[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)
		{
			ctx->iocp = hFile;
			ctx->wd = wd;
			wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
			if (wd->iothread == NULL)
				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;
				ResetEvent(fov->o.hEvent);
				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;
	DWORD iobytes;
	BOOL success = FALSE;
	int rc = FIO_Q_COMPLETED;

	fio_ro_check(td, io_u);

	lpOvl->Internal = STATUS_PENDING;
	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, &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) {
			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) {
		CloseHandle(o->o.hEvent);
		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 = CreateEvent(NULL, TRUE, FALSE, NULL);
	if (o->o.hEvent == NULL) {
		log_err("windowsaio: failed to create event handle\n");
		free(o);
		return 1;
	}

	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	ctx->iocp = hFile;
	117	ctx->wd = wd;
	118	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	119	if (wd->iothread == NULL)
	120	log_err("windowsaio: failed to create io completion thread\n");
	121	}
	122
	123	if (rc \|\| wd->iothread == NULL)
	124	rc = 1;
	125	}
	126
	127	return rc;
	128	}
	129
	130	static void fio_windowsaio_cleanup(struct thread_data *td)
	131	{
	132	struct windowsaio_data *wd;
	133
	134	wd = td->io_ops->data;
	135
	136	if (wd != NULL) {
	137	wd->iothread_running = FALSE;
	138	WaitForSingleObject(wd->iothread, INFINITE);
	139
	140	CloseHandle(wd->iothread);
	141	CloseHandle(wd->iocomplete_event);
	142
	143	free(wd->aio_events);
	144	free(wd);
	145
	146	td->io_ops->data = NULL;
	147	}
	148	}
	149
	150
	151	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	152	{
	153	int rc = 0;
	154	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
	155	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	156	DWORD openmode = OPEN_ALWAYS;
	157	DWORD access;
	158
	159	dprint(FD_FILE, "fd open %s\n", f->file_name);
	160
	161	if (f->filetype == FIO_TYPE_PIPE) {
	162	log_err("windowsaio: pipes are not supported\n");
	163	return 1;
	164	}
	165
	166	if (!strcmp(f->file_name, "-")) {
	167	log_err("windowsaio: can't read/write to stdin/out\n");
	168	return 1;
	169	}
	170
	171	if (td->o.odirect)
	172	flags \|= FILE_FLAG_NO_BUFFERING;
	173	if (td->o.sync_io)
	174	flags \|= FILE_FLAG_WRITE_THROUGH;
	175
	176	/*
	177	* Inform Windows whether we're going to be doing sequential or
	178	* random io so it can tune the Cache Manager
	179	*/
	180	if (td->o.td_ddir == TD_DDIR_READ \|\|
	181	td->o.td_ddir == TD_DDIR_WRITE)
	182	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
	183	else
	184	flags \|= FILE_FLAG_RANDOM_ACCESS;
	185
	186	if (!td_write(td) \|\| read_only)
	187	access = GENERIC_READ;
	188	else
	189	access = (GENERIC_READ \| GENERIC_WRITE);
	190
	191	if (td->o.create_on_open)
	192	openmode = OPEN_ALWAYS;
	193	else
	194	openmode = OPEN_EXISTING;
	195
	196	f->hFile = CreateFile(f->file_name, access, sharemode,
	197	NULL, openmode, flags, NULL);
	198
	199	if (f->hFile == INVALID_HANDLE_VALUE) {
	200	log_err("windowsaio: failed to open file \"%s\"\n", f->file_name);
	201	rc = 1;
	202	}
	203
	204	/* Only set up the completion port and thread if we're not just
	205	* querying the device size */
	206	if (!rc && td->io_ops->data != NULL) {
	207	struct windowsaio_data *wd;
	208
	209	wd = td->io_ops->data;
	210
	211	if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL) {
	212	log_err("windowsaio: failed to create io completion port\n");
	213	rc = 1;
	214	}
	215	}
	216
	217	return rc;
	218	}
	219
	220	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	221	{
	222	int rc = 0;
	223
	224	dprint(FD_FILE, "fd close %s\n", f->file_name);
	225
	226	if (f->hFile != INVALID_HANDLE_VALUE) {
	227	if (!CloseHandle(f->hFile)) {
	228	log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name);
	229	rc = 1;
	230	}
	231	}
	232
	233	f->hFile = INVALID_HANDLE_VALUE;
	234	return rc;
	235	}
	236
	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,
	260	const struct timespec *t)
	261	{
	262	struct windowsaio_data *wd = td->io_ops->data;
	263	unsigned int dequeued = 0;
	264	struct io_u *io_u;
	265	int i;
	266	struct fio_overlapped *fov;
	267	DWORD start_count = 0;
	268	DWORD end_count = 0;
	269	DWORD status;
	270	DWORD mswait = 250;
	271
	272	if (t != NULL) {
	273	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	274	start_count = GetTickCount();
	275	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	276	}
	277
	278	do {
	279	io_u_qiter(&td->io_u_all, io_u, i) {
	280	if (!(io_u->flags & IO_U_F_FLIGHT))
	281	continue;
	282
	283	fov = (struct fio_overlapped*)io_u->engine_data;
	284
	285	if (fov->io_complete) {
	286	fov->io_complete = FALSE;
	287	ResetEvent(fov->o.hEvent);
	288	wd->aio_events[dequeued] = io_u;
	289	dequeued++;
	290	}
	291
	292	if (dequeued >= min)
	293	break;
	294	}
	295
	296	if (dequeued < min) {
	297	status = WaitForSingleObject(wd->iocomplete_event, mswait);
	298	if (status != WAIT_OBJECT_0 && dequeued >= min)
	299	break;
	300	}
	301
	302	if (dequeued >= min \|\| (t != NULL && timeout_expired(start_count, end_count)))
	303	break;
	304	} while (1);
	305
	306	return dequeued;
	307	}
	308
	309	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
	310	{
	311	struct fio_overlapped *o = io_u->engine_data;
	312	LPOVERLAPPED lpOvl = &o->o;
	313	DWORD iobytes;
	314	BOOL success = FALSE;
	315	int rc = FIO_Q_COMPLETED;
	316
	317	fio_ro_check(td, io_u);
	318
	319	lpOvl->Internal = STATUS_PENDING;
	320	lpOvl->InternalHigh = 0;
	321	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	322	lpOvl->OffsetHigh = io_u->offset >> 32;
	323
	324	switch (io_u->ddir) {
	325	case DDIR_WRITE:
	326	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	327	break;
	328	case DDIR_READ:
	329	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	330	break;
	331	case DDIR_SYNC:
	332	case DDIR_DATASYNC:
	333	case DDIR_SYNC_FILE_RANGE:
	334	success = FlushFileBuffers(io_u->file->hFile);
	335	if (!success) {
	336	log_err("windowsaio: failed to flush file buffers\n");
	337	io_u->error = win_to_posix_error(GetLastError());
	338	}
	339
	340	return FIO_Q_COMPLETED;
	341	break;
	342	case DDIR_TRIM:
	343	log_err("windowsaio: manual TRIM isn't supported on Windows\n");
	344	io_u->error = 1;
	345	io_u->resid = io_u->xfer_buflen;
	346	return FIO_Q_COMPLETED;
	347	break;
	348	default:
	349	assert(0);
	350	break;
	351	}
	352
	353	if (success \|\| GetLastError() == ERROR_IO_PENDING)
	354	rc = FIO_Q_QUEUED;
	355	else {
	356	io_u->error = win_to_posix_error(GetLastError());
	357	io_u->resid = io_u->xfer_buflen;
	358	}
	359
	360	return rc;
	361	}
	362
	363	/* Runs as a thread and waits for queued IO to complete */
	364	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	365	{
	366	OVERLAPPED *ovl;
	367	struct fio_overlapped *fov;
	368	struct io_u *io_u;
	369	struct windowsaio_data *wd;
	370	struct thread_ctx *ctx;
	371	ULONG_PTR ulKey = 0;
	372	DWORD bytes;
	373
	374	ctx = (struct thread_ctx*)lpParameter;
	375	wd = ctx->wd;
	376
	377	do {
	378	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
	379	continue;
	380
	381	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	382	io_u = fov->io_u;
	383
	384	if (ovl->Internal == ERROR_SUCCESS) {
	385	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	386	io_u->error = 0;
	387	} else {
	388	io_u->resid = io_u->xfer_buflen;
	389	io_u->error = win_to_posix_error(GetLastError());
	390	}
	391
	392	fov->io_complete = TRUE;
	393	SetEvent(wd->iocomplete_event);
	394	} while (ctx->wd->iothread_running);
	395
	396	CloseHandle(ctx->iocp);
	397	free(ctx);
	398	return 0;
	399	}
	400
	401	static void fio_windowsaio_io_u_free(struct thread_data td, struct io_u io_u)
	402	{
	403	struct fio_overlapped *o = io_u->engine_data;
	404
	405	if (o) {
	406	CloseHandle(o->o.hEvent);
	407	io_u->engine_data = NULL;
	408	free(o);
	409	}
	410	}
	411
	412	static int fio_windowsaio_io_u_init(struct thread_data td, struct io_u io_u)
	413	{
	414	struct fio_overlapped *o;
	415
	416	o = malloc(sizeof(*o));
	417	o->io_complete = FALSE;
	418	o->io_u = io_u;
	419	o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
	420	if (o->o.hEvent == NULL) {
	421	log_err("windowsaio: failed to create event handle\n");
	422	free(o);
	423	return 1;
	424	}
	425
	426	io_u->engine_data = o;
	427	return 0;
	428	}
	429
	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	.getevents = fio_windowsaio_getevents,
	436	.event = fio_windowsaio_event,
	437	.cleanup = fio_windowsaio_cleanup,
	438	.open_file = fio_windowsaio_open_file,
	439	.close_file = fio_windowsaio_close_file,
	440	.get_file_size = generic_get_file_size,
	441	.io_u_init = fio_windowsaio_io_u_init,
	442	.io_u_free = fio_windowsaio_io_u_free,
	443	};
	444
	445	static void fio_init fio_windowsaio_register(void)
	446	{
	447	register_ioengine(&ioengine);
	448	}
	449
	450	static void fio_exit fio_windowsaio_unregister(void)
	451	{
	452	unregister_ioengine(&ioengine);
	453	}