[fio.git] / engines / windowsaio.c

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

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

#include "../fio.h"

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

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;
	CANCELIOEX pCancelIoEx;
	BOOL useIOCP;
};

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);

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 or newer, we can cancel individual IO requests */
	if (wd->pCancelIoEx != NULL) {
		FIO_OVERLAPPED *ovl = io_u->engine_data;
		if (!wd->pCancelIoEx(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 == 0) {
				io_u->seen = 1;
				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;
	LPOVERLAPPED lpOvl;
	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]);

	if (wd->useIOCP) {
	    lpOvl = &wd->ovls[ind].o;

    	lpOvl->Internal = STATUS_PENDING;
    	lpOvl->InternalHigh = 0;
    	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
    	lpOvl->OffsetHigh = io_u->offset >> 32;
    	lpOvl->hEvent = wd->io_handles[ind];
    	lpOvl->Pointer = NULL;
    	wd->ovls[ind].io_u = io_u;
	} else {
	    lpOvl = NULL;
    }

	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, lpOvl);
	} else if (io_u->ddir == DDIR_READ) {
		success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	} 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("Manual TRIM isn't supported on Windows");
		return FIO_Q_COMPLETED;
	} else
		assert(0);

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

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

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

        /* We sometimes get an IO request that hasn't completed yet. Ignore it. */
        if (ovl->Internal == STATUS_PENDING)
            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;
		}

		SetEvent(wd->iocomplete_event);
	} while (ctx->wd->iothread_running);

	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;
	HANDLE hKernel32Dll;
	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(td->o.iodepth * sizeof(HANDLE));
		if (wd->io_handles == NULL)
			rc = 1;
	}

	if (!rc) {
		wd->ovls = malloc(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) {
		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);
		}
	}

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

	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_FLAG_POSIX_SEMANTICS;
	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 && !td->o.sync_io && td->io_ops->data != NULL)
	    flags |= FILE_FLAG_OVERLAPPED;

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


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

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

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

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

	if (f->hFile == INVALID_HANDLE_VALUE) {
		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 && !td->o.odirect && !td->o.sync_io) {
		struct thread_ctx *ctx;
        struct windowsaio_data *wd;
		hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);


        wd = td->io_ops->data;

        if (!td->o.odirect && !td->o.sync_io)
            wd->useIOCP = 1;
        else
            wd->useIOCP = 0;

		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 < 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) 2011 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 BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);
	16
	17	typedef struct {
	18	OVERLAPPED o;
	19	struct io_u *io_u;
	20	} FIO_OVERLAPPED;
	21
	22	struct windowsaio_data {
	23	HANDLE *io_handles;
	24	unsigned int io_index;
	25	FIO_OVERLAPPED *ovls;
	26
	27	HANDLE iothread;
	28	HANDLE iothread_stopped;
	29	BOOL iothread_running;
	30
	31	struct io_u **aio_events;
	32	HANDLE iocomplete_event;
	33	CANCELIOEX pCancelIoEx;
	34	BOOL useIOCP;
	35	};
	36
	37	struct thread_ctx {
	38	HANDLE iocp;
	39	struct windowsaio_data *wd;
	40	};
	41
	42	static void PrintError(LPCSTR lpszFunction);
	43	static int fio_windowsaio_cancel(struct thread_data *td,
	44	struct io_u *io_u);
	45	static BOOL TimedOut(DWORD start_count, DWORD end_count);
	46	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	47	unsigned int max, struct timespec *t);
	48	static struct io_u fio_windowsaio_event(struct thread_data td, int event);
	49	static int fio_windowsaio_queue(struct thread_data *td,
	50	struct io_u *io_u);
	51	static void fio_windowsaio_cleanup(struct thread_data *td);
	52	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
	53	static int fio_windowsaio_init(struct thread_data *td);
	54	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f);
	55	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f);
	56
	57	int sync_file_range(int fd, off64_t offset, off64_t nbytes,
	58	unsigned int flags)
	59	{
	60	errno = ENOSYS;
	61	return -1;
	62	}
	63
	64	static void PrintError(LPCSTR lpszFunction)
	65	{
	66	// Retrieve the system error message for the last-error code
	67
	68	LPSTR lpMsgBuf;
	69	DWORD dw = GetLastError();
	70
	71	FormatMessage(
	72	FORMAT_MESSAGE_ALLOCATE_BUFFER \|
	73	FORMAT_MESSAGE_FROM_SYSTEM \|
	74	FORMAT_MESSAGE_IGNORE_INSERTS,
	75	NULL,
	76	dw,
	77	MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
	78	(LPTSTR)&lpMsgBuf,
	79	0, NULL );
	80
	81	log_err("%s - %s", lpszFunction, lpMsgBuf);
	82	LocalFree(lpMsgBuf);
	83	}
	84
	85	static int fio_windowsaio_cancel(struct thread_data *td,
	86	struct io_u *io_u)
	87	{
	88	int rc = 0;
	89
	90	struct windowsaio_data *wd = td->io_ops->data;
	91
	92	/* If we're running on Vista or newer, we can cancel individual IO requests */
	93	if (wd->pCancelIoEx != NULL) {
	94	FIO_OVERLAPPED *ovl = io_u->engine_data;
	95	if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
	96	rc = 1;
	97	} else
	98	rc = 1;
	99
	100	return rc;
	101	}
	102
	103	static BOOL TimedOut(DWORD start_count, DWORD end_count)
	104	{
	105	BOOL expired = FALSE;
	106	DWORD current_time;
	107
	108	current_time = GetTickCount();
	109
	110	if ((end_count > start_count) && current_time >= end_count)
	111	expired = TRUE;
	112	else if (current_time < start_count && current_time > end_count)
	113	expired = TRUE;
	114
	115	return expired;
	116	}
	117
	118	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	119	unsigned int max, struct timespec *t)
	120	{
	121	struct windowsaio_data *wd = td->io_ops->data;
	122	struct flist_head *entry;
	123	unsigned int dequeued = 0;
	124	struct io_u *io_u;
	125	DWORD start_count = 0, end_count = 0;
	126	BOOL timedout = FALSE;
	127	unsigned int mswait = 100;
	128
	129	if (t != NULL) {
	130	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	131	start_count = GetTickCount();
	132	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	133	}
	134
	135	while (dequeued < min && !timedout) {
	136	flist_for_each(entry, &td->io_u_busylist) {
	137	io_u = flist_entry(entry, struct io_u, list);
	138
	139	if (io_u->seen == 0) {
	140	io_u->seen = 1;
	141	wd->aio_events[dequeued] = io_u;
	142	dequeued++;
	143	}
	144
	145	if (dequeued == max)
	146	break;
	147	}
	148
	149	if (dequeued < min)
	150	WaitForSingleObject(wd->iocomplete_event, mswait);
	151
	152	if (t != NULL && TimedOut(start_count, end_count))
	153	timedout = TRUE;
	154	}
	155
	156	return dequeued;
	157	}
	158
	159	static struct io_u fio_windowsaio_event(struct thread_data td, int event)
	160	{
	161	struct windowsaio_data *wd = td->io_ops->data;
	162	return wd->aio_events[event];
	163	}
	164
	165	static int fio_windowsaio_queue(struct thread_data *td,
	166	struct io_u *io_u)
	167	{
	168	struct windowsaio_data *wd;
	169	LPOVERLAPPED lpOvl;
	170	DWORD iobytes;
	171	BOOL success = TRUE;
	172	int ind;
	173	int rc;
	174
	175	fio_ro_check(td, io_u);
	176
	177	wd = td->io_ops->data;
	178	ind = wd->io_index;
	179
	180	ResetEvent(wd->io_handles[ind]);
	181
	182	if (wd->useIOCP) {
	183	lpOvl = &wd->ovls[ind].o;
	184
	185	lpOvl->Internal = STATUS_PENDING;
	186	lpOvl->InternalHigh = 0;
	187	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	188	lpOvl->OffsetHigh = io_u->offset >> 32;
	189	lpOvl->hEvent = wd->io_handles[ind];
	190	lpOvl->Pointer = NULL;
	191	wd->ovls[ind].io_u = io_u;
	192	} else {
	193	lpOvl = NULL;
	194	}
	195
	196	io_u->engine_data = &wd->ovls[ind];
	197	io_u->seen = 0;
	198
	199	if (io_u->ddir == DDIR_WRITE) {
	200	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	201	} else if (io_u->ddir == DDIR_READ) {
	202	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	203	} else if (io_u->ddir == DDIR_SYNC \|\|
	204	io_u->ddir == DDIR_DATASYNC \|\|
	205	io_u->ddir == DDIR_SYNC_FILE_RANGE)
	206	{
	207	FlushFileBuffers(io_u->file->hFile);
	208	return FIO_Q_COMPLETED;
	209	} else if (io_u->ddir == DDIR_TRIM) {
	210	log_info("Manual TRIM isn't supported on Windows");
	211	return FIO_Q_COMPLETED;
	212	} else
	213	assert(0);
	214
	215	if (wd->useIOCP && (success \|\| GetLastError() == ERROR_IO_PENDING)) {
	216	wd->io_index = (wd->io_index + 1) % td->o.iodepth;
	217	rc = FIO_Q_QUEUED;
	218	} else if (success && !wd->useIOCP) {
	219	io_u->resid = io_u->xfer_buflen - iobytes;
	220	io_u->error = 0;
	221	rc = FIO_Q_COMPLETED;
	222	} else {
	223	PrintError(__func__);
	224	io_u->error = GetLastError();
	225	io_u->resid = io_u->xfer_buflen;
	226	rc = FIO_Q_COMPLETED;
	227	}
	228
	229	return rc;
	230	}
	231
	232	static void fio_windowsaio_cleanup(struct thread_data *td)
	233	{
	234	int i;
	235	struct windowsaio_data *wd;
	236
	237	wd = td->io_ops->data;
	238
	239	WaitForSingleObject(wd->iothread_stopped, INFINITE);
	240
	241	if (wd != NULL) {
	242	CloseHandle(wd->iothread);
	243	CloseHandle(wd->iothread_stopped);
	244	CloseHandle(wd->iocomplete_event);
	245
	246	for (i = 0; i < td->o.iodepth; i++) {
	247	CloseHandle(wd->io_handles[i]);
	248	}
	249
	250	free(wd->aio_events);
	251	free(wd->io_handles);
	252	free(wd->ovls);
	253	free(wd);
	254
	255	td->io_ops->data = NULL;
	256	}
	257	}
	258
	259	/* Runs as a thread and waits for queued IO to complete */
	260	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	261	{
	262	OVERLAPPED *ovl;
	263	FIO_OVERLAPPED *fov;
	264	struct io_u *io_u;
	265	struct windowsaio_data *wd;
	266	struct thread_ctx *ctx;
	267	ULONG_PTR ulKey = 0;
	268	DWORD bytes;
	269
	270	ctx = (struct thread_ctx*)lpParameter;
	271	wd = ctx->wd;
	272
	273	do {
	274	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250))
	275	continue;
	276
	277	fov = CONTAINING_RECORD(ovl, FIO_OVERLAPPED, o);
	278	io_u = fov->io_u;
	279
	280	/* We sometimes get an IO request that hasn't completed yet. Ignore it. */
	281	if (ovl->Internal == STATUS_PENDING)
	282	continue;
	283
	284	if (ovl->Internal == ERROR_SUCCESS) {
	285	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	286	io_u->error = 0;
	287	} else {
	288	io_u->resid = io_u->xfer_buflen;
	289	io_u->error = ovl->Internal;
	290	}
	291
	292	SetEvent(wd->iocomplete_event);
	293	} while (ctx->wd->iothread_running);
	294
	295	CloseHandle(ctx->iocp);
	296	SetEvent(ctx->wd->iothread_stopped);
	297	free(ctx);
	298
	299	return 0;
	300	}
	301
	302	static int fio_windowsaio_init(struct thread_data *td)
	303	{
	304	struct windowsaio_data *wd;
	305	HANDLE hKernel32Dll;
	306	int rc = 0;
	307
	308	wd = malloc(sizeof(struct windowsaio_data));
	309	if (wd != NULL)
	310	ZeroMemory(wd, sizeof(struct windowsaio_data));
	311	else
	312	rc = 1;
	313
	314	if (!rc) {
	315	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	316	if (wd->aio_events == NULL)
	317	rc = 1;
	318	}
	319
	320	if (!rc) {
	321	wd->io_handles = malloc(td->o.iodepth * sizeof(HANDLE));
	322	if (wd->io_handles == NULL)
	323	rc = 1;
	324	}
	325
	326	if (!rc) {
	327	wd->ovls = malloc(td->o.iodepth * sizeof(FIO_OVERLAPPED));
	328	if (wd->ovls == NULL)
	329	rc = 1;
	330	}
	331
	332	if (!rc) {
	333	/* Create an auto-reset event */
	334	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	335	if (wd->iocomplete_event == NULL)
	336	rc = 1;
	337	}
	338
	339	if (rc) {
	340	PrintError(__func__);
	341	if (wd != NULL) {
	342	if (wd->ovls != NULL)
	343	free(wd->ovls);
	344	if (wd->io_handles != NULL)
	345	free(wd->io_handles);
	346	if (wd->aio_events != NULL)
	347	free(wd->aio_events);
	348
	349	free(wd);
	350	}
	351	}
	352
	353	hKernel32Dll = GetModuleHandle("kernel32.dll");
	354	wd->pCancelIoEx = GetProcAddress(hKernel32Dll, "CancelIoEx");
	355
	356	td->io_ops->data = wd;
	357	return 0;
	358	}
	359
	360	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	361	{
	362	int rc = 0;
	363	HANDLE hFile;
	364	DWORD flags = FILE_FLAG_POSIX_SEMANTICS;
	365	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	366	DWORD openmode = OPEN_ALWAYS;
	367	DWORD access;
	368	int i;
	369
	370	dprint(FD_FILE, "fd open %s\n", f->file_name);
	371
	372	if (f->filetype == FIO_TYPE_PIPE) {
	373	log_err("fio: windowsaio doesn't support pipes\n");
	374	return 1;
	375	}
	376
	377	if (!strcmp(f->file_name, "-")) {
	378	log_err("fio: can't read/write to stdin/out\n");
	379	return 1;
	380	}
	381
	382	if (!td->o.odirect && !td->o.sync_io && td->io_ops->data != NULL)
	383	flags \|= FILE_FLAG_OVERLAPPED;
	384
	385	if (td->o.odirect)
	386	flags \|= FILE_FLAG_NO_BUFFERING;
	387	if (td->o.sync_io)
	388	flags \|= FILE_FLAG_WRITE_THROUGH;
	389
	390
	391	if (td->o.td_ddir == TD_DDIR_READ \|\|
	392	td->o.td_ddir == TD_DDIR_WRITE)
	393	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
	394	else
	395	flags \|= FILE_FLAG_RANDOM_ACCESS;
	396
	397	if (!td_write(td) \|\| read_only)
	398	access = GENERIC_READ;
	399	else
	400	access = (GENERIC_READ \| GENERIC_WRITE);
	401
	402	if (td->o.create_on_open > 0)
	403	openmode = OPEN_ALWAYS;
	404	else
	405	openmode = OPEN_EXISTING;
	406
	407	f->hFile = CreateFile(f->file_name, access, sharemode,
	408	NULL, openmode, flags, NULL);
	409
	410	if (f->hFile == INVALID_HANDLE_VALUE) {
	411	PrintError(__func__);
	412	rc = 1;
	413	}
	414
	415	/* Only set up the competion port and thread if we're not just
	416	* querying the device size */
	417	if (!rc && td->io_ops->data != NULL && !td->o.odirect && !td->o.sync_io) {
	418	struct thread_ctx *ctx;
	419	struct windowsaio_data *wd;
	420	hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
	421
	422
	423	wd = td->io_ops->data;
	424
	425	if (!td->o.odirect && !td->o.sync_io)
	426	wd->useIOCP = 1;
	427	else
	428	wd->useIOCP = 0;
	429
	430	wd->io_index = 0;
	431	wd->iothread_running = TRUE;
	432	/* Create a manual-reset event */
	433	wd->iothread_stopped = CreateEvent(NULL, TRUE, FALSE, NULL);
	434
	435	if (wd->iothread_stopped == NULL)
	436	rc = 1;
	437
	438	if (!rc) {
	439	for (i = 0; i < td->o.iodepth; i++) {
	440	/* Create a manual-reset event for putting in OVERLAPPED */
	441	wd->io_handles[i] = CreateEvent(NULL, TRUE, FALSE, NULL);
	442	if (wd->io_handles[i] == NULL) {
	443	PrintError(__func__);
	444	rc = 1;
	445	break;
	446	}
	447	}
	448	}
	449
	450	if (!rc) {
	451	ctx = malloc(sizeof(struct thread_ctx));
	452	ctx->iocp = hFile;
	453	ctx->wd = wd;
	454
	455	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	456	}
	457
	458	if (rc \|\| wd->iothread == NULL) {
	459	PrintError(__func__);
	460	rc = 1;
	461	}
	462	}
	463
	464	return rc;
	465	}
	466
	467	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	468	{
	469	struct windowsaio_data *wd;
	470
	471	dprint(FD_FILE, "fd close %s\n", f->file_name);
	472
	473	if (td->io_ops->data != NULL) {
	474	wd = td->io_ops->data;
	475	wd->iothread_running = FALSE;
	476	WaitForSingleObject(wd->iothread_stopped, INFINITE);
	477	}
	478
	479	if (f->hFile != INVALID_HANDLE_VALUE) {
	480	if (!CloseHandle(f->hFile))
	481	PrintError(__func__);
	482	}
	483
	484	f->hFile = INVALID_HANDLE_VALUE;
	485	return 0;
	486	}
	487
	488	static struct ioengine_ops ioengine = {
	489	.name = "windowsaio",
	490	.version = FIO_IOOPS_VERSION,
	491	.init = fio_windowsaio_init,
	492	.queue = fio_windowsaio_queue,
	493	.cancel = fio_windowsaio_cancel,
	494	.getevents = fio_windowsaio_getevents,
	495	.event = fio_windowsaio_event,
	496	.cleanup = fio_windowsaio_cleanup,
	497	.open_file = fio_windowsaio_open_file,
	498	.close_file = fio_windowsaio_close_file,
	499	.get_file_size = generic_get_file_size
	500	};
	501
	502	static void fio_init fio_posixaio_register(void)
	503	{
	504	register_ioengine(&ioengine);
	505	}
	506
	507	static void fio_exit fio_posixaio_unregister(void)
	508	{
	509	unregister_ioengine(&ioengine);
	510	}