[fio.git] / engines / windowsaio.c

/*
 * windowsaio engine
 *
 * IO engine using Windows IO Completion Ports.
 */

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

#include "../fio.h"

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

int geterrno_from_win_error (DWORD code, int deferrno);

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

struct windowsaio_data {
	struct io_u **aio_events;
	HANDLE iocp;
	HANDLE iothread;
	HANDLE iocomplete_event;
	BOOL iothread_running;
};

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

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, const 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);

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

	wd = calloc(1, sizeof(struct windowsaio_data));
	if (wd == NULL) {
		 log_err("windowsaio: failed to allocate memory for engine data\n");
		rc = 1;
	}

	if (!rc) {
		wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
		if (wd->aio_events == NULL) {
			log_err("windowsaio: failed to allocate memory for aio events list\n");
			rc = 1;
		}
	}

	if (!rc) {
		/* Create an auto-reset event */
		wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
		if (wd->iocomplete_event == NULL) {
			log_err("windowsaio: failed to create io complete event handle\n");
			rc = 1;
		}
	}

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

			free(wd);
		}
	}

	td->io_ops_data = wd;

	if (!rc) {
		struct thread_ctx *ctx;
		struct windowsaio_data *wd;
		HANDLE hFile;

		hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0);
		if (hFile == INVALID_HANDLE_VALUE) {
			log_err("windowsaio: failed to create io completion port\n");
			rc = 1;
		}

		wd = td->io_ops_data;
		wd->iothread_running = TRUE;
		wd->iocp = hFile;

		if (!rc)
			ctx = malloc(sizeof(struct thread_ctx));

		if (!rc && ctx == NULL)
		{
			log_err("windowsaio: failed to allocate memory for thread context structure\n");
			CloseHandle(hFile);
			rc = 1;
		}

		if (!rc)
		{
			DWORD threadid;

			ctx->iocp = hFile;
			ctx->wd = wd;
			wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, &threadid);

			if (wd->iothread != NULL)
				fio_setaffinity(threadid, td->o.cpumask);
			else
				log_err("windowsaio: failed to create io completion thread\n");
		}

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

	return rc;
}

static void fio_windowsaio_cleanup(struct thread_data *td)
{
	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);

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

		td->io_ops_data = NULL;
	}
}


static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
{
	int rc = 0;
	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("windowsaio: pipes are not supported\n");
		return 1;
	}

	if (!strcmp(f->file_name, "-")) {
		log_err("windowsaio: 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;

	/*
	 * Inform Windows whether we're going to be doing sequential or
	 * random io so it can tune the Cache Manager
	 */
	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)
		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) {
		log_err("windowsaio: failed to open file \"%s\"\n", f->file_name);
		rc = 1;
	}

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

		wd = td->io_ops_data;

		if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL) {
			log_err("windowsaio: failed to create io completion port\n");
			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)) {
			log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name);
			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,
				    const struct timespec *t)
{
	struct windowsaio_data *wd = td->io_ops_data;
	unsigned int dequeued = 0;
	struct io_u *io_u;
	int i;
	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 {
		io_u_qiter(&td->io_u_all, io_u, i) {
			if (!(io_u->flags & IO_U_F_FLIGHT))
				continue;

			fov = (struct fio_overlapped*)io_u->engine_data;

			if (fov->io_complete) {
				fov->io_complete = FALSE;
				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 >= min)
				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)
{
	struct fio_overlapped *o = io_u->engine_data;
	LPOVERLAPPED lpOvl = &o->o;
	BOOL success = FALSE;
	int rc = FIO_Q_COMPLETED;

	fio_ro_check(td, io_u);

	lpOvl->Internal = 0;
	lpOvl->InternalHigh = 0;
	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	lpOvl->OffsetHigh = io_u->offset >> 32;

	switch (io_u->ddir) {
	case DDIR_WRITE:
		success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, NULL, lpOvl);
		break;
	case DDIR_READ:
		success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, NULL, lpOvl);
		break;
	case DDIR_SYNC:
	case DDIR_DATASYNC:
	case DDIR_SYNC_FILE_RANGE:
		success = FlushFileBuffers(io_u->file->hFile);
		if (!success) {
			log_err("windowsaio: failed to flush file buffers\n");
			io_u->error = win_to_posix_error(GetLastError());
		}

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

	if (success || GetLastError() == ERROR_IO_PENDING)
		rc = FIO_Q_QUEUED;
	else {
		io_u->error = win_to_posix_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) && ovl == NULL)
			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 = win_to_posix_error(GetLastError());
		}

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

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

static void fio_windowsaio_io_u_free(struct thread_data *td, struct io_u *io_u)
{
	struct fio_overlapped *o = io_u->engine_data;

	if (o) {
		io_u->engine_data = NULL;
		free(o);
	}
}

static int fio_windowsaio_io_u_init(struct thread_data *td, struct io_u *io_u)
{
	struct fio_overlapped *o;

	o = malloc(sizeof(*o));
	o->io_complete = FALSE;
	o->io_u = io_u;
	o->o.hEvent = NULL;
	io_u->engine_data = o;
	return 0;
}

static struct ioengine_ops ioengine = {
	.name		= "windowsaio",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_windowsaio_init,
	.queue		= fio_windowsaio_queue,
	.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,
	.io_u_init	= fio_windowsaio_io_u_init,
	.io_u_free	= fio_windowsaio_io_u_free,
};

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

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