[fio.git] / engines / windowsaio.c

/*
 * Native Windows async IO engine
 * Copyright (C) 2012 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);

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

struct windowsaio_data {
	struct fio_overlapped *ovls;
	struct io_u **aio_events;
	HANDLE iothread;
	HANDLE iocomplete_event;
	CANCELIOEX pCancelIoEx;
	BOOL iothread_running;
};

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

static int fio_windowsaio_cancel(struct thread_data *td,
			       struct io_u *io_u);
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, 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 int fio_windowsaio_init(struct thread_data *td)
{
	struct windowsaio_data *wd;
	HANDLE hKernel32Dll;
	int rc = 0;
	int i;

	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->ovls = malloc(td->o.iodepth * sizeof(struct fio_overlapped));
		if (wd->ovls == NULL)
			rc = 1;
	}

	if (!rc) {
		for (i = 0; i < td->o.iodepth; i++) {
			wd->ovls[i].io_free = TRUE;
			wd->ovls[i].io_complete = FALSE;

			wd->ovls[i].o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
			if (wd->ovls[i].o.hEvent == NULL) {
				rc = 1;
				break;
			}
		}
	}

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

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

			free(wd);
		}
	}

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

	return rc;
}

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

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

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

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


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 | 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("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;

	/*
	 * 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)
		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 thread_ctx *ctx;
		struct windowsaio_data *wd;

		hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);

		wd = td->io_ops->data;
		wd->iothread_running = TRUE;

		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)
			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))
			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, struct timespec *t)
{
	struct windowsaio_data *wd = td->io_ops->data;
	struct flist_head *entry;
	unsigned int dequeued = 0;
	struct io_u *io_u;
	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 {
		flist_for_each(entry, &td->io_u_busylist) {
			io_u = flist_entry(entry, struct io_u, list);
			fov = (struct fio_overlapped*)io_u->engine_data;

			if (fov->io_complete) {
				fov->io_complete = FALSE;
				fov->io_free  = TRUE;
				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)
{
	LPOVERLAPPED lpOvl = NULL;
	struct windowsaio_data *wd;
	DWORD iobytes;
	BOOL success = FALSE;
	int index;
	int rc = FIO_Q_COMPLETED;

	fio_ro_check(td, io_u);

	wd = td->io_ops->data;

	for (index = 0; index < td->o.iodepth; index++) {
		if (wd->ovls[index].io_free)
			break;
	}

	assert(index < td->o.iodepth);

	wd->ovls[index].io_free = FALSE;
	wd->ovls[index].io_u = io_u;
	lpOvl = &wd->ovls[index].o;
	lpOvl->Internal = STATUS_PENDING;
	lpOvl->InternalHigh = 0;
	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	lpOvl->OffsetHigh = io_u->offset >> 32;
	io_u->engine_data = &wd->ovls[index];

	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)
		    io_u->error = GetLastError();

		return FIO_Q_COMPLETED;
		break;
	case DDIR_TRIM:
		log_err("manual TRIM isn't supported on Windows");
		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 = 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))
			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 = ovl->Internal;
		}

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

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

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) {
		struct fio_overlapped *ovl = io_u->engine_data;

		if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
			rc = 1;
	} else
		rc = 1;

	return rc;
}

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
ecc314ba BC	1	/*
ecc314ba BC	2	* Native Windows async IO engine
93bcfd20	3	* Copyright (C) 2012 Bruce Cran <bruce@cran.org.uk>
ecc314ba BC	4	*/
ecc314ba BC	5
ecc314ba BC	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
ea4500d8 BC	15	typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);
ea4500d8 BC	16
67897036 BC	17	struct fio_overlapped {
	18	OVERLAPPED o;
	19	struct io_u *io_u;
	20	BOOL io_complete;
40c5db35	21	BOOL io_free;
67897036	22	};
ecc314ba	23
9b836561	24	struct windowsaio_data {
67897036	25	struct fio_overlapped *ovls;
9b836561	26	struct io_u **aio_events;
67897036	27	HANDLE iothread;
9b836561	28	HANDLE iocomplete_event;
ea4500d8	29	CANCELIOEX pCancelIoEx;
67897036	30	BOOL iothread_running;
9b836561 BC	31	};
9b836561 BC	32
ecc314ba	33	struct thread_ctx {
9b836561	34	HANDLE iocp;
ecc314ba BC	35	struct windowsaio_data *wd;
	36	};
	37
ecc314ba BC	38	static int fio_windowsaio_cancel(struct thread_data *td,
ecc314ba BC	39	struct io_u *io_u);
67897036	40	static BOOL timeout_expired(DWORD start_count, DWORD end_count);
ecc314ba BC	41	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	42	unsigned int max, struct timespec *t);
	43	static struct io_u fio_windowsaio_event(struct thread_data td, int event);
	44	static int fio_windowsaio_queue(struct thread_data *td,
	45	struct io_u *io_u);
	46	static void fio_windowsaio_cleanup(struct thread_data *td);
	47	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
	48	static int fio_windowsaio_init(struct thread_data *td);
	49	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f);
	50	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f);
	51
9b836561 BC	52	int sync_file_range(int fd, off64_t offset, off64_t nbytes,
	53	unsigned int flags)
	54	{
	55	errno = ENOSYS;
	56	return -1;
	57	}
	58
ecc314ba BC	59	static int fio_windowsaio_init(struct thread_data *td)
ecc314ba BC	60	{
ecc314ba	61	struct windowsaio_data *wd;
ea4500d8	62	HANDLE hKernel32Dll;
9b836561	63	int rc = 0;
67897036	64	int i;
ecc314ba	65
ecc314ba	66	wd = malloc(sizeof(struct windowsaio_data));
9b836561 BC	67	if (wd != NULL)
	68	ZeroMemory(wd, sizeof(struct windowsaio_data));
	69	else
	70	rc = 1;
ecc314ba	71
9b836561 BC	72	if (!rc) {
	73	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	74	if (wd->aio_events == NULL)
	75	rc = 1;
e4db9fec BC	76	}
e4db9fec BC	77
9b836561	78	if (!rc) {
67897036 BC	79	wd->ovls = malloc(td->o.iodepth * sizeof(struct fio_overlapped));
67897036 BC	80	if (wd->ovls == NULL)
9b836561	81	rc = 1;
e4db9fec BC	82	}
e4db9fec BC	83
9b836561	84	if (!rc) {
f9a58c2a BC	85	for (i = 0; i < td->o.iodepth; i++) {
	86	wd->ovls[i].io_free = TRUE;
	87	wd->ovls[i].io_complete = FALSE;
67897036 BC	88
	89	wd->ovls[i].o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
	90	if (wd->ovls[i].o.hEvent == NULL) {
	91	rc = 1;
	92	break;
	93	}
f9a58c2a	94	}
9b836561	95	}
e4db9fec	96
9b836561 BC	97	if (!rc) {
	98	/* Create an auto-reset event */
	99	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	100	if (wd->iocomplete_event == NULL)
	101	rc = 1;
	102	}
	103
9b836561	104	if (rc) {
9b836561 BC	105	if (wd != NULL) {
	106	if (wd->ovls != NULL)
	107	free(wd->ovls);
9b836561 BC	108	if (wd->aio_events != NULL)
	109	free(wd->aio_events);
	110
	111	free(wd);
	112	}
	113	}
ecc314ba	114
ea4500d8	115	hKernel32Dll = GetModuleHandle("kernel32.dll");
93bcfd20	116	wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
ecc314ba	117	td->io_ops->data = wd;
93bcfd20	118
4e79098f	119	return rc;
ecc314ba BC	120	}
ecc314ba BC	121
67897036 BC	122	static void fio_windowsaio_cleanup(struct thread_data *td)
	123	{
	124	int i;
	125	struct windowsaio_data *wd;
	126
	127	wd = td->io_ops->data;
	128
	129	if (wd != NULL) {
40c5db35 JA	130	wd->iothread_running = FALSE;
40c5db35 JA	131	WaitForSingleObject(wd->iothread, INFINITE);
67897036 BC	132
	133	CloseHandle(wd->iothread);
	134	CloseHandle(wd->iocomplete_event);
	135
93bcfd20	136	for (i = 0; i < td->o.iodepth; i++) {
67897036	137	CloseHandle(wd->ovls[i].o.hEvent);
93bcfd20	138	}
67897036 BC	139
	140	free(wd->aio_events);
	141	free(wd->ovls);
	142	free(wd);
	143
	144	td->io_ops->data = NULL;
	145	}
	146	}
	147
	148
ecc314ba BC	149	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	150	{
	151	int rc = 0;
	152	HANDLE hFile;
4e79098f	153	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
ecc314ba BC	154	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	155	DWORD openmode = OPEN_ALWAYS;
	156	DWORD access;
	157
	158	dprint(FD_FILE, "fd open %s\n", f->file_name);
	159
ecc314ba BC	160	if (f->filetype == FIO_TYPE_PIPE) {
	161	log_err("fio: windowsaio doesn't support pipes\n");
	162	return 1;
	163	}
	164
	165	if (!strcmp(f->file_name, "-")) {
	166	log_err("fio: can't read/write to stdin/out\n");
	167	return 1;
	168	}
	169
	170	if (td->o.odirect)
	171	flags \|= FILE_FLAG_NO_BUFFERING;
	172	if (td->o.sync_io)
	173	flags \|= FILE_FLAG_WRITE_THROUGH;
	174
93bcfd20 BC	175	/*
	176	* Inform Windows whether we're going to be doing sequential or
	177	* random io so it can tune the Cache Manager
	178	*/
ecc314ba	179	if (td->o.td_ddir == TD_DDIR_READ \|\|
ea4500d8	180	td->o.td_ddir == TD_DDIR_WRITE)
ecc314ba	181	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
ecc314ba	182	else
ecc314ba	183	flags \|= FILE_FLAG_RANDOM_ACCESS;
ecc314ba	184
ea4500d8	185	if (!td_write(td) \|\| read_only)
ecc314ba BC	186	access = GENERIC_READ;
	187	else
	188	access = (GENERIC_READ \| GENERIC_WRITE);
	189
93bcfd20	190	if (td->o.create_on_open)
ecc314ba BC	191	openmode = OPEN_ALWAYS;
	192	else
	193	openmode = OPEN_EXISTING;
	194
	195	f->hFile = CreateFile(f->file_name, access, sharemode,
	196	NULL, openmode, flags, NULL);
	197
4e79098f	198	if (f->hFile == INVALID_HANDLE_VALUE)
ecc314ba	199	rc = 1;
ecc314ba	200
93bcfd20	201	/* Only set up the completion port and thread if we're not just
ecc314ba	202	* querying the device size */
40c5db35	203	if (!rc && td->io_ops->data != NULL) {
ecc314ba	204	struct thread_ctx *ctx;
40c5db35	205	struct windowsaio_data *wd;
ecc314ba	206
40c5db35	207	hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
ea4500d8	208
40c5db35	209	wd = td->io_ops->data;
9b836561	210	wd->iothread_running = TRUE;
ecc314ba	211
9b836561 BC	212	if (!rc) {
	213	ctx = malloc(sizeof(struct thread_ctx));
	214	ctx->iocp = hFile;
	215	ctx->wd = wd;
	216
	217	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	218	}
ecc314ba	219
4e79098f	220	if (rc \|\| wd->iothread == NULL)
ecc314ba	221	rc = 1;
ecc314ba BC	222	}
ecc314ba BC	223
ecc314ba BC	224	return rc;
	225	}
	226
	227	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	228	{
4e79098f	229	int rc = 0;
1e42cc3d	230
9b836561 BC	231	dprint(FD_FILE, "fd close %s\n", f->file_name);
9b836561 BC	232
ecc314ba	233	if (f->hFile != INVALID_HANDLE_VALUE) {
9b836561	234	if (!CloseHandle(f->hFile))
4e79098f	235	rc = 1;
ecc314ba BC	236	}
	237
	238	f->hFile = INVALID_HANDLE_VALUE;
4e79098f	239	return rc;
ecc314ba BC	240	}
ecc314ba BC	241
67897036 BC	242	static BOOL timeout_expired(DWORD start_count, DWORD end_count)
	243	{
	244	BOOL expired = FALSE;
	245	DWORD current_time;
	246
	247	current_time = GetTickCount();
	248
	249	if ((end_count > start_count) && current_time >= end_count)
	250	expired = TRUE;
	251	else if (current_time < start_count && current_time > end_count)
	252	expired = TRUE;
	253
	254	return expired;
	255	}
	256
	257	static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
	258	{
	259	struct windowsaio_data *wd = td->io_ops->data;
	260	return wd->aio_events[event];
	261	}
	262
	263	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	264	unsigned int max, struct timespec *t)
	265	{
	266	struct windowsaio_data *wd = td->io_ops->data;
	267	struct flist_head *entry;
	268	unsigned int dequeued = 0;
	269	struct io_u *io_u;
	270	struct fio_overlapped *fov;
	271	DWORD start_count = 0;
	272	DWORD end_count = 0;
	273	DWORD status;
	274	DWORD mswait = 250;
	275
	276	if (t != NULL) {
	277	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	278	start_count = GetTickCount();
	279	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	280	}
	281
	282	do {
	283	flist_for_each(entry, &td->io_u_busylist) {
	284	io_u = flist_entry(entry, struct io_u, list);
	285	fov = (struct fio_overlapped*)io_u->engine_data;
	286
	287	if (fov->io_complete) {
40c5db35	288	fov->io_complete = FALSE;
67897036	289	fov->io_free = TRUE;
f9a58c2a	290	ResetEvent(fov->o.hEvent);
67897036 BC	291	wd->aio_events[dequeued] = io_u;
	292	dequeued++;
	293	}
	294
	295	if (dequeued >= min)
	296	break;
	297	}
	298
40c5db35	299	if (dequeued < min) {
67897036	300	status = WaitForSingleObject(wd->iocomplete_event, mswait);
f9a58c2a	301	if (status != WAIT_OBJECT_0 && dequeued >= min)
67897036 BC	302	break;
	303	}
	304
	305	if (dequeued >= min \|\| (t != NULL && timeout_expired(start_count, end_count)))
	306	break;
	307	} while (1);
	308
	309	return dequeued;
	310	}
	311
40c5db35	312	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
67897036	313	{
1e42cc3d	314	LPOVERLAPPED lpOvl = NULL;
67897036 BC	315	struct windowsaio_data *wd;
67897036 BC	316	DWORD iobytes;
93bcfd20	317	BOOL success = FALSE;
67897036 BC	318	int index;
	319	int rc = FIO_Q_COMPLETED;
	320
	321	fio_ro_check(td, io_u);
	322
	323	wd = td->io_ops->data;
	324
1e42cc3d	325	for (index = 0; index < td->o.iodepth; index++) {
f9a58c2a	326	if (wd->ovls[index].io_free)
1e42cc3d	327	break;
1e42cc3d	328	}
67897036	329
1e42cc3d	330	assert(index < td->o.iodepth);
67897036	331
f9a58c2a	332	wd->ovls[index].io_free = FALSE;
1e42cc3d	333	wd->ovls[index].io_u = io_u;
f9a58c2a	334	lpOvl = &wd->ovls[index].o;
4e79098f BC	335	lpOvl->Internal = STATUS_PENDING;
	336	lpOvl->InternalHigh = 0;
	337	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	338	lpOvl->OffsetHigh = io_u->offset >> 32;
1e42cc3d	339	io_u->engine_data = &wd->ovls[index];
67897036 BC	340
67897036 BC	341	switch (io_u->ddir) {
40c5db35	342	case DDIR_WRITE:
67897036 BC	343	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	344	break;
	345	case DDIR_READ:
	346	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	347	break;
	348	case DDIR_SYNC:
	349	case DDIR_DATASYNC:
	350	case DDIR_SYNC_FILE_RANGE:
	351	success = FlushFileBuffers(io_u->file->hFile);
	352	if (!success)
	353	io_u->error = GetLastError();
	354
	355	return FIO_Q_COMPLETED;
	356	break;
	357	case DDIR_TRIM:
	358	log_err("manual TRIM isn't supported on Windows");
	359	io_u->error = 1;
	360	io_u->resid = io_u->xfer_buflen;
	361	return FIO_Q_COMPLETED;
	362	break;
	363	default:
	364	assert(0);
93bcfd20	365	break;
67897036 BC	366	}
67897036 BC	367
40c5db35	368	if (success \|\| GetLastError() == ERROR_IO_PENDING)
67897036	369	rc = FIO_Q_QUEUED;
40c5db35	370	else {
67897036 BC	371	io_u->error = GetLastError();
	372	io_u->resid = io_u->xfer_buflen;
	373	}
	374
	375	return rc;
	376	}
	377
	378	/* Runs as a thread and waits for queued IO to complete */
	379	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	380	{
	381	OVERLAPPED *ovl;
	382	struct fio_overlapped *fov;
	383	struct io_u *io_u;
	384	struct windowsaio_data *wd;
	385	struct thread_ctx *ctx;
	386	ULONG_PTR ulKey = 0;
	387	DWORD bytes;
	388
	389	ctx = (struct thread_ctx*)lpParameter;
	390	wd = ctx->wd;
	391
	392	do {
	393	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250))
	394	continue;
	395
	396	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	397	io_u = fov->io_u;
	398
	399	if (ovl->Internal == ERROR_SUCCESS) {
	400	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	401	io_u->error = 0;
	402	} else {
	403	io_u->resid = io_u->xfer_buflen;
	404	io_u->error = ovl->Internal;
	405	}
	406
40c5db35	407	fov->io_complete = TRUE;
67897036 BC	408	SetEvent(wd->iocomplete_event);
	409	} while (ctx->wd->iothread_running);
	410
	411	CloseHandle(ctx->iocp);
	412	free(ctx);
	413	return 0;
	414	}
	415
	416	static int fio_windowsaio_cancel(struct thread_data *td,
	417	struct io_u *io_u)
	418	{
	419	int rc = 0;
	420
	421	struct windowsaio_data *wd = td->io_ops->data;
	422
	423	/* If we're running on Vista or newer, we can cancel individual IO requests */
	424	if (wd->pCancelIoEx != NULL) {
	425	struct fio_overlapped *ovl = io_u->engine_data;
40c5db35	426
67897036 BC	427	if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
	428	rc = 1;
	429	} else
	430	rc = 1;
	431
	432	return rc;
	433	}
	434
ecc314ba BC	435	static struct ioengine_ops ioengine = {
	436	.name = "windowsaio",
	437	.version = FIO_IOOPS_VERSION,
	438	.init = fio_windowsaio_init,
	439	.queue = fio_windowsaio_queue,
	440	.cancel = fio_windowsaio_cancel,
	441	.getevents = fio_windowsaio_getevents,
	442	.event = fio_windowsaio_event,
	443	.cleanup = fio_windowsaio_cleanup,
	444	.open_file = fio_windowsaio_open_file,
	445	.close_file = fio_windowsaio_close_file,
	446	.get_file_size = generic_get_file_size
	447	};
	448
	449	static void fio_init fio_posixaio_register(void)
	450	{
	451	register_ioengine(&ioengine);
	452	}
	453
	454	static void fio_exit fio_posixaio_unregister(void)
	455	{
	456	unregister_ioengine(&ioengine);
	457	}