[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);

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;
	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);
static int win_to_posix_error(DWORD winerr);

static int win_to_posix_error(DWORD winerr)
{
	switch (winerr)
	{
	case ERROR_FILE_NOT_FOUND:		return ENOENT;
	case ERROR_PATH_NOT_FOUND:		return ENOENT;
	case ERROR_ACCESS_DENIED:		return EACCES;
	case ERROR_INVALID_HANDLE:		return EBADF;
	case ERROR_NOT_ENOUGH_MEMORY:	return ENOMEM;
	case ERROR_INVALID_DATA:		return EINVAL;
	case ERROR_OUTOFMEMORY:			return ENOMEM;
	case ERROR_INVALID_DRIVE:		return ENODEV;
	case ERROR_NOT_SAME_DEVICE:		return EXDEV;
	case ERROR_WRITE_PROTECT:		return EROFS;
	case ERROR_BAD_UNIT:			return ENODEV;
	case ERROR_SHARING_VIOLATION:	return EACCES;
	case ERROR_LOCK_VIOLATION:		return EACCES;
	case ERROR_SHARING_BUFFER_EXCEEDED:	return ENOLCK;
	case ERROR_HANDLE_DISK_FULL:	return ENOSPC;
	case ERROR_NOT_SUPPORTED:		return ENOSYS;
	case ERROR_FILE_EXISTS:			return EEXIST;
	case ERROR_CANNOT_MAKE:			return EPERM;
	case ERROR_INVALID_PARAMETER:	return EINVAL;
	case ERROR_NO_PROC_SLOTS:		return EAGAIN;
	case ERROR_BROKEN_PIPE:			return EPIPE;
	case ERROR_OPEN_FAILED:			return EIO;
	case ERROR_NO_MORE_SEARCH_HANDLES:	return ENFILE;
	case ERROR_CALL_NOT_IMPLEMENTED:	return ENOSYS;
	case ERROR_INVALID_NAME:		return ENOENT;
	case ERROR_WAIT_NO_CHILDREN:	return ECHILD;
	case ERROR_CHILD_NOT_COMPLETE:	return EBUSY;
	case ERROR_DIR_NOT_EMPTY:		return ENOTEMPTY;
	case ERROR_SIGNAL_REFUSED:		return EIO;
	case ERROR_BAD_PATHNAME:		return ENOENT;
	case ERROR_SIGNAL_PENDING:		return EBUSY;
	case ERROR_MAX_THRDS_REACHED:	return EAGAIN;
	case ERROR_BUSY:				return EBUSY;
	case ERROR_ALREADY_EXISTS:		return EEXIST;
	case ERROR_NO_SIGNAL_SENT:		return EIO;
	case ERROR_FILENAME_EXCED_RANGE:	return EINVAL;
	case ERROR_META_EXPANSION_TOO_LONG:	return EINVAL;
	case ERROR_INVALID_SIGNAL_NUMBER:	return EINVAL;
	case ERROR_THREAD_1_INACTIVE:	return EINVAL;
	case ERROR_BAD_PIPE:			return EINVAL;
	case ERROR_PIPE_BUSY:			return EBUSY;
	case ERROR_NO_DATA:				return EPIPE;
	case ERROR_MORE_DATA:			return EAGAIN;
	case ERROR_DIRECTORY:			return ENOTDIR;
	case ERROR_PIPE_CONNECTED:		return EBUSY;
	case ERROR_NO_TOKEN:			return EINVAL;
	case ERROR_PROCESS_ABORTED:		return EFAULT;
	case ERROR_BAD_DEVICE:			return ENODEV;
	case ERROR_BAD_USERNAME:		return EINVAL;
	case ERROR_OPEN_FILES:			return EAGAIN;
	case ERROR_ACTIVE_CONNECTIONS:	return EAGAIN;
	case ERROR_DEVICE_IN_USE:		return EAGAIN;
	case ERROR_INVALID_AT_INTERRUPT_TIME:	return EINTR;
	case ERROR_IO_DEVICE:			return EIO;
	case ERROR_NOT_OWNER:			return EPERM;
	case ERROR_END_OF_MEDIA:		return ENOSPC;
	case ERROR_EOM_OVERFLOW:		return ENOSPC;
	case ERROR_BEGINNING_OF_MEDIA:	return ESPIPE;
	case ERROR_SETMARK_DETECTED:	return ESPIPE;
	case ERROR_NO_DATA_DETECTED:	return ENOSPC;
	case ERROR_POSSIBLE_DEADLOCK:	return EDEADLOCK;
	case ERROR_CRC:					return EIO;
	case ERROR_NEGATIVE_SEEK:		return EINVAL;
	case ERROR_DISK_FULL:			return ENOSPC;
	case ERROR_NOACCESS:			return EFAULT;
	case ERROR_FILE_INVALID:		return ENXIO;
	}

	return winerr;
}

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) {
		/* 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->aio_events != NULL)
				free(wd->aio_events);

			free(wd);
		}
	}

	hKernel32Dll = GetModuleHandle("kernel32.dll");
	wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
	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)
			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("fio: out of memory in windowsaio\n");
			CloseHandle(hFile);
			rc = 1;
		}

		if (!rc)
		{
			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 void fio_windowsaio_cleanup(struct thread_data *td)
{
	struct windowsaio_data *wd;

	wd = td->io_ops->data;

	if (wd != NULL) {
		wd->iothread_running = FALSE;
		WaitForSingleObject(wd->iothread, INFINITE);

		CloseHandle(wd->iothread);
		CloseHandle(wd->iocomplete_event);

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

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


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

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

	if (f->filetype == FIO_TYPE_PIPE) {
		log_err("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 windowsaio_data *wd;

		wd = td->io_ops->data;

		if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == 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;
				ResetEvent(fov->o.hEvent);
				wd->aio_events[dequeued] = io_u;
				dequeued++;
			}

			if (dequeued >= min)
				break;
		}

		if (dequeued < min) {
			status = WaitForSingleObject(wd->iocomplete_event, mswait);
			if (status != WAIT_OBJECT_0 && dequeued >= min)
			    break;
		}

		if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count)))
			break;
	} while (1);

	return dequeued;
}

static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u)
{
	struct fio_overlapped *o = io_u->engine_data;
	LPOVERLAPPED lpOvl = &o->o;
	DWORD iobytes;
	BOOL success = FALSE;
	int rc = FIO_Q_COMPLETED;

	fio_ro_check(td, io_u);

	lpOvl->Internal = STATUS_PENDING;
	lpOvl->InternalHigh = 0;
	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	lpOvl->OffsetHigh = io_u->offset >> 32;

	switch (io_u->ddir) {
	case DDIR_WRITE:
		success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
		break;
	case DDIR_READ:
		success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
		break;
	case DDIR_SYNC:
	case DDIR_DATASYNC:
	case DDIR_SYNC_FILE_RANGE:
		success = FlushFileBuffers(io_u->file->hFile);
		if (!success)
		    io_u->error = win_to_posix_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 = win_to_posix_error(GetLastError());
		io_u->resid = io_u->xfer_buflen;
	}

	return rc;
}

/* Runs as a thread and waits for queued IO to complete */
static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
{
	OVERLAPPED *ovl;
	struct fio_overlapped *fov;
	struct io_u *io_u;
	struct windowsaio_data *wd;
	struct thread_ctx *ctx;
	ULONG_PTR ulKey = 0;
	DWORD bytes;

	ctx = (struct thread_ctx*)lpParameter;
	wd = ctx->wd;

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

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

		if (ovl->Internal == ERROR_SUCCESS) {
			io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
			io_u->error = 0;
		} else {
			io_u->resid = io_u->xfer_buflen;
			io_u->error = win_to_posix_error(GetLastError());
		}

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

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

static 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 void fio_windowsaio_io_u_free(struct thread_data *td, struct io_u *io_u)
{
	struct fio_overlapped *o = io_u->engine_data;

	if (o) {
		CloseHandle(o->o.hEvent);
		io_u->engine_data = NULL;
		free(o);
	}
}

static int fio_windowsaio_io_u_init(struct thread_data *td, struct io_u *io_u)
{
	struct fio_overlapped *o;

	o = malloc(sizeof(*o));
	o->io_complete = FALSE;
	o->io_u = io_u;
	o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
	if (!o->o.hEvent) {
		free(o);
		return 1;
	}

	io_u->engine_data = o;
	return 0;
}

static struct ioengine_ops ioengine = {
	.name		= "windowsaio",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_windowsaio_init,
	.queue		= fio_windowsaio_queue,
	.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,
	.io_u_init	= fio_windowsaio_io_u_init,
	.io_u_free	= fio_windowsaio_io_u_free,
};

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
66c098b8 BC	17	int geterrno_from_win_error (DWORD code, int deferrno);
66c098b8 BC	18
67897036 BC	19	struct fio_overlapped {
	20	OVERLAPPED o;
	21	struct io_u *io_u;
	22	BOOL io_complete;
67897036	23	};
ecc314ba	24
9b836561	25	struct windowsaio_data {
9b836561	26	struct io_u **aio_events;
5a90bb5f	27	HANDLE iocp;
67897036	28	HANDLE iothread;
9b836561	29	HANDLE iocomplete_event;
ea4500d8	30	CANCELIOEX pCancelIoEx;
67897036	31	BOOL iothread_running;
9b836561 BC	32	};
9b836561 BC	33
ecc314ba	34	struct thread_ctx {
9b836561	35	HANDLE iocp;
ecc314ba BC	36	struct windowsaio_data *wd;
	37	};
	38
ecc314ba BC	39	static int fio_windowsaio_cancel(struct thread_data *td,
ecc314ba BC	40	struct io_u *io_u);
67897036	41	static BOOL timeout_expired(DWORD start_count, DWORD end_count);
ecc314ba BC	42	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	43	unsigned int max, struct timespec *t);
	44	static struct io_u fio_windowsaio_event(struct thread_data td, int event);
	45	static int fio_windowsaio_queue(struct thread_data *td,
	46	struct io_u *io_u);
	47	static void fio_windowsaio_cleanup(struct thread_data *td);
	48	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
	49	static int fio_windowsaio_init(struct thread_data *td);
	50	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f);
	51	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f);
2277d5d5	52	static int win_to_posix_error(DWORD winerr);
66c098b8	53
2277d5d5	54	static int win_to_posix_error(DWORD winerr)
66c098b8 BC	55	{
	56	switch (winerr)
	57	{
	58	case ERROR_FILE_NOT_FOUND: return ENOENT;
	59	case ERROR_PATH_NOT_FOUND: return ENOENT;
	60	case ERROR_ACCESS_DENIED: return EACCES;
	61	case ERROR_INVALID_HANDLE: return EBADF;
	62	case ERROR_NOT_ENOUGH_MEMORY: return ENOMEM;
	63	case ERROR_INVALID_DATA: return EINVAL;
	64	case ERROR_OUTOFMEMORY: return ENOMEM;
	65	case ERROR_INVALID_DRIVE: return ENODEV;
	66	case ERROR_NOT_SAME_DEVICE: return EXDEV;
	67	case ERROR_WRITE_PROTECT: return EROFS;
	68	case ERROR_BAD_UNIT: return ENODEV;
	69	case ERROR_SHARING_VIOLATION: return EACCES;
	70	case ERROR_LOCK_VIOLATION: return EACCES;
	71	case ERROR_SHARING_BUFFER_EXCEEDED: return ENOLCK;
	72	case ERROR_HANDLE_DISK_FULL: return ENOSPC;
	73	case ERROR_NOT_SUPPORTED: return ENOSYS;
	74	case ERROR_FILE_EXISTS: return EEXIST;
	75	case ERROR_CANNOT_MAKE: return EPERM;
	76	case ERROR_INVALID_PARAMETER: return EINVAL;
	77	case ERROR_NO_PROC_SLOTS: return EAGAIN;
	78	case ERROR_BROKEN_PIPE: return EPIPE;
	79	case ERROR_OPEN_FAILED: return EIO;
	80	case ERROR_NO_MORE_SEARCH_HANDLES: return ENFILE;
	81	case ERROR_CALL_NOT_IMPLEMENTED: return ENOSYS;
	82	case ERROR_INVALID_NAME: return ENOENT;
	83	case ERROR_WAIT_NO_CHILDREN: return ECHILD;
	84	case ERROR_CHILD_NOT_COMPLETE: return EBUSY;
	85	case ERROR_DIR_NOT_EMPTY: return ENOTEMPTY;
	86	case ERROR_SIGNAL_REFUSED: return EIO;
	87	case ERROR_BAD_PATHNAME: return ENOENT;
	88	case ERROR_SIGNAL_PENDING: return EBUSY;
	89	case ERROR_MAX_THRDS_REACHED: return EAGAIN;
	90	case ERROR_BUSY: return EBUSY;
	91	case ERROR_ALREADY_EXISTS: return EEXIST;
	92	case ERROR_NO_SIGNAL_SENT: return EIO;
	93	case ERROR_FILENAME_EXCED_RANGE: return EINVAL;
	94	case ERROR_META_EXPANSION_TOO_LONG: return EINVAL;
	95	case ERROR_INVALID_SIGNAL_NUMBER: return EINVAL;
	96	case ERROR_THREAD_1_INACTIVE: return EINVAL;
	97	case ERROR_BAD_PIPE: return EINVAL;
	98	case ERROR_PIPE_BUSY: return EBUSY;
	99	case ERROR_NO_DATA: return EPIPE;
	100	case ERROR_MORE_DATA: return EAGAIN;
	101	case ERROR_DIRECTORY: return ENOTDIR;
	102	case ERROR_PIPE_CONNECTED: return EBUSY;
	103	case ERROR_NO_TOKEN: return EINVAL;
	104	case ERROR_PROCESS_ABORTED: return EFAULT;
	105	case ERROR_BAD_DEVICE: return ENODEV;
	106	case ERROR_BAD_USERNAME: return EINVAL;
	107	case ERROR_OPEN_FILES: return EAGAIN;
	108	case ERROR_ACTIVE_CONNECTIONS: return EAGAIN;
	109	case ERROR_DEVICE_IN_USE: return EAGAIN;
	110	case ERROR_INVALID_AT_INTERRUPT_TIME: return EINTR;
	111	case ERROR_IO_DEVICE: return EIO;
	112	case ERROR_NOT_OWNER: return EPERM;
	113	case ERROR_END_OF_MEDIA: return ENOSPC;
	114	case ERROR_EOM_OVERFLOW: return ENOSPC;
	115	case ERROR_BEGINNING_OF_MEDIA: return ESPIPE;
	116	case ERROR_SETMARK_DETECTED: return ESPIPE;
	117	case ERROR_NO_DATA_DETECTED: return ENOSPC;
	118	case ERROR_POSSIBLE_DEADLOCK: return EDEADLOCK;
119	case ERROR_CRC: return EIO;
120	case ERROR_NEGATIVE_SEEK: return EINVAL;
121	case ERROR_DISK_FULL: return ENOSPC;
122	case ERROR_NOACCESS: return EFAULT;
123	case ERROR_FILE_INVALID: return ENXIO;
124	}
125
126	return winerr;
127	}
ecc314ba	128
ecc314ba BC	129	static int fio_windowsaio_init(struct thread_data *td)
ecc314ba BC	130	{
ecc314ba	131	struct windowsaio_data *wd;
ea4500d8	132	HANDLE hKernel32Dll;
9b836561	133	int rc = 0;
ecc314ba	134
ecc314ba	135	wd = malloc(sizeof(struct windowsaio_data));
9b836561 BC	136	if (wd != NULL)
	137	ZeroMemory(wd, sizeof(struct windowsaio_data));
	138	else
	139	rc = 1;
ecc314ba	140
9b836561 BC	141	if (!rc) {
	142	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	143	if (wd->aio_events == NULL)
	144	rc = 1;
e4db9fec BC	145	}
e4db9fec BC	146
9b836561 BC	147	if (!rc) {
	148	/* Create an auto-reset event */
	149	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	150	if (wd->iocomplete_event == NULL)
	151	rc = 1;
	152	}
	153
9b836561	154	if (rc) {
9b836561	155	if (wd != NULL) {
9b836561 BC	156	if (wd->aio_events != NULL)
	157	free(wd->aio_events);
	158
	159	free(wd);
	160	}
	161	}
ecc314ba	162
ea4500d8	163	hKernel32Dll = GetModuleHandle("kernel32.dll");
93bcfd20	164	wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
ecc314ba	165	td->io_ops->data = wd;
93bcfd20	166
5a90bb5f BC	167
	168	if (!rc) {
	169	struct thread_ctx *ctx;
	170	struct windowsaio_data *wd;
	171	HANDLE hFile;
	172
	173	hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0);
	174	if (hFile == INVALID_HANDLE_VALUE)
	175	rc = 1;
	176
	177	wd = td->io_ops->data;
	178	wd->iothread_running = TRUE;
	179	wd->iocp = hFile;
	180
	181	if (!rc)
	182	ctx = malloc(sizeof(struct thread_ctx));
	183
	184	if (!rc && ctx == NULL)
	185	{
	186	log_err("fio: out of memory in windowsaio\n");
	187	CloseHandle(hFile);
	188	rc = 1;
	189	}
	190
	191	if (!rc)
	192	{
	193	ctx->iocp = hFile;
	194	ctx->wd = wd;
	195	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	196	}
	197
	198	if (rc \|\| wd->iothread == NULL)
	199	rc = 1;
	200	}
	201
4e79098f	202	return rc;
ecc314ba BC	203	}
ecc314ba BC	204
67897036 BC	205	static void fio_windowsaio_cleanup(struct thread_data *td)
67897036 BC	206	{
67897036 BC	207	struct windowsaio_data *wd;
	208
	209	wd = td->io_ops->data;
	210
	211	if (wd != NULL) {
40c5db35 JA	212	wd->iothread_running = FALSE;
40c5db35 JA	213	WaitForSingleObject(wd->iothread, INFINITE);
67897036 BC	214
	215	CloseHandle(wd->iothread);
	216	CloseHandle(wd->iocomplete_event);
	217
67897036	218	free(wd->aio_events);
67897036 BC	219	free(wd);
	220
	221	td->io_ops->data = NULL;
	222	}
	223	}
	224
	225
ecc314ba BC	226	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	227	{
	228	int rc = 0;
4e79098f	229	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
ecc314ba BC	230	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	231	DWORD openmode = OPEN_ALWAYS;
	232	DWORD access;
	233
	234	dprint(FD_FILE, "fd open %s\n", f->file_name);
	235
ecc314ba BC	236	if (f->filetype == FIO_TYPE_PIPE) {
	237	log_err("fio: windowsaio doesn't support pipes\n");
	238	return 1;
	239	}
	240
	241	if (!strcmp(f->file_name, "-")) {
	242	log_err("fio: can't read/write to stdin/out\n");
	243	return 1;
	244	}
	245
	246	if (td->o.odirect)
	247	flags \|= FILE_FLAG_NO_BUFFERING;
	248	if (td->o.sync_io)
	249	flags \|= FILE_FLAG_WRITE_THROUGH;
	250
93bcfd20 BC	251	/*
	252	* Inform Windows whether we're going to be doing sequential or
	253	* random io so it can tune the Cache Manager
	254	*/
ecc314ba	255	if (td->o.td_ddir == TD_DDIR_READ \|\|
ea4500d8	256	td->o.td_ddir == TD_DDIR_WRITE)
ecc314ba	257	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
ecc314ba	258	else
ecc314ba	259	flags \|= FILE_FLAG_RANDOM_ACCESS;
ecc314ba	260
ea4500d8	261	if (!td_write(td) \|\| read_only)
ecc314ba BC	262	access = GENERIC_READ;
	263	else
	264	access = (GENERIC_READ \| GENERIC_WRITE);
	265
93bcfd20	266	if (td->o.create_on_open)
ecc314ba BC	267	openmode = OPEN_ALWAYS;
	268	else
	269	openmode = OPEN_EXISTING;
	270
	271	f->hFile = CreateFile(f->file_name, access, sharemode,
	272	NULL, openmode, flags, NULL);
	273
4e79098f	274	if (f->hFile == INVALID_HANDLE_VALUE)
ecc314ba	275	rc = 1;
ecc314ba	276
93bcfd20	277	/* Only set up the completion port and thread if we're not just
ecc314ba	278	* querying the device size */
40c5db35	279	if (!rc && td->io_ops->data != NULL) {
40c5db35	280	struct windowsaio_data *wd;
ecc314ba	281
40c5db35	282	wd = td->io_ops->data;
ecc314ba	283
5a90bb5f	284	if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL)
ecc314ba	285	rc = 1;
ecc314ba BC	286	}
ecc314ba BC	287
ecc314ba BC	288	return rc;
	289	}
	290
	291	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	292	{
4e79098f	293	int rc = 0;
1e42cc3d	294
9b836561 BC	295	dprint(FD_FILE, "fd close %s\n", f->file_name);
9b836561 BC	296
ecc314ba	297	if (f->hFile != INVALID_HANDLE_VALUE) {
9b836561	298	if (!CloseHandle(f->hFile))
4e79098f	299	rc = 1;
ecc314ba BC	300	}
	301
	302	f->hFile = INVALID_HANDLE_VALUE;
4e79098f	303	return rc;
ecc314ba BC	304	}
ecc314ba BC	305
67897036 BC	306	static BOOL timeout_expired(DWORD start_count, DWORD end_count)
	307	{
	308	BOOL expired = FALSE;
	309	DWORD current_time;
	310
	311	current_time = GetTickCount();
	312
	313	if ((end_count > start_count) && current_time >= end_count)
	314	expired = TRUE;
	315	else if (current_time < start_count && current_time > end_count)
	316	expired = TRUE;
	317
	318	return expired;
	319	}
	320
	321	static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
	322	{
	323	struct windowsaio_data *wd = td->io_ops->data;
	324	return wd->aio_events[event];
	325	}
	326
	327	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	328	unsigned int max, struct timespec *t)
	329	{
	330	struct windowsaio_data *wd = td->io_ops->data;
	331	struct flist_head *entry;
	332	unsigned int dequeued = 0;
	333	struct io_u *io_u;
	334	struct fio_overlapped *fov;
	335	DWORD start_count = 0;
	336	DWORD end_count = 0;
	337	DWORD status;
	338	DWORD mswait = 250;
	339
	340	if (t != NULL) {
	341	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	342	start_count = GetTickCount();
	343	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	344	}
	345
	346	do {
	347	flist_for_each(entry, &td->io_u_busylist) {
	348	io_u = flist_entry(entry, struct io_u, list);
	349	fov = (struct fio_overlapped*)io_u->engine_data;
	350
	351	if (fov->io_complete) {
40c5db35	352	fov->io_complete = FALSE;
f9a58c2a	353	ResetEvent(fov->o.hEvent);
67897036 BC	354	wd->aio_events[dequeued] = io_u;
	355	dequeued++;
	356	}
	357
	358	if (dequeued >= min)
	359	break;
	360	}
	361
40c5db35	362	if (dequeued < min) {
67897036	363	status = WaitForSingleObject(wd->iocomplete_event, mswait);
f9a58c2a	364	if (status != WAIT_OBJECT_0 && dequeued >= min)
67897036 BC	365	break;
	366	}
	367
	368	if (dequeued >= min \|\| (t != NULL && timeout_expired(start_count, end_count)))
	369	break;
	370	} while (1);
	371
	372	return dequeued;
	373	}
	374
40c5db35	375	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
67897036	376	{
c73ed246 JA	377	struct fio_overlapped *o = io_u->engine_data;
c73ed246 JA	378	LPOVERLAPPED lpOvl = &o->o;
67897036	379	DWORD iobytes;
93bcfd20	380	BOOL success = FALSE;
67897036 BC	381	int rc = FIO_Q_COMPLETED;
	382
	383	fio_ro_check(td, io_u);
	384
4e79098f BC	385	lpOvl->Internal = STATUS_PENDING;
	386	lpOvl->InternalHigh = 0;
	387	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	388	lpOvl->OffsetHigh = io_u->offset >> 32;
67897036 BC	389
67897036 BC	390	switch (io_u->ddir) {
40c5db35	391	case DDIR_WRITE:
67897036 BC	392	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	393	break;
	394	case DDIR_READ:
	395	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	396	break;
	397	case DDIR_SYNC:
	398	case DDIR_DATASYNC:
	399	case DDIR_SYNC_FILE_RANGE:
	400	success = FlushFileBuffers(io_u->file->hFile);
	401	if (!success)
2277d5d5	402	io_u->error = win_to_posix_error(GetLastError());
67897036 BC	403
	404	return FIO_Q_COMPLETED;
	405	break;
	406	case DDIR_TRIM:
	407	log_err("manual TRIM isn't supported on Windows");
	408	io_u->error = 1;
	409	io_u->resid = io_u->xfer_buflen;
	410	return FIO_Q_COMPLETED;
	411	break;
	412	default:
	413	assert(0);
93bcfd20	414	break;
67897036 BC	415	}
67897036 BC	416
40c5db35	417	if (success \|\| GetLastError() == ERROR_IO_PENDING)
67897036	418	rc = FIO_Q_QUEUED;
40c5db35	419	else {
2277d5d5	420	io_u->error = win_to_posix_error(GetLastError());
67897036 BC	421	io_u->resid = io_u->xfer_buflen;
	422	}
	423
	424	return rc;
	425	}
	426
	427	/* Runs as a thread and waits for queued IO to complete */
	428	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	429	{
	430	OVERLAPPED *ovl;
	431	struct fio_overlapped *fov;
	432	struct io_u *io_u;
	433	struct windowsaio_data *wd;
	434	struct thread_ctx *ctx;
	435	ULONG_PTR ulKey = 0;
	436	DWORD bytes;
	437
	438	ctx = (struct thread_ctx*)lpParameter;
	439	wd = ctx->wd;
	440
	441	do {
66c098b8	442	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
67897036 BC	443	continue;
	444
	445	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	446	io_u = fov->io_u;
	447
	448	if (ovl->Internal == ERROR_SUCCESS) {
	449	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	450	io_u->error = 0;
	451	} else {
	452	io_u->resid = io_u->xfer_buflen;
2277d5d5	453	io_u->error = win_to_posix_error(GetLastError());
67897036 BC	454	}
67897036 BC	455
40c5db35	456	fov->io_complete = TRUE;
67897036 BC	457	SetEvent(wd->iocomplete_event);
	458	} while (ctx->wd->iothread_running);
	459
	460	CloseHandle(ctx->iocp);
	461	free(ctx);
	462	return 0;
	463	}
	464
	465	static int fio_windowsaio_cancel(struct thread_data *td,
	466	struct io_u *io_u)
	467	{
	468	int rc = 0;
	469
	470	struct windowsaio_data *wd = td->io_ops->data;
	471
	472	/* If we're running on Vista or newer, we can cancel individual IO requests */
	473	if (wd->pCancelIoEx != NULL) {
	474	struct fio_overlapped *ovl = io_u->engine_data;
40c5db35	475
67897036 BC	476	if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
	477	rc = 1;
	478	} else
	479	rc = 1;
	480
	481	return rc;
	482	}
	483
c73ed246 JA	484	static void fio_windowsaio_io_u_free(struct thread_data td, struct io_u io_u)
	485	{
	486	struct fio_overlapped *o = io_u->engine_data;
	487
	488	if (o) {
	489	CloseHandle(o->o.hEvent);
	490	io_u->engine_data = NULL;
	491	free(o);
	492	}
	493	}
	494
	495	static int fio_windowsaio_io_u_init(struct thread_data td, struct io_u io_u)
	496	{
	497	struct fio_overlapped *o;
	498
	499	o = malloc(sizeof(*o));
b0106419	500	o->io_complete = FALSE;
c73ed246 JA	501	o->io_u = io_u;
	502	o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
	503	if (!o->o.hEvent) {
	504	free(o);
	505	return 1;
	506	}
	507
	508	io_u->engine_data = o;
	509	return 0;
	510	}
	511
ecc314ba BC	512	static struct ioengine_ops ioengine = {
	513	.name = "windowsaio",
	514	.version = FIO_IOOPS_VERSION,
	515	.init = fio_windowsaio_init,
	516	.queue = fio_windowsaio_queue,
	517	.cancel = fio_windowsaio_cancel,
	518	.getevents = fio_windowsaio_getevents,
	519	.event = fio_windowsaio_event,
	520	.cleanup = fio_windowsaio_cleanup,
	521	.open_file = fio_windowsaio_open_file,
	522	.close_file = fio_windowsaio_close_file,
c73ed246 JA	523	.get_file_size = generic_get_file_size,
	524	.io_u_init = fio_windowsaio_io_u_init,
	525	.io_u_free = fio_windowsaio_io_u_free,
ecc314ba BC	526	};
	527
	528	static void fio_init fio_posixaio_register(void)
	529	{
	530	register_ioengine(&ioengine);
	531	}
	532
	533	static void fio_exit fio_posixaio_unregister(void)
	534	{
	535	unregister_ioengine(&ioengine);
	536	}