	return rc;
ecc314ba BC	235	}
ecc314ba BC	236
67897036 BC	237	static BOOL timeout_expired(DWORD start_count, DWORD end_count)
	238	{
	239	BOOL expired = FALSE;
	240	DWORD current_time;
	241
	242	current_time = GetTickCount();
	243
	244	if ((end_count > start_count) && current_time >= end_count)
	245	expired = TRUE;
	246	else if (current_time < start_count && current_time > end_count)
	247	expired = TRUE;
	248
	249	return expired;
	250	}
	251
	252	static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
	253	{
	254	struct windowsaio_data *wd = td->io_ops->data;
	255	return wd->aio_events[event];
	256	}
	257
	258	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	259	unsigned int max, struct timespec *t)
	260	{
	261	struct windowsaio_data *wd = td->io_ops->data;
	262	struct flist_head *entry;
	263	unsigned int dequeued = 0;
	264	struct io_u *io_u;
	265	struct fio_overlapped *fov;
	266	DWORD start_count = 0;
	267	DWORD end_count = 0;
	268	DWORD status;
	269	DWORD mswait = 250;
	270
	271	if (t != NULL) {
	272	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	273	start_count = GetTickCount();
	274	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	275	}
	276
	277	do {
	278	flist_for_each(entry, &td->io_u_busylist) {
	279	io_u = flist_entry(entry, struct io_u, list);
	280	fov = (struct fio_overlapped*)io_u->engine_data;
	281
	282	if (fov->io_complete) {
40c5db35	283	fov->io_complete = FALSE;
67897036 BC	284	fov->io_free = TRUE;
	285	wd->aio_events[dequeued] = io_u;
	286	dequeued++;
	287	}
	288
	289	if (dequeued >= min)
	290	break;
	291	}
	292
40c5db35	293	if (dequeued < min) {
67897036 BC	294	status = WaitForSingleObject(wd->iocomplete_event, mswait);
	295	if (status != WAIT_OBJECT_0 && dequeued > 0)
	296	break;
	297	}
	298
	299	if (dequeued >= min \|\| (t != NULL && timeout_expired(start_count, end_count)))
	300	break;
	301	} while (1);
	302
	303	return dequeued;
	304	}
	305
40c5db35	306	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
67897036	307	{
1e42cc3d	308	LPOVERLAPPED lpOvl = NULL;
67897036 BC	309	struct windowsaio_data *wd;
	310	DWORD iobytes;
	311	BOOL success;
	312	int index;
	313	int rc = FIO_Q_COMPLETED;
	314
	315	fio_ro_check(td, io_u);
	316
	317	wd = td->io_ops->data;
	318
1e42cc3d BC	319	for (index = 0; index < td->o.iodepth; index++) {
	320	if (wd->ovls[index].io_free) {
	321	wd->ovls[index].io_free = FALSE;
	322	ResetEvent(wd->ovls[index].o.hEvent);
	323	break;
	324	}
	325	}
67897036	326
1e42cc3d	327	assert(index < td->o.iodepth);
67897036	328
1e42cc3d BC	329	lpOvl = &wd->ovls[index].o;
1e42cc3d BC	330	wd->ovls[index].io_u = io_u;
4e79098f BC	331	lpOvl->Internal = STATUS_PENDING;
	332	lpOvl->InternalHigh = 0;
	333	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	334	lpOvl->OffsetHigh = io_u->offset >> 32;
1e42cc3d	335	io_u->engine_data = &wd->ovls[index];
67897036 BC	336
67897036 BC	337	switch (io_u->ddir) {
40c5db35	338	case DDIR_WRITE:
67897036 BC	339	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	340	break;
	341	case DDIR_READ:
	342	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	343	break;
	344	case DDIR_SYNC:
	345	case DDIR_DATASYNC:
	346	case DDIR_SYNC_FILE_RANGE:
	347	success = FlushFileBuffers(io_u->file->hFile);
	348	if (!success)
	349	io_u->error = GetLastError();
	350
	351	return FIO_Q_COMPLETED;
	352	break;
	353	case DDIR_TRIM:
	354	log_err("manual TRIM isn't supported on Windows");
	355	io_u->error = 1;
	356	io_u->resid = io_u->xfer_buflen;
	357	return FIO_Q_COMPLETED;
	358	break;
	359	default:
	360	assert(0);
	361	}
	362
40c5db35	363	if (success \|\| GetLastError() == ERROR_IO_PENDING)
67897036	364	rc = FIO_Q_QUEUED;
40c5db35	365	else {
67897036 BC	366	io_u->error = GetLastError();
	367	io_u->resid = io_u->xfer_buflen;
	368	}
	369
	370	return rc;
	371	}
	372
	373	/* Runs as a thread and waits for queued IO to complete */
	374	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	375	{
	376	OVERLAPPED *ovl;
	377	struct fio_overlapped *fov;
	378	struct io_u *io_u;
	379	struct windowsaio_data *wd;
	380	struct thread_ctx *ctx;
	381	ULONG_PTR ulKey = 0;
	382	DWORD bytes;
	383
	384	ctx = (struct thread_ctx*)lpParameter;
	385	wd = ctx->wd;
	386
	387	do {
	388	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250))
	389	continue;
	390
	391	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	392	io_u = fov->io_u;
	393
	394	if (ovl->Internal == ERROR_SUCCESS) {
	395	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	396	io_u->error = 0;
	397	} else {
	398	io_u->resid = io_u->xfer_buflen;
	399	io_u->error = ovl->Internal;
	400	}
	401
40c5db35	402	fov->io_complete = TRUE;
67897036 BC	403	SetEvent(wd->iocomplete_event);
	404	} while (ctx->wd->iothread_running);
	405
	406	CloseHandle(ctx->iocp);
	407	free(ctx);
	408	return 0;
	409	}
	410
	411	static int fio_windowsaio_cancel(struct thread_data *td,
	412	struct io_u *io_u)
	413	{
	414	int rc = 0;
	415
	416	struct windowsaio_data *wd = td->io_ops->data;
	417
	418	/* If we're running on Vista or newer, we can cancel individual IO requests */
	419	if (wd->pCancelIoEx != NULL) {
	420	struct fio_overlapped *ovl = io_u->engine_data;
40c5db35	421
67897036 BC	422	if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
	423	rc = 1;
	424	} else
	425	rc = 1;
	426
	427	return rc;
	428	}
	429
ecc314ba BC	430	static struct ioengine_ops ioengine = {
	431	.name = "windowsaio",
	432	.version = FIO_IOOPS_VERSION,
	433	.init = fio_windowsaio_init,
	434	.queue = fio_windowsaio_queue,
	435	.cancel = fio_windowsaio_cancel,
	436	.getevents = fio_windowsaio_getevents,
	437	.event = fio_windowsaio_event,
	438	.cleanup = fio_windowsaio_cleanup,
	439	.open_file = fio_windowsaio_open_file,
	440	.close_file = fio_windowsaio_close_file,
	441	.get_file_size = generic_get_file_size
	442	};
	443
	444	static void fio_init fio_posixaio_register(void)
	445	{
	446	register_ioengine(&ioengine);
	447	}
	448
	449	static void fio_exit fio_posixaio_unregister(void)
	450	{
	451	unregister_ioengine(&ioengine);
	452	}