[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 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;
}

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;

	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;

	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;
	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);
		if (hFile == INVALID_HANDLE_VALUE)
			rc = 1;

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

		if (!rc)
			ctx = malloc(sizeof(struct thread_ctx));

		if (!rc && ctx == NULL)
		{
			log_err("fio: out of memory in windowsaio\n");
			CloseHandle(hFile);
			CloseHandle(f->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 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
	1	/*
	2	* Native Windows async IO engine
	3	* Copyright (C) 2012 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	int geterrno_from_win_error (DWORD code, int deferrno);
	18
	19	struct fio_overlapped {
	20	OVERLAPPED o;
	21	struct io_u *io_u;
	22	BOOL io_complete;
	23	};
	24
	25	struct windowsaio_data {
	26	struct io_u **aio_events;
	27	HANDLE iothread;
	28	HANDLE iocomplete_event;
	29	CANCELIOEX pCancelIoEx;
	30	BOOL iothread_running;
	31	};
	32
	33	struct thread_ctx {
	34	HANDLE iocp;
	35	struct windowsaio_data *wd;
	36	};
	37
	38	static int fio_windowsaio_cancel(struct thread_data *td,
	39	struct io_u *io_u);
	40	static BOOL timeout_expired(DWORD start_count, DWORD end_count);
	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	static int win_to_posix_error(DWORD winerr);
	52
	53	static int win_to_posix_error(DWORD winerr)
	54	{
	55	switch (winerr)
	56	{
	57	case ERROR_FILE_NOT_FOUND: return ENOENT;
	58	case ERROR_PATH_NOT_FOUND: return ENOENT;
	59	case ERROR_ACCESS_DENIED: return EACCES;
	60	case ERROR_INVALID_HANDLE: return EBADF;
	61	case ERROR_NOT_ENOUGH_MEMORY: return ENOMEM;
	62	case ERROR_INVALID_DATA: return EINVAL;
	63	case ERROR_OUTOFMEMORY: return ENOMEM;
	64	case ERROR_INVALID_DRIVE: return ENODEV;
	65	case ERROR_NOT_SAME_DEVICE: return EXDEV;
	66	case ERROR_WRITE_PROTECT: return EROFS;
	67	case ERROR_BAD_UNIT: return ENODEV;
	68	case ERROR_SHARING_VIOLATION: return EACCES;
	69	case ERROR_LOCK_VIOLATION: return EACCES;
	70	case ERROR_SHARING_BUFFER_EXCEEDED: return ENOLCK;
	71	case ERROR_HANDLE_DISK_FULL: return ENOSPC;
	72	case ERROR_NOT_SUPPORTED: return ENOSYS;
	73	case ERROR_FILE_EXISTS: return EEXIST;
	74	case ERROR_CANNOT_MAKE: return EPERM;
	75	case ERROR_INVALID_PARAMETER: return EINVAL;
	76	case ERROR_NO_PROC_SLOTS: return EAGAIN;
	77	case ERROR_BROKEN_PIPE: return EPIPE;
	78	case ERROR_OPEN_FAILED: return EIO;
	79	case ERROR_NO_MORE_SEARCH_HANDLES: return ENFILE;
	80	case ERROR_CALL_NOT_IMPLEMENTED: return ENOSYS;
	81	case ERROR_INVALID_NAME: return ENOENT;
	82	case ERROR_WAIT_NO_CHILDREN: return ECHILD;
	83	case ERROR_CHILD_NOT_COMPLETE: return EBUSY;
	84	case ERROR_DIR_NOT_EMPTY: return ENOTEMPTY;
	85	case ERROR_SIGNAL_REFUSED: return EIO;
	86	case ERROR_BAD_PATHNAME: return ENOENT;
	87	case ERROR_SIGNAL_PENDING: return EBUSY;
	88	case ERROR_MAX_THRDS_REACHED: return EAGAIN;
	89	case ERROR_BUSY: return EBUSY;
	90	case ERROR_ALREADY_EXISTS: return EEXIST;
	91	case ERROR_NO_SIGNAL_SENT: return EIO;
	92	case ERROR_FILENAME_EXCED_RANGE: return EINVAL;
	93	case ERROR_META_EXPANSION_TOO_LONG: return EINVAL;
	94	case ERROR_INVALID_SIGNAL_NUMBER: return EINVAL;
	95	case ERROR_THREAD_1_INACTIVE: return EINVAL;
	96	case ERROR_BAD_PIPE: return EINVAL;
	97	case ERROR_PIPE_BUSY: return EBUSY;
	98	case ERROR_NO_DATA: return EPIPE;
	99	case ERROR_MORE_DATA: return EAGAIN;
	100	case ERROR_DIRECTORY: return ENOTDIR;
	101	case ERROR_PIPE_CONNECTED: return EBUSY;
	102	case ERROR_NO_TOKEN: return EINVAL;
	103	case ERROR_PROCESS_ABORTED: return EFAULT;
	104	case ERROR_BAD_DEVICE: return ENODEV;
	105	case ERROR_BAD_USERNAME: return EINVAL;
	106	case ERROR_OPEN_FILES: return EAGAIN;
	107	case ERROR_ACTIVE_CONNECTIONS: return EAGAIN;
	108	case ERROR_DEVICE_IN_USE: return EAGAIN;
	109	case ERROR_INVALID_AT_INTERRUPT_TIME: return EINTR;
	110	case ERROR_IO_DEVICE: return EIO;
	111	case ERROR_NOT_OWNER: return EPERM;
	112	case ERROR_END_OF_MEDIA: return ENOSPC;
	113	case ERROR_EOM_OVERFLOW: return ENOSPC;
	114	case ERROR_BEGINNING_OF_MEDIA: return ESPIPE;
	115	case ERROR_SETMARK_DETECTED: return ESPIPE;
	116	case ERROR_NO_DATA_DETECTED: return ENOSPC;
	117	case ERROR_POSSIBLE_DEADLOCK: return EDEADLOCK;
	118	case ERROR_CRC: return EIO;
	119	case ERROR_NEGATIVE_SEEK: return EINVAL;
	120	case ERROR_DISK_FULL: return ENOSPC;
	121	case ERROR_NOACCESS: return EFAULT;
	122	case ERROR_FILE_INVALID: return ENXIO;
	123	}
	124
	125	return winerr;
	126	}
	127
	128	int sync_file_range(int fd, off64_t offset, off64_t nbytes,
	129	unsigned int flags)
	130	{
	131	errno = ENOSYS;
	132	return -1;
	133	}
	134
	135	static int fio_windowsaio_init(struct thread_data *td)
	136	{
	137	struct windowsaio_data *wd;
	138	HANDLE hKernel32Dll;
	139	int rc = 0;
	140
	141	wd = malloc(sizeof(struct windowsaio_data));
	142	if (wd != NULL)
	143	ZeroMemory(wd, sizeof(struct windowsaio_data));
	144	else
	145	rc = 1;
	146
	147	if (!rc) {
	148	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	149	if (wd->aio_events == NULL)
	150	rc = 1;
	151	}
	152
	153	if (!rc) {
	154	/* Create an auto-reset event */
	155	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	156	if (wd->iocomplete_event == NULL)
	157	rc = 1;
	158	}
	159
	160	if (rc) {
	161	if (wd != NULL) {
	162	if (wd->aio_events != NULL)
	163	free(wd->aio_events);
	164
	165	free(wd);
	166	}
	167	}
	168
	169	hKernel32Dll = GetModuleHandle("kernel32.dll");
	170	wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
	171	td->io_ops->data = wd;
	172
	173	return rc;
	174	}
	175
	176	static void fio_windowsaio_cleanup(struct thread_data *td)
	177	{
	178	struct windowsaio_data *wd;
	179
	180	wd = td->io_ops->data;
	181
	182	if (wd != NULL) {
	183	wd->iothread_running = FALSE;
	184	WaitForSingleObject(wd->iothread, INFINITE);
	185
	186	CloseHandle(wd->iothread);
	187	CloseHandle(wd->iocomplete_event);
	188
	189	free(wd->aio_events);
	190	free(wd);
	191
	192	td->io_ops->data = NULL;
	193	}
	194	}
	195
	196
	197	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	198	{
	199	int rc = 0;
	200	HANDLE hFile;
	201	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
	202	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	203	DWORD openmode = OPEN_ALWAYS;
	204	DWORD access;
	205
	206	dprint(FD_FILE, "fd open %s\n", f->file_name);
	207
	208	if (f->filetype == FIO_TYPE_PIPE) {
	209	log_err("fio: windowsaio doesn't support pipes\n");
	210	return 1;
	211	}
	212
	213	if (!strcmp(f->file_name, "-")) {
	214	log_err("fio: can't read/write to stdin/out\n");
	215	return 1;
	216	}
	217
	218	if (td->o.odirect)
	219	flags \|= FILE_FLAG_NO_BUFFERING;
	220	if (td->o.sync_io)
	221	flags \|= FILE_FLAG_WRITE_THROUGH;
	222
	223	/*
	224	* Inform Windows whether we're going to be doing sequential or
	225	* random io so it can tune the Cache Manager
	226	*/
	227	if (td->o.td_ddir == TD_DDIR_READ \|\|
	228	td->o.td_ddir == TD_DDIR_WRITE)
	229	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
	230	else
	231	flags \|= FILE_FLAG_RANDOM_ACCESS;
	232
	233	if (!td_write(td) \|\| read_only)
	234	access = GENERIC_READ;
	235	else
	236	access = (GENERIC_READ \| GENERIC_WRITE);
	237
	238	if (td->o.create_on_open)
	239	openmode = OPEN_ALWAYS;
	240	else
	241	openmode = OPEN_EXISTING;
	242
	243	f->hFile = CreateFile(f->file_name, access, sharemode,
	244	NULL, openmode, flags, NULL);
	245
	246	if (f->hFile == INVALID_HANDLE_VALUE)
	247	rc = 1;
	248
	249	/* Only set up the completion port and thread if we're not just
	250	* querying the device size */
	251	if (!rc && td->io_ops->data != NULL) {
	252	struct thread_ctx *ctx;
	253	struct windowsaio_data *wd;
	254
	255	hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
	256	if (hFile == INVALID_HANDLE_VALUE)
	257	rc = 1;
	258
	259	wd = td->io_ops->data;
	260	wd->iothread_running = TRUE;
	261
	262	if (!rc)
	263	ctx = malloc(sizeof(struct thread_ctx));
	264
	265	if (!rc && ctx == NULL)
	266	{
	267	log_err("fio: out of memory in windowsaio\n");
	268	CloseHandle(hFile);
	269	CloseHandle(f->hFile);
	270	rc = 1;
	271	}
	272
	273	if (!rc)
	274	{
	275	ctx->iocp = hFile;
	276	ctx->wd = wd;
	277	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	278	}
	279
	280	if (rc \|\| wd->iothread == NULL)
	281	rc = 1;
	282	}
	283
	284	return rc;
	285	}
	286
	287	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	288	{
	289	int rc = 0;
	290
	291	dprint(FD_FILE, "fd close %s\n", f->file_name);
	292
	293	if (f->hFile != INVALID_HANDLE_VALUE) {
	294	if (!CloseHandle(f->hFile))
	295	rc = 1;
	296	}
	297
	298	f->hFile = INVALID_HANDLE_VALUE;
	299	return rc;
	300	}
	301
	302	static BOOL timeout_expired(DWORD start_count, DWORD end_count)
	303	{
	304	BOOL expired = FALSE;
	305	DWORD current_time;
	306
	307	current_time = GetTickCount();
	308
	309	if ((end_count > start_count) && current_time >= end_count)
	310	expired = TRUE;
	311	else if (current_time < start_count && current_time > end_count)
	312	expired = TRUE;
	313
	314	return expired;
	315	}
	316
	317	static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
	318	{
	319	struct windowsaio_data *wd = td->io_ops->data;
	320	return wd->aio_events[event];
	321	}
	322
	323	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	324	unsigned int max, struct timespec *t)
	325	{
	326	struct windowsaio_data *wd = td->io_ops->data;
	327	struct flist_head *entry;
	328	unsigned int dequeued = 0;
	329	struct io_u *io_u;
	330	struct fio_overlapped *fov;
	331	DWORD start_count = 0;
	332	DWORD end_count = 0;
	333	DWORD status;
	334	DWORD mswait = 250;
	335
	336	if (t != NULL) {
	337	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	338	start_count = GetTickCount();
	339	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	340	}
	341
	342	do {
	343	flist_for_each(entry, &td->io_u_busylist) {
	344	io_u = flist_entry(entry, struct io_u, list);
	345	fov = (struct fio_overlapped*)io_u->engine_data;
	346
	347	if (fov->io_complete) {
	348	fov->io_complete = FALSE;
	349	ResetEvent(fov->o.hEvent);
	350	wd->aio_events[dequeued] = io_u;
	351	dequeued++;
	352	}
	353
	354	if (dequeued >= min)
	355	break;
	356	}
	357
	358	if (dequeued < min) {
	359	status = WaitForSingleObject(wd->iocomplete_event, mswait);
	360	if (status != WAIT_OBJECT_0 && dequeued >= min)
	361	break;
	362	}
	363
	364	if (dequeued >= min \|\| (t != NULL && timeout_expired(start_count, end_count)))
	365	break;
	366	} while (1);
	367
	368	return dequeued;
	369	}
	370
	371	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
	372	{
	373	struct fio_overlapped *o = io_u->engine_data;
	374	LPOVERLAPPED lpOvl = &o->o;
	375	DWORD iobytes;
	376	BOOL success = FALSE;
	377	int rc = FIO_Q_COMPLETED;
	378
	379	fio_ro_check(td, io_u);
	380
	381	lpOvl->Internal = STATUS_PENDING;
	382	lpOvl->InternalHigh = 0;
	383	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	384	lpOvl->OffsetHigh = io_u->offset >> 32;
	385
	386	switch (io_u->ddir) {
	387	case DDIR_WRITE:
	388	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	389	break;
	390	case DDIR_READ:
	391	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	392	break;
	393	case DDIR_SYNC:
	394	case DDIR_DATASYNC:
	395	case DDIR_SYNC_FILE_RANGE:
	396	success = FlushFileBuffers(io_u->file->hFile);
	397	if (!success)
	398	io_u->error = win_to_posix_error(GetLastError());
	399
	400	return FIO_Q_COMPLETED;
	401	break;
	402	case DDIR_TRIM:
	403	log_err("manual TRIM isn't supported on Windows");
	404	io_u->error = 1;
	405	io_u->resid = io_u->xfer_buflen;
	406	return FIO_Q_COMPLETED;
	407	break;
	408	default:
	409	assert(0);
	410	break;
	411	}
	412
	413	if (success \|\| GetLastError() == ERROR_IO_PENDING)
	414	rc = FIO_Q_QUEUED;
	415	else {
	416	io_u->error = win_to_posix_error(GetLastError());
	417	io_u->resid = io_u->xfer_buflen;
	418	}
	419
	420	return rc;
	421	}
	422
	423	/* Runs as a thread and waits for queued IO to complete */
	424	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	425	{
	426	OVERLAPPED *ovl;
	427	struct fio_overlapped *fov;
	428	struct io_u *io_u;
	429	struct windowsaio_data *wd;
	430	struct thread_ctx *ctx;
	431	ULONG_PTR ulKey = 0;
	432	DWORD bytes;
	433
	434	ctx = (struct thread_ctx*)lpParameter;
	435	wd = ctx->wd;
	436
	437	do {
	438	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
	439	continue;
	440
	441	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	442	io_u = fov->io_u;
	443
	444	if (ovl->Internal == ERROR_SUCCESS) {
	445	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	446	io_u->error = 0;
	447	} else {
	448	io_u->resid = io_u->xfer_buflen;
	449	io_u->error = win_to_posix_error(GetLastError());
	450	}
	451
	452	fov->io_complete = TRUE;
	453	SetEvent(wd->iocomplete_event);
	454	} while (ctx->wd->iothread_running);
	455
	456	CloseHandle(ctx->iocp);
	457	free(ctx);
	458	return 0;
	459	}
	460
	461	static int fio_windowsaio_cancel(struct thread_data *td,
	462	struct io_u *io_u)
	463	{
	464	int rc = 0;
	465
	466	struct windowsaio_data *wd = td->io_ops->data;
	467
	468	/* If we're running on Vista or newer, we can cancel individual IO requests */
	469	if (wd->pCancelIoEx != NULL) {
	470	struct fio_overlapped *ovl = io_u->engine_data;
	471
	472	if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
	473	rc = 1;
	474	} else
	475	rc = 1;
	476
	477	return rc;
	478	}
	479
	480	static void fio_windowsaio_io_u_free(struct thread_data td, struct io_u io_u)
	481	{
	482	struct fio_overlapped *o = io_u->engine_data;
	483
	484	if (o) {
	485	CloseHandle(o->o.hEvent);
	486	io_u->engine_data = NULL;
	487	free(o);
	488	}
	489	}
	490
	491	static int fio_windowsaio_io_u_init(struct thread_data td, struct io_u io_u)
	492	{
	493	struct fio_overlapped *o;
	494
	495	o = malloc(sizeof(*o));
	496	o->io_complete = FALSE:
	497	o->io_u = io_u;
	498	o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
	499	if (!o->o.hEvent) {
	500	free(o);
	501	return 1;
	502	}
	503
	504	io_u->engine_data = o;
	505	return 0;
	506	}
	507
	508	static struct ioengine_ops ioengine = {
	509	.name = "windowsaio",
	510	.version = FIO_IOOPS_VERSION,
	511	.init = fio_windowsaio_init,
	512	.queue = fio_windowsaio_queue,
	513	.cancel = fio_windowsaio_cancel,
	514	.getevents = fio_windowsaio_getevents,
	515	.event = fio_windowsaio_event,
	516	.cleanup = fio_windowsaio_cleanup,
	517	.open_file = fio_windowsaio_open_file,
	518	.close_file = fio_windowsaio_close_file,
	519	.get_file_size = generic_get_file_size,
	520	.io_u_init = fio_windowsaio_io_u_init,
	521	.io_u_free = fio_windowsaio_io_u_free,
	522	};
	523
	524	static void fio_init fio_posixaio_register(void)
	525	{
	526	register_ioengine(&ioengine);
	527	}
	528
	529	static void fio_exit fio_posixaio_unregister(void)
	530	{
	531	unregister_ioengine(&ioengine);
	532	}