[fio.git] / engines / windowsaio.c

/*
 * Native Windows async IO engine
 * Copyright (C) 2010 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 struct {
 OVERLAPPED o;
 struct io_u *io_u;
} FIO_OVERLAPPED;

struct windowsaio_data {
	HANDLE *io_handles;
	unsigned int io_index;
	FIO_OVERLAPPED *ovls;

	HANDLE iothread;
	HANDLE iothread_stopped;
	BOOL iothread_running;

	struct io_u **aio_events;
	HANDLE iocomplete_event;
	BOOL have_cancelioex;
};

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

static void PrintError(LPCSTR lpszFunction);
static int fio_windowsaio_cancel(struct thread_data *td,
			       struct io_u *io_u);
static BOOL TimedOut(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);

/* CancelIoEx isn't in Cygwin's w32api */
BOOL WINAPI CancelIoEx(
  HANDLE hFile,
  LPOVERLAPPED lpOverlapped
);


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

static void PrintError(LPCSTR lpszFunction)
{
	// Retrieve the system error message for the last-error code

	LPSTR lpMsgBuf;
	DWORD dw = GetLastError();

	FormatMessage(
		FORMAT_MESSAGE_ALLOCATE_BUFFER |
		FORMAT_MESSAGE_FROM_SYSTEM |
		FORMAT_MESSAGE_IGNORE_INSERTS,
		NULL,
		dw,
		MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
		(LPTSTR)&lpMsgBuf,
		0, NULL );

	log_err("%s - %s", lpszFunction, lpMsgBuf);
	LocalFree(lpMsgBuf);
}

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, we can cancel individual IO requests */
	if (wd->have_cancelioex) {
		FIO_OVERLAPPED *ovl = io_u->engine_data;
		if (!CancelIoEx(io_u->file->hFile, &ovl->o))
			rc = 1;
	} else
		rc = 1;

	return rc;
}

static BOOL TimedOut(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 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;
	DWORD start_count = 0, end_count = 0;
	BOOL timedout = FALSE;
	unsigned int mswait = 100;

	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);
	}

	while (dequeued < min && !timedout) {
		flist_for_each(entry, &td->io_u_busylist) {
			io_u = flist_entry(entry, struct io_u, list);

			if (io_u->seen == 1) {
				io_u->seen = 2;
				wd->aio_events[dequeued] = io_u;
				dequeued++;
			}

			if (dequeued == max)
				break;
		}

		if (dequeued < min) {
			WaitForSingleObject(wd->iocomplete_event, mswait);
		}

		if (t != NULL && TimedOut(start_count, end_count))
			timedout = TRUE;
	}

	return dequeued;
}

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_queue(struct thread_data *td,
			      struct io_u *io_u)
{
	struct windowsaio_data *wd;
	DWORD iobytes;
	BOOL success = TRUE;
	int ind;
	int rc;

	fio_ro_check(td, io_u);

	wd = td->io_ops->data;
	ind = wd->io_index;

	ResetEvent(wd->io_handles[ind]);
	wd->ovls[ind].o.Internal = 0;
	wd->ovls[ind].o.InternalHigh = 0;
	wd->ovls[ind].o.Offset = io_u->offset & 0xFFFFFFFF;
	wd->ovls[ind].o.OffsetHigh = io_u->offset >> 32;
	wd->ovls[ind].o.hEvent = wd->io_handles[ind];
	wd->ovls[ind].io_u = io_u;

	io_u->engine_data = &wd->ovls[ind];
	io_u->seen = 0;

	if (io_u->ddir == DDIR_WRITE) {
		success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, &wd->ovls[ind].o);
	} else if (io_u->ddir == DDIR_READ) {
		success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, &wd->ovls[ind].o);
	} else if (io_u->ddir == DDIR_SYNC     ||
			 io_u->ddir == DDIR_DATASYNC ||
			 io_u->ddir == DDIR_SYNC_FILE_RANGE)
	{
		FlushFileBuffers(io_u->file->hFile);
		return FIO_Q_COMPLETED;
	} else if (io_u->ddir == DDIR_TRIM) {
		log_info("explicit TRIM isn't supported on Windows");
		return FIO_Q_COMPLETED;
	} else
		assert(0);

	if (success) {
		io_u->seen = 1;
		io_u->resid = io_u->xfer_buflen - iobytes;
		io_u->error = 0;
		rc = FIO_Q_COMPLETED;
	} else if (!success && GetLastError() == ERROR_IO_PENDING) {
		wd->io_index = (wd->io_index + 1) % (2 * td->o.iodepth);
		rc = FIO_Q_QUEUED;
	} else {
		PrintError(__func__);
		io_u->error = GetLastError();
		io_u->resid = io_u->xfer_buflen;
		rc = FIO_Q_COMPLETED;
	}

	return rc;
}

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

	wd = td->io_ops->data;

	WaitForSingleObject(wd->iothread_stopped, INFINITE);

	if (wd != NULL) {
		CloseHandle(wd->iothread);
		CloseHandle(wd->iothread_stopped);
		CloseHandle(wd->iocomplete_event);

		for (i = 0; i < 2 * td->o.iodepth; i++) {
			CloseHandle(wd->io_handles[i]);
		}

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

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

/* Runs as a thread and waits for queued IO to complete */
static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
{
	OVERLAPPED *ovl;
	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;

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

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

		if (io_u->seen != 0)
			continue;

		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;
		}

		io_u->seen = 1;
		SetEvent(wd->iocomplete_event);
	}

	CloseHandle(ctx->iocp);
	SetEvent(ctx->wd->iothread_stopped);
	free(ctx);

	return 0;
}

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

	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->io_handles = malloc(2 * td->o.iodepth * sizeof(HANDLE));
		if (wd->io_handles == NULL)
			rc = 1;
	}

	if (!rc) {
		wd->ovls = malloc(2 * td->o.iodepth * sizeof(FIO_OVERLAPPED));
		if (wd->ovls == NULL)
			rc = 1;
	}

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

	if (!rc) {
		osInfo.dwOSVersionInfoSize = sizeof(OSVERSIONINFO);
		GetVersionEx(&osInfo);

		if (osInfo.dwMajorVersion >= 6)
			wd->have_cancelioex = TRUE;
		else
			wd->have_cancelioex = FALSE;
	}

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

			free(wd);
		}
	}

	td->io_ops->data = wd;
	return 0;
}

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

	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 ||
		td->o.td_ddir == TD_DDIR_RANDRW)
	{
		flags |= FILE_FLAG_SEQUENTIAL_SCAN;
	}
	else
	{
		flags |= FILE_FLAG_RANDOM_ACCESS;
	}

	if (td_read(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) {
		PrintError(__func__);
		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 windowsaio_data *wd;
		struct thread_ctx *ctx;
		hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);

		wd = td->io_ops->data;

		wd->io_index = 0;
		wd->iothread_running = TRUE;
		/* Create a manual-reset event */
		wd->iothread_stopped = CreateEvent(NULL, TRUE, FALSE, NULL);

		if (wd->iothread_stopped == NULL)
			rc = 1;

		if (!rc) {
			for (i = 0; i < 2 * td->o.iodepth; i++) {
				/* Create a manual-reset event for putting in OVERLAPPED */
				wd->io_handles[i] = CreateEvent(NULL, TRUE, FALSE, NULL);
				if (wd->io_handles[i] == NULL) {
					PrintError(__func__);
					rc = 1;
					break;
				}
			}
		}

		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) {
			PrintError(__func__);
			rc = 1;
		}
	}

	return rc;
}

static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f)
{
	struct windowsaio_data *wd;

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

	if (td->io_ops->data != NULL) {
		wd = td->io_ops->data;
		wd->iothread_running = FALSE;
		WaitForSingleObject(wd->iothread_stopped, INFINITE);
	}

	if (f->hFile != INVALID_HANDLE_VALUE) {
		if (!CloseHandle(f->hFile))
			PrintError(__func__);
	}

	f->hFile = INVALID_HANDLE_VALUE;
	return 0;
}

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
	1	/*
	2	* Native Windows async IO engine
	3	* Copyright (C) 2010 Bruce Cran <bruce@cran.org.uk>
	4	*/
	5
	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
	15	typedef struct {
	16	OVERLAPPED o;
	17	struct io_u *io_u;
	18	} FIO_OVERLAPPED;
	19
	20	struct windowsaio_data {
	21	HANDLE *io_handles;
	22	unsigned int io_index;
	23	FIO_OVERLAPPED *ovls;
	24
	25	HANDLE iothread;
	26	HANDLE iothread_stopped;
	27	BOOL iothread_running;
	28
	29	struct io_u **aio_events;
	30	HANDLE iocomplete_event;
	31	BOOL have_cancelioex;
	32	};
	33
	34	struct thread_ctx {
	35	HANDLE iocp;
	36	struct windowsaio_data *wd;
	37	};
	38
	39	static void PrintError(LPCSTR lpszFunction);
	40	static int fio_windowsaio_cancel(struct thread_data *td,
	41	struct io_u *io_u);
	42	static BOOL TimedOut(DWORD start_count, DWORD end_count);
	43	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	44	unsigned int max, struct timespec *t);
	45	static struct io_u fio_windowsaio_event(struct thread_data td, int event);
	46	static int fio_windowsaio_queue(struct thread_data *td,
	47	struct io_u *io_u);
	48	static void fio_windowsaio_cleanup(struct thread_data *td);
	49	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
	50	static int fio_windowsaio_init(struct thread_data *td);
	51	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f);
	52	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f);
	53
	54	/* CancelIoEx isn't in Cygwin's w32api */
	55	BOOL WINAPI CancelIoEx(
	56	HANDLE hFile,
	57	LPOVERLAPPED lpOverlapped
	58	);
	59
	60
	61
	62	int sync_file_range(int fd, off64_t offset, off64_t nbytes,
	63	unsigned int flags)
	64	{
	65	errno = ENOSYS;
	66	return -1;
	67	}
	68
	69	static void PrintError(LPCSTR lpszFunction)
	70	{
	71	// Retrieve the system error message for the last-error code
	72
	73	LPSTR lpMsgBuf;
	74	DWORD dw = GetLastError();
	75
	76	FormatMessage(
	77	FORMAT_MESSAGE_ALLOCATE_BUFFER \|
	78	FORMAT_MESSAGE_FROM_SYSTEM \|
	79	FORMAT_MESSAGE_IGNORE_INSERTS,
	80	NULL,
	81	dw,
	82	MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
	83	(LPTSTR)&lpMsgBuf,
	84	0, NULL );
	85
	86	log_err("%s - %s", lpszFunction, lpMsgBuf);
	87	LocalFree(lpMsgBuf);
	88	}
	89
	90	static int fio_windowsaio_cancel(struct thread_data *td,
	91	struct io_u *io_u)
	92	{
	93	int rc = 0;
	94
	95	struct windowsaio_data *wd = td->io_ops->data;
	96
	97	/* If we're running on Vista, we can cancel individual IO requests */
	98	if (wd->have_cancelioex) {
	99	FIO_OVERLAPPED *ovl = io_u->engine_data;
	100	if (!CancelIoEx(io_u->file->hFile, &ovl->o))
	101	rc = 1;
	102	} else
	103	rc = 1;
	104
	105	return rc;
	106	}
	107
	108	static BOOL TimedOut(DWORD start_count, DWORD end_count)
	109	{
	110	BOOL expired = FALSE;
	111	DWORD current_time;
	112
	113	current_time = GetTickCount();
	114
	115	if ((end_count > start_count) && current_time >= end_count)
	116	expired = TRUE;
	117	else if (current_time < start_count && current_time > end_count)
	118	expired = TRUE;
	119
	120	return expired;
	121	}
	122
	123	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	124	unsigned int max, struct timespec *t)
	125	{
	126	struct windowsaio_data *wd = td->io_ops->data;
	127	struct flist_head *entry;
	128	unsigned int dequeued = 0;
	129	struct io_u *io_u;
	130	DWORD start_count = 0, end_count = 0;
	131	BOOL timedout = FALSE;
	132	unsigned int mswait = 100;
	133
	134	if (t != NULL) {
	135	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	136	start_count = GetTickCount();
	137	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	138	}
	139
	140	while (dequeued < min && !timedout) {
	141	flist_for_each(entry, &td->io_u_busylist) {
	142	io_u = flist_entry(entry, struct io_u, list);
	143
	144	if (io_u->seen == 1) {
	145	io_u->seen = 2;
	146	wd->aio_events[dequeued] = io_u;
	147	dequeued++;
	148	}
	149
	150	if (dequeued == max)
	151	break;
	152	}
	153
	154	if (dequeued < min) {
	155	WaitForSingleObject(wd->iocomplete_event, mswait);
	156	}
	157
	158	if (t != NULL && TimedOut(start_count, end_count))
	159	timedout = TRUE;
	160	}
	161
	162	return dequeued;
	163	}
	164
	165	static struct io_u fio_windowsaio_event(struct thread_data td, int event)
	166	{
	167	struct windowsaio_data *wd = td->io_ops->data;
	168	return wd->aio_events[event];
	169	}
	170
	171	static int fio_windowsaio_queue(struct thread_data *td,
	172	struct io_u *io_u)
	173	{
	174	struct windowsaio_data *wd;
	175	DWORD iobytes;
	176	BOOL success = TRUE;
	177	int ind;
	178	int rc;
	179
	180	fio_ro_check(td, io_u);
	181
	182	wd = td->io_ops->data;
	183	ind = wd->io_index;
	184
	185	ResetEvent(wd->io_handles[ind]);
	186	wd->ovls[ind].o.Internal = 0;
	187	wd->ovls[ind].o.InternalHigh = 0;
	188	wd->ovls[ind].o.Offset = io_u->offset & 0xFFFFFFFF;
	189	wd->ovls[ind].o.OffsetHigh = io_u->offset >> 32;
	190	wd->ovls[ind].o.hEvent = wd->io_handles[ind];
	191	wd->ovls[ind].io_u = io_u;
	192
	193	io_u->engine_data = &wd->ovls[ind];
	194	io_u->seen = 0;
	195
	196	if (io_u->ddir == DDIR_WRITE) {
	197	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, &wd->ovls[ind].o);
	198	} else if (io_u->ddir == DDIR_READ) {
	199	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, &wd->ovls[ind].o);
	200	} else if (io_u->ddir == DDIR_SYNC \|\|
	201	io_u->ddir == DDIR_DATASYNC \|\|
	202	io_u->ddir == DDIR_SYNC_FILE_RANGE)
	203	{
	204	FlushFileBuffers(io_u->file->hFile);
	205	return FIO_Q_COMPLETED;
	206	} else if (io_u->ddir == DDIR_TRIM) {
	207	log_info("explicit TRIM isn't supported on Windows");
	208	return FIO_Q_COMPLETED;
	209	} else
	210	assert(0);
	211
	212	if (success) {
	213	io_u->seen = 1;
	214	io_u->resid = io_u->xfer_buflen - iobytes;
	215	io_u->error = 0;
	216	rc = FIO_Q_COMPLETED;
	217	} else if (!success && GetLastError() == ERROR_IO_PENDING) {
	218	wd->io_index = (wd->io_index + 1) % (2 * td->o.iodepth);
	219	rc = FIO_Q_QUEUED;
	220	} else {
	221	PrintError(__func__);
	222	io_u->error = GetLastError();
	223	io_u->resid = io_u->xfer_buflen;
	224	rc = FIO_Q_COMPLETED;
	225	}
	226
	227	return rc;
	228	}
	229
	230	static void fio_windowsaio_cleanup(struct thread_data *td)
	231	{
	232	int i;
	233	struct windowsaio_data *wd;
	234
	235	wd = td->io_ops->data;
	236
	237	WaitForSingleObject(wd->iothread_stopped, INFINITE);
	238
	239	if (wd != NULL) {
	240	CloseHandle(wd->iothread);
	241	CloseHandle(wd->iothread_stopped);
	242	CloseHandle(wd->iocomplete_event);
	243
	244	for (i = 0; i < 2 * td->o.iodepth; i++) {
	245	CloseHandle(wd->io_handles[i]);
	246	}
	247
	248	free(wd->aio_events);
	249	free(wd->io_handles);
	250	free(wd->ovls);
	251	free(wd);
	252
	253	td->io_ops->data = NULL;
	254	}
	255	}
	256
	257	/* Runs as a thread and waits for queued IO to complete */
	258	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	259	{
	260	OVERLAPPED *ovl;
	261	FIO_OVERLAPPED *fov;
	262	struct io_u *io_u;
	263	struct windowsaio_data *wd;
	264	struct thread_ctx *ctx;
	265	ULONG_PTR ulKey = 0;
	266	DWORD bytes;
	267
	268	ctx = (struct thread_ctx*)lpParameter;
	269	wd = ctx->wd;
	270
	271	while (ctx->wd->iothread_running) {
	272	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250))
	273	continue;
	274
	275	fov = CONTAINING_RECORD(ovl, FIO_OVERLAPPED, o);
	276	io_u = fov->io_u;
	277
	278	if (io_u->seen != 0)
	279	continue;
	280
	281	if (ovl->Internal == ERROR_SUCCESS) {
	282	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	283	io_u->error = 0;
	284	} else {
	285	io_u->resid = io_u->xfer_buflen;
	286	io_u->error = ovl->Internal;
	287	}
	288
	289	io_u->seen = 1;
	290	SetEvent(wd->iocomplete_event);
	291	}
	292
	293	CloseHandle(ctx->iocp);
	294	SetEvent(ctx->wd->iothread_stopped);
	295	free(ctx);
	296
	297	return 0;
	298	}
	299
	300	static int fio_windowsaio_init(struct thread_data *td)
	301	{
	302	struct windowsaio_data *wd;
	303	OSVERSIONINFO osInfo;
	304	int rc = 0;
	305
	306	wd = malloc(sizeof(struct windowsaio_data));
	307	if (wd != NULL)
	308	ZeroMemory(wd, sizeof(struct windowsaio_data));
	309	else
	310	rc = 1;
	311
	312	if (!rc) {
	313	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	314	if (wd->aio_events == NULL)
	315	rc = 1;
	316	}
	317
	318	if (!rc) {
	319	wd->io_handles = malloc(2 * td->o.iodepth * sizeof(HANDLE));
	320	if (wd->io_handles == NULL)
	321	rc = 1;
	322	}
	323
	324	if (!rc) {
	325	wd->ovls = malloc(2 * td->o.iodepth * sizeof(FIO_OVERLAPPED));
	326	if (wd->ovls == NULL)
	327	rc = 1;
	328	}
	329
	330	if (!rc) {
	331	/* Create an auto-reset event */
	332	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	333	if (wd->iocomplete_event == NULL)
	334	rc = 1;
	335	}
	336
	337	if (!rc) {
	338	osInfo.dwOSVersionInfoSize = sizeof(OSVERSIONINFO);
	339	GetVersionEx(&osInfo);
	340
	341	if (osInfo.dwMajorVersion >= 6)
	342	wd->have_cancelioex = TRUE;
	343	else
	344	wd->have_cancelioex = FALSE;
	345	}
	346
	347	if (rc) {
	348	PrintError(__func__);
	349	if (wd != NULL) {
	350	if (wd->ovls != NULL)
	351	free(wd->ovls);
	352	if (wd->io_handles != NULL)
	353	free(wd->io_handles);
	354	if (wd->aio_events != NULL)
	355	free(wd->aio_events);
	356
	357	free(wd);
	358	}
	359	}
	360
	361	td->io_ops->data = wd;
	362	return 0;
	363	}
	364
	365	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	366	{
	367	int rc = 0;
	368	HANDLE hFile;
	369	DWORD flags = FILE_ATTRIBUTE_TEMPORARY \| FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
	370	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	371	DWORD openmode = OPEN_ALWAYS;
	372	DWORD access;
	373	int i;
	374
	375	dprint(FD_FILE, "fd open %s\n", f->file_name);
	376
	377	if (f->filetype == FIO_TYPE_PIPE) {
	378	log_err("fio: windowsaio doesn't support pipes\n");
	379	return 1;
	380	}
	381
	382	if (!strcmp(f->file_name, "-")) {
	383	log_err("fio: can't read/write to stdin/out\n");
	384	return 1;
	385	}
	386
	387	if (td->o.odirect)
	388	flags \|= FILE_FLAG_NO_BUFFERING;
	389	if (td->o.sync_io)
	390	flags \|= FILE_FLAG_WRITE_THROUGH;
	391
	392
	393	if (td->o.td_ddir == TD_DDIR_READ \|\|
	394	td->o.td_ddir == TD_DDIR_WRITE \|\|
	395	td->o.td_ddir == TD_DDIR_RANDRW)
	396	{
	397	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
	398	}
	399	else
	400	{
	401	flags \|= FILE_FLAG_RANDOM_ACCESS;
	402	}
	403
	404	if (td_read(td) \|\| read_only)
	405	access = GENERIC_READ;
	406	else
	407	access = (GENERIC_READ \| GENERIC_WRITE);
	408
	409	if (td->o.create_on_open > 0)
	410	openmode = OPEN_ALWAYS;
	411	else
	412	openmode = OPEN_EXISTING;
	413
	414	f->hFile = CreateFile(f->file_name, access, sharemode,
	415	NULL, openmode, flags, NULL);
	416
	417	if (f->hFile == INVALID_HANDLE_VALUE) {
	418	PrintError(__func__);
	419	rc = 1;
	420	}
	421
	422	/* Only set up the competion port and thread if we're not just
	423	* querying the device size */
	424	if (!rc && td->io_ops->data != NULL) {
	425	struct windowsaio_data *wd;
	426	struct thread_ctx *ctx;
	427	hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
	428
	429	wd = td->io_ops->data;
	430
	431	wd->io_index = 0;
	432	wd->iothread_running = TRUE;
	433	/* Create a manual-reset event */
	434	wd->iothread_stopped = CreateEvent(NULL, TRUE, FALSE, NULL);
	435
	436	if (wd->iothread_stopped == NULL)
	437	rc = 1;
	438
	439	if (!rc) {
	440	for (i = 0; i < 2 * td->o.iodepth; i++) {
	441	/* Create a manual-reset event for putting in OVERLAPPED */
	442	wd->io_handles[i] = CreateEvent(NULL, TRUE, FALSE, NULL);
	443	if (wd->io_handles[i] == NULL) {
	444	PrintError(__func__);
	445	rc = 1;
	446	break;
	447	}
	448	}
	449	}
	450
	451	if (!rc) {
	452	ctx = malloc(sizeof(struct thread_ctx));
	453	ctx->iocp = hFile;
	454	ctx->wd = wd;
	455
	456	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	457	}
	458
	459	if (rc \|\| wd->iothread == NULL) {
	460	PrintError(__func__);
	461	rc = 1;
	462	}
	463	}
	464
	465	return rc;
	466	}
	467
	468	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	469	{
	470	struct windowsaio_data *wd;
	471
	472	dprint(FD_FILE, "fd close %s\n", f->file_name);
	473
	474	if (td->io_ops->data != NULL) {
	475	wd = td->io_ops->data;
	476	wd->iothread_running = FALSE;
	477	WaitForSingleObject(wd->iothread_stopped, INFINITE);
	478	}
	479
	480	if (f->hFile != INVALID_HANDLE_VALUE) {
	481	if (!CloseHandle(f->hFile))
	482	PrintError(__func__);
	483	}
	484
	485	f->hFile = INVALID_HANDLE_VALUE;
	486	return 0;
	487	}
	488
	489	static struct ioengine_ops ioengine = {
	490	.name = "windowsaio",
	491	.version = FIO_IOOPS_VERSION,
	492	.init = fio_windowsaio_init,
	493	.queue = fio_windowsaio_queue,
	494	.cancel = fio_windowsaio_cancel,
	495	.getevents = fio_windowsaio_getevents,
	496	.event = fio_windowsaio_event,
	497	.cleanup = fio_windowsaio_cleanup,
	498	.open_file = fio_windowsaio_open_file,
	499	.close_file = fio_windowsaio_close_file,
	500	.get_file_size = generic_get_file_size
	501	};
	502
	503	static void fio_init fio_posixaio_register(void)
	504	{
	505	register_ioengine(&ioengine);
	506	}
	507
	508	static void fio_exit fio_posixaio_unregister(void)
	509	{
	510	unregister_ioengine(&ioengine);
	511	}