[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 DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);

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 windowsaio_invalidate_cache(struct fio_file *f)
{
	DWORD error;
	DWORD isharemode = (FILE_SHARE_DELETE | FILE_SHARE_READ |
				FILE_SHARE_WRITE);
	HANDLE ihFile;
	int rc = 0;

	/*
	 * Encourage Windows to drop cached parts of a file by temporarily
	 * opening it for non-buffered access. Note: this will only work when
	 * the following is the only thing with the file open on the whole
	 * system.
	 */
	dprint(FD_IO, "windowaio: attempt invalidate cache for %s\n",
			f->file_name);
	ihFile = CreateFile(f->file_name, 0, isharemode, NULL, OPEN_EXISTING,
			FILE_FLAG_NO_BUFFERING, NULL);

	if (ihFile != INVALID_HANDLE_VALUE) {
		if (!CloseHandle(ihFile)) {
			error = GetLastError();
			log_info("windowsaio: invalidation fd close %s "
				 "failed: error %d\n", f->file_name, error);
			rc = 1;
		}
	} else {
		error = GetLastError();
		if (error != ERROR_FILE_NOT_FOUND) {
			log_info("windowsaio: cache invalidation of %s failed: "
					"error %d\n", f->file_name, error);
			rc = 1;
		}
	}

	return rc;
}

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
	 */
	switch (td->o.fadvise_hint) {
	case F_ADV_TYPE:
		if (td_random(td))
			flags |= FILE_FLAG_RANDOM_ACCESS;
		else
			flags |= FILE_FLAG_SEQUENTIAL_SCAN;
		break;
	case F_ADV_RANDOM:
		flags |= FILE_FLAG_RANDOM_ACCESS;
		break;
	case F_ADV_SEQUENTIAL:
		flags |= FILE_FLAG_SEQUENTIAL_SCAN;
		break;
	case F_ADV_NONE:
		break;
	default:
		log_err("fio: unknown fadvise type %d\n", td->o.fadvise_hint);
	}

	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;

	/* If we're going to use direct I/O, Windows will try and invalidate
	 * its cache at that point so there's no need to do it here */
	if (td->o.invalidate_cache && !td->o.odirect)
		windowsaio_invalidate_cache(f);

	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;
	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;
	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 {
		BOOL ret;

		ret = GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey,
						&ovl, 250);
		if (!ret && 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 DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
	39
	40	static int fio_windowsaio_init(struct thread_data *td)
	41	{
	42	struct windowsaio_data *wd;
	43	int rc = 0;
	44
	45	wd = calloc(1, sizeof(struct windowsaio_data));
	46	if (wd == NULL) {
	47	log_err("windowsaio: failed to allocate memory for engine data\n");
	48	rc = 1;
	49	}
	50
	51	if (!rc) {
	52	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	53	if (wd->aio_events == NULL) {
	54	log_err("windowsaio: failed to allocate memory for aio events list\n");
	55	rc = 1;
	56	}
	57	}
	58
	59	if (!rc) {
	60	/* Create an auto-reset event */
	61	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	62	if (wd->iocomplete_event == NULL) {
	63	log_err("windowsaio: failed to create io complete event handle\n");
	64	rc = 1;
	65	}
	66	}
	67
	68	if (rc) {
	69	if (wd != NULL) {
	70	if (wd->aio_events != NULL)
	71	free(wd->aio_events);
	72
	73	free(wd);
	74	}
	75	}
	76
	77	td->io_ops_data = wd;
	78
	79	if (!rc) {
	80	struct thread_ctx *ctx;
	81	struct windowsaio_data *wd;
	82	HANDLE hFile;
	83
	84	hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0);
	85	if (hFile == INVALID_HANDLE_VALUE) {
	86	log_err("windowsaio: failed to create io completion port\n");
	87	rc = 1;
	88	}
	89
	90	wd = td->io_ops_data;
	91	wd->iothread_running = TRUE;
	92	wd->iocp = hFile;
	93
	94	if (!rc)
	95	ctx = malloc(sizeof(struct thread_ctx));
	96
	97	if (!rc && ctx == NULL) {
	98	log_err("windowsaio: failed to allocate memory for thread context structure\n");
	99	CloseHandle(hFile);
	100	rc = 1;
	101	}
	102
	103	if (!rc) {
	104	DWORD threadid;
	105
	106	ctx->iocp = hFile;
	107	ctx->wd = wd;
	108	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, &threadid);
	109
	110	if (wd->iothread != NULL)
	111	fio_setaffinity(threadid, td->o.cpumask);
	112	else
	113	log_err("windowsaio: failed to create io completion thread\n");
	114	}
	115
	116	if (rc \|\| wd->iothread == NULL)
	117	rc = 1;
	118	}
	119
	120	return rc;
	121	}
	122
	123	static void fio_windowsaio_cleanup(struct thread_data *td)
	124	{
	125	struct windowsaio_data *wd;
	126
	127	wd = td->io_ops_data;
	128
	129	if (wd != NULL) {
	130	wd->iothread_running = FALSE;
	131	WaitForSingleObject(wd->iothread, INFINITE);
	132
	133	CloseHandle(wd->iothread);
	134	CloseHandle(wd->iocomplete_event);
	135
	136	free(wd->aio_events);
	137	free(wd);
	138
	139	td->io_ops_data = NULL;
	140	}
	141	}
	142
	143	static int windowsaio_invalidate_cache(struct fio_file *f)
	144	{
	145	DWORD error;
	146	DWORD isharemode = (FILE_SHARE_DELETE \| FILE_SHARE_READ \|
	147	FILE_SHARE_WRITE);
	148	HANDLE ihFile;
	149	int rc = 0;
	150
	151	/*
	152	* Encourage Windows to drop cached parts of a file by temporarily
	153	* opening it for non-buffered access. Note: this will only work when
	154	* the following is the only thing with the file open on the whole
	155	* system.
	156	*/
	157	dprint(FD_IO, "windowaio: attempt invalidate cache for %s\n",
	158	f->file_name);
	159	ihFile = CreateFile(f->file_name, 0, isharemode, NULL, OPEN_EXISTING,
	160	FILE_FLAG_NO_BUFFERING, NULL);
	161
	162	if (ihFile != INVALID_HANDLE_VALUE) {
	163	if (!CloseHandle(ihFile)) {
	164	error = GetLastError();
	165	log_info("windowsaio: invalidation fd close %s "
	166	"failed: error %d\n", f->file_name, error);
	167	rc = 1;
	168	}
	169	} else {
	170	error = GetLastError();
	171	if (error != ERROR_FILE_NOT_FOUND) {
	172	log_info("windowsaio: cache invalidation of %s failed: "
	173	"error %d\n", f->file_name, error);
	174	rc = 1;
	175	}
	176	}
	177
	178	return rc;
	179	}
	180
	181	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	182	{
	183	int rc = 0;
	184	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
	185	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	186	DWORD openmode = OPEN_ALWAYS;
	187	DWORD access;
	188
	189	dprint(FD_FILE, "fd open %s\n", f->file_name);
	190
	191	if (f->filetype == FIO_TYPE_PIPE) {
	192	log_err("windowsaio: pipes are not supported\n");
	193	return 1;
	194	}
	195
	196	if (!strcmp(f->file_name, "-")) {
	197	log_err("windowsaio: can't read/write to stdin/out\n");
	198	return 1;
	199	}
	200
	201	if (td->o.odirect)
	202	flags \|= FILE_FLAG_NO_BUFFERING;
	203	if (td->o.sync_io)
	204	flags \|= FILE_FLAG_WRITE_THROUGH;
	205
	206	/*
	207	* Inform Windows whether we're going to be doing sequential or
	208	* random IO so it can tune the Cache Manager
	209	*/
	210	switch (td->o.fadvise_hint) {
	211	case F_ADV_TYPE:
	212	if (td_random(td))
	213	flags \|= FILE_FLAG_RANDOM_ACCESS;
	214	else
	215	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
	216	break;
	217	case F_ADV_RANDOM:
	218	flags \|= FILE_FLAG_RANDOM_ACCESS;
	219	break;
	220	case F_ADV_SEQUENTIAL:
	221	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
	222	break;
	223	case F_ADV_NONE:
	224	break;
	225	default:
	226	log_err("fio: unknown fadvise type %d\n", td->o.fadvise_hint);
	227	}
	228
	229	if (!td_write(td) \|\| read_only)
	230	access = GENERIC_READ;
	231	else
	232	access = (GENERIC_READ \| GENERIC_WRITE);
	233
	234	if (td->o.create_on_open)
	235	openmode = OPEN_ALWAYS;
	236	else
	237	openmode = OPEN_EXISTING;
	238
	239	/* If we're going to use direct I/O, Windows will try and invalidate
	240	* its cache at that point so there's no need to do it here */
	241	if (td->o.invalidate_cache && !td->o.odirect)
	242	windowsaio_invalidate_cache(f);
	243
	244	f->hFile = CreateFile(f->file_name, access, sharemode,
	245	NULL, openmode, flags, NULL);
	246
	247	if (f->hFile == INVALID_HANDLE_VALUE) {
	248	log_err("windowsaio: failed to open file \"%s\"\n", f->file_name);
	249	rc = 1;
	250	}
	251
	252	/* Only set up the completion port and thread if we're not just
	253	* querying the device size */
	254	if (!rc && td->io_ops_data != NULL) {
	255	struct windowsaio_data *wd;
	256
	257	wd = td->io_ops_data;
	258
	259	if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL) {
	260	log_err("windowsaio: failed to create io completion port\n");
	261	rc = 1;
	262	}
	263	}
	264
	265	return rc;
	266	}
	267
	268	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	269	{
	270	int rc = 0;
	271
	272	dprint(FD_FILE, "fd close %s\n", f->file_name);
	273
	274	if (f->hFile != INVALID_HANDLE_VALUE) {
	275	if (!CloseHandle(f->hFile)) {
	276	log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name);
	277	rc = 1;
	278	}
	279	}
	280
	281	f->hFile = INVALID_HANDLE_VALUE;
	282	return rc;
	283	}
	284
	285	static BOOL timeout_expired(DWORD start_count, DWORD end_count)
	286	{
	287	BOOL expired = FALSE;
	288	DWORD current_time;
	289
	290	current_time = GetTickCount();
	291
	292	if ((end_count > start_count) && current_time >= end_count)
	293	expired = TRUE;
	294	else if (current_time < start_count && current_time > end_count)
	295	expired = TRUE;
	296
	297	return expired;
	298	}
	299
	300	static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
	301	{
	302	struct windowsaio_data *wd = td->io_ops_data;
	303	return wd->aio_events[event];
	304	}
	305
	306	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	307	unsigned int max,
	308	const struct timespec *t)
	309	{
	310	struct windowsaio_data *wd = td->io_ops_data;
	311	unsigned int dequeued = 0;
	312	struct io_u *io_u;
	313	int i;
	314	struct fio_overlapped *fov;
	315	DWORD start_count = 0;
	316	DWORD end_count = 0;
	317	DWORD status;
	318	DWORD mswait = 250;
	319
	320	if (t != NULL) {
	321	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	322	start_count = GetTickCount();
	323	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	324	}
	325
	326	do {
	327	io_u_qiter(&td->io_u_all, io_u, i) {
	328	if (!(io_u->flags & IO_U_F_FLIGHT))
	329	continue;
	330
	331	fov = (struct fio_overlapped*)io_u->engine_data;
	332
	333	if (fov->io_complete) {
	334	fov->io_complete = FALSE;
	335	wd->aio_events[dequeued] = io_u;
	336	dequeued++;
	337	}
	338
	339	}
	340	if (dequeued >= min)
	341	break;
	342
	343	if (dequeued < min) {
	344	status = WaitForSingleObject(wd->iocomplete_event, mswait);
	345	if (status != WAIT_OBJECT_0 && dequeued >= min)
	346	break;
	347	}
	348
	349	if (dequeued >= min \|\|
	350	(t != NULL && timeout_expired(start_count, end_count)))
	351	break;
	352	} while (1);
	353
	354	return dequeued;
	355	}
	356
	357	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
	358	{
	359	struct fio_overlapped *o = io_u->engine_data;
	360	LPOVERLAPPED lpOvl = &o->o;
	361	BOOL success = FALSE;
	362	int rc = FIO_Q_COMPLETED;
	363
	364	fio_ro_check(td, io_u);
	365
	366	lpOvl->Internal = 0;
	367	lpOvl->InternalHigh = 0;
	368	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	369	lpOvl->OffsetHigh = io_u->offset >> 32;
	370
	371	switch (io_u->ddir) {
	372	case DDIR_WRITE:
	373	success = WriteFile(io_u->file->hFile, io_u->xfer_buf,
	374	io_u->xfer_buflen, NULL, lpOvl);
	375	break;
	376	case DDIR_READ:
	377	success = ReadFile(io_u->file->hFile, io_u->xfer_buf,
	378	io_u->xfer_buflen, NULL, lpOvl);
	379	break;
	380	case DDIR_SYNC:
	381	case DDIR_DATASYNC:
	382	case DDIR_SYNC_FILE_RANGE:
	383	success = FlushFileBuffers(io_u->file->hFile);
	384	if (!success) {
	385	log_err("windowsaio: failed to flush file buffers\n");
	386	io_u->error = win_to_posix_error(GetLastError());
	387	}
	388
	389	return FIO_Q_COMPLETED;
	390	case DDIR_TRIM:
	391	log_err("windowsaio: manual TRIM isn't supported on Windows\n");
	392	io_u->error = 1;
	393	io_u->resid = io_u->xfer_buflen;
	394	return FIO_Q_COMPLETED;
	395	default:
	396	assert(0);
	397	break;
	398	}
	399
	400	if (success \|\| GetLastError() == ERROR_IO_PENDING)
	401	rc = FIO_Q_QUEUED;
	402	else {
	403	io_u->error = win_to_posix_error(GetLastError());
	404	io_u->resid = io_u->xfer_buflen;
	405	}
	406
	407	return rc;
	408	}
	409
	410	/* Runs as a thread and waits for queued IO to complete */
	411	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	412	{
	413	OVERLAPPED *ovl;
	414	struct fio_overlapped *fov;
	415	struct io_u *io_u;
	416	struct windowsaio_data *wd;
	417	struct thread_ctx *ctx;
	418	ULONG_PTR ulKey = 0;
	419	DWORD bytes;
	420
	421	ctx = (struct thread_ctx*)lpParameter;
	422	wd = ctx->wd;
	423
	424	do {
	425	BOOL ret;
	426
	427	ret = GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey,
	428	&ovl, 250);
	429	if (!ret && ovl == NULL)
	430	continue;
	431
	432	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	433	io_u = fov->io_u;
	434
	435	if (ovl->Internal == ERROR_SUCCESS) {
	436	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	437	io_u->error = 0;
	438	} else {
	439	io_u->resid = io_u->xfer_buflen;
	440	io_u->error = win_to_posix_error(GetLastError());
	441	}
	442
	443	fov->io_complete = TRUE;
	444	SetEvent(wd->iocomplete_event);
	445	} while (ctx->wd->iothread_running);
	446
	447	CloseHandle(ctx->iocp);
	448	free(ctx);
	449	return 0;
	450	}
	451
	452	static void fio_windowsaio_io_u_free(struct thread_data td, struct io_u io_u)
	453	{
	454	struct fio_overlapped *o = io_u->engine_data;
	455
	456	if (o) {
	457	io_u->engine_data = NULL;
	458	free(o);
	459	}
	460	}
	461
	462	static int fio_windowsaio_io_u_init(struct thread_data td, struct io_u io_u)
	463	{
	464	struct fio_overlapped *o;
	465
	466	o = malloc(sizeof(*o));
	467	o->io_complete = FALSE;
	468	o->io_u = io_u;
	469	o->o.hEvent = NULL;
	470	io_u->engine_data = o;
	471	return 0;
	472	}
	473
	474	static struct ioengine_ops ioengine = {
	475	.name = "windowsaio",
	476	.version = FIO_IOOPS_VERSION,
	477	.init = fio_windowsaio_init,
	478	.queue = fio_windowsaio_queue,
	479	.getevents = fio_windowsaio_getevents,
	480	.event = fio_windowsaio_event,
	481	.cleanup = fio_windowsaio_cleanup,
	482	.open_file = fio_windowsaio_open_file,
	483	.close_file = fio_windowsaio_close_file,
	484	.get_file_size = generic_get_file_size,
	485	.io_u_init = fio_windowsaio_io_u_init,
	486	.io_u_free = fio_windowsaio_io_u_free,
	487	};
	488
	489	static void fio_init fio_windowsaio_register(void)
	490	{
	491	register_ioengine(&ioengine);
	492	}
	493
	494	static void fio_exit fio_windowsaio_unregister(void)
	495	{
	496	unregister_ioengine(&ioengine);
	497	}