[fio.git] / engines / windowsaio.c

/*
 * windowsaio engine
 *
 * IO engine using Windows IO Completion Ports.
 */

#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 = calloc(1, sizeof(struct windowsaio_data));
	if (wd == NULL) {
		 log_err("windowsaio: failed to allocate memory for engine data\n");
		rc = 1;
	}

	if (!rc) {
		wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
		if (wd->aio_events == NULL) {
			log_err("windowsaio: failed to allocate memory for aio events list\n");
			rc = 1;
		}
	}

	if (!rc) {
		/* Create an auto-reset event */
		wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
		if (wd->iocomplete_event == NULL) {
			log_err("windowsaio: failed to create io complete event handle\n");
			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) {
			log_err("windowsaio: failed to create io completion port\n");
			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("windowsaio: failed to allocate memory for thread context structure\n");
			CloseHandle(hFile);
			rc = 1;
		}

		if (!rc)
		{
			ctx->iocp = hFile;
			ctx->wd = wd;
			wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
			if (wd->iothread == NULL)
				log_err("windowsaio: failed to create io completion thread\n");
		}

		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("windowsaio: pipes are not supported\n");
		return 1;
	}

	if (!strcmp(f->file_name, "-")) {
		log_err("windowsaio: 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) {
		log_err("windowsaio: failed to open file \"%s\"\n", f->file_name);
		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) {
			log_err("windowsaio: failed to create io completion port\n");
			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)) {
			log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name);
			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) {
			log_err("windowsaio: failed to flush file buffers\n");
			io_u->error = win_to_posix_error(GetLastError());
		}

		return FIO_Q_COMPLETED;
		break;
	case DDIR_TRIM:
		log_err("windowsaio: manual TRIM isn't supported on Windows\n");
		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)) {
			log_err("windowsaio: failed to cancel io\n");
			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 == NULL) {
		log_err("windowsaio: failed to create event handle\n");
		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_windowsaio_register(void)
{
	register_ioengine(&ioengine);
}

static void fio_exit fio_windowsaio_unregister(void)
{
	unregister_ioengine(&ioengine);
}
Commit	Line	Data
ecc314ba	1	/*
03244c19 BC	2	* windowsaio engine
	3	*
	4	* IO engine using Windows IO Completion Ports.
ecc314ba BC	5	*/
ecc314ba BC	6
ecc314ba BC	7	#include <stdio.h>
	8	#include <stdlib.h>
	9	#include <unistd.h>
	10	#include <signal.h>
	11	#include <errno.h>
	12	#include <windows.h>
	13
	14	#include "../fio.h"
	15
ea4500d8 BC	16	typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);
ea4500d8 BC	17
66c098b8 BC	18	int geterrno_from_win_error (DWORD code, int deferrno);
66c098b8 BC	19
67897036 BC	20	struct fio_overlapped {
	21	OVERLAPPED o;
	22	struct io_u *io_u;
	23	BOOL io_complete;
67897036	24	};
ecc314ba	25
9b836561	26	struct windowsaio_data {
9b836561	27	struct io_u **aio_events;
5a90bb5f	28	HANDLE iocp;
67897036	29	HANDLE iothread;
9b836561	30	HANDLE iocomplete_event;
ea4500d8	31	CANCELIOEX pCancelIoEx;
67897036	32	BOOL iothread_running;
9b836561 BC	33	};
9b836561 BC	34
ecc314ba	35	struct thread_ctx {
9b836561	36	HANDLE iocp;
ecc314ba BC	37	struct windowsaio_data *wd;
	38	};
	39
ecc314ba	40	static int fio_windowsaio_cancel(struct thread_data *td,
03244c19	41	struct io_u *io_u);
67897036	42	static BOOL timeout_expired(DWORD start_count, DWORD end_count);
ecc314ba	43	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
03244c19	44	unsigned int max, struct timespec *t);
ecc314ba BC	45	static struct io_u fio_windowsaio_event(struct thread_data td, int event);
ecc314ba BC	46	static int fio_windowsaio_queue(struct thread_data *td,
03244c19	47	struct io_u *io_u);
ecc314ba BC	48	static void fio_windowsaio_cleanup(struct thread_data *td);
	49	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
	50	static int fio_windowsaio_init(struct thread_data *td);
	51	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f);
	52	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f);
2277d5d5	53	static int win_to_posix_error(DWORD winerr);
66c098b8	54
2277d5d5	55	static int win_to_posix_error(DWORD winerr)
66c098b8 BC	56	{
	57	switch (winerr)
	58	{
	59	case ERROR_FILE_NOT_FOUND: return ENOENT;
	60	case ERROR_PATH_NOT_FOUND: return ENOENT;
	61	case ERROR_ACCESS_DENIED: return EACCES;
	62	case ERROR_INVALID_HANDLE: return EBADF;
	63	case ERROR_NOT_ENOUGH_MEMORY: return ENOMEM;
	64	case ERROR_INVALID_DATA: return EINVAL;
	65	case ERROR_OUTOFMEMORY: return ENOMEM;
	66	case ERROR_INVALID_DRIVE: return ENODEV;
	67	case ERROR_NOT_SAME_DEVICE: return EXDEV;
	68	case ERROR_WRITE_PROTECT: return EROFS;
	69	case ERROR_BAD_UNIT: return ENODEV;
	70	case ERROR_SHARING_VIOLATION: return EACCES;
	71	case ERROR_LOCK_VIOLATION: return EACCES;
	72	case ERROR_SHARING_BUFFER_EXCEEDED: return ENOLCK;
	73	case ERROR_HANDLE_DISK_FULL: return ENOSPC;
	74	case ERROR_NOT_SUPPORTED: return ENOSYS;
	75	case ERROR_FILE_EXISTS: return EEXIST;
	76	case ERROR_CANNOT_MAKE: return EPERM;
	77	case ERROR_INVALID_PARAMETER: return EINVAL;
	78	case ERROR_NO_PROC_SLOTS: return EAGAIN;
	79	case ERROR_BROKEN_PIPE: return EPIPE;
	80	case ERROR_OPEN_FAILED: return EIO;
	81	case ERROR_NO_MORE_SEARCH_HANDLES: return ENFILE;
	82	case ERROR_CALL_NOT_IMPLEMENTED: return ENOSYS;
	83	case ERROR_INVALID_NAME: return ENOENT;
	84	case ERROR_WAIT_NO_CHILDREN: return ECHILD;
	85	case ERROR_CHILD_NOT_COMPLETE: return EBUSY;
	86	case ERROR_DIR_NOT_EMPTY: return ENOTEMPTY;
	87	case ERROR_SIGNAL_REFUSED: return EIO;
	88	case ERROR_BAD_PATHNAME: return ENOENT;
	89	case ERROR_SIGNAL_PENDING: return EBUSY;
	90	case ERROR_MAX_THRDS_REACHED: return EAGAIN;
	91	case ERROR_BUSY: return EBUSY;
	92	case ERROR_ALREADY_EXISTS: return EEXIST;
	93	case ERROR_NO_SIGNAL_SENT: return EIO;
	94	case ERROR_FILENAME_EXCED_RANGE: return EINVAL;
	95	case ERROR_META_EXPANSION_TOO_LONG: return EINVAL;
	96	case ERROR_INVALID_SIGNAL_NUMBER: return EINVAL;
	97	case ERROR_THREAD_1_INACTIVE: return EINVAL;
	98	case ERROR_BAD_PIPE: return EINVAL;
	99	case ERROR_PIPE_BUSY: return EBUSY;
	100	case ERROR_NO_DATA: return EPIPE;
	101	case ERROR_MORE_DATA: return EAGAIN;
	102	case ERROR_DIRECTORY: return ENOTDIR;
	103	case ERROR_PIPE_CONNECTED: return EBUSY;
	104	case ERROR_NO_TOKEN: return EINVAL;
	105	case ERROR_PROCESS_ABORTED: return EFAULT;
	106	case ERROR_BAD_DEVICE: return ENODEV;
	107	case ERROR_BAD_USERNAME: return EINVAL;
	108	case ERROR_OPEN_FILES: return EAGAIN;
	109	case ERROR_ACTIVE_CONNECTIONS: return EAGAIN;
	110	case ERROR_DEVICE_IN_USE: return EAGAIN;
	111	case ERROR_INVALID_AT_INTERRUPT_TIME: return EINTR;
	112	case ERROR_IO_DEVICE: return EIO;
	113	case ERROR_NOT_OWNER: return EPERM;
	114	case ERROR_END_OF_MEDIA: return ENOSPC;
	115	case ERROR_EOM_OVERFLOW: return ENOSPC;
	116	case ERROR_BEGINNING_OF_MEDIA: return ESPIPE;
	117	case ERROR_SETMARK_DETECTED: return ESPIPE;
	118	case ERROR_NO_DATA_DETECTED: return ENOSPC;
	119	case ERROR_POSSIBLE_DEADLOCK: return EDEADLOCK;
120	case ERROR_CRC: return EIO;
121	case ERROR_NEGATIVE_SEEK: return EINVAL;
122	case ERROR_DISK_FULL: return ENOSPC;
123	case ERROR_NOACCESS: return EFAULT;
124	case ERROR_FILE_INVALID: return ENXIO;
125	}
126
127	return winerr;
128	}
ecc314ba	129
ecc314ba BC	130	static int fio_windowsaio_init(struct thread_data *td)
ecc314ba BC	131	{
ecc314ba	132	struct windowsaio_data *wd;
ea4500d8	133	HANDLE hKernel32Dll;
9b836561	134	int rc = 0;
ecc314ba	135
03244c19 BC	136	wd = calloc(1, sizeof(struct windowsaio_data));
	137	if (wd == NULL) {
	138	log_err("windowsaio: failed to allocate memory for engine data\n");
9b836561	139	rc = 1;
03244c19	140	}
ecc314ba	141
9b836561 BC	142	if (!rc) {
9b836561 BC	143	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
03244c19 BC	144	if (wd->aio_events == NULL) {
03244c19 BC	145	log_err("windowsaio: failed to allocate memory for aio events list\n");
9b836561	146	rc = 1;
03244c19	147	}
e4db9fec BC	148	}
e4db9fec BC	149
9b836561 BC	150	if (!rc) {
	151	/* Create an auto-reset event */
	152	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
03244c19 BC	153	if (wd->iocomplete_event == NULL) {
03244c19 BC	154	log_err("windowsaio: failed to create io complete event handle\n");
9b836561	155	rc = 1;
03244c19	156	}
9b836561 BC	157	}
9b836561 BC	158
9b836561	159	if (rc) {
9b836561	160	if (wd != NULL) {
9b836561 BC	161	if (wd->aio_events != NULL)
	162	free(wd->aio_events);
	163
	164	free(wd);
	165	}
	166	}
ecc314ba	167
ea4500d8	168	hKernel32Dll = GetModuleHandle("kernel32.dll");
93bcfd20	169	wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
ecc314ba	170	td->io_ops->data = wd;
93bcfd20	171
5a90bb5f BC	172	if (!rc) {
	173	struct thread_ctx *ctx;
	174	struct windowsaio_data *wd;
	175	HANDLE hFile;
	176
	177	hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0);
03244c19 BC	178	if (hFile == INVALID_HANDLE_VALUE) {
03244c19 BC	179	log_err("windowsaio: failed to create io completion port\n");
5a90bb5f	180	rc = 1;
03244c19	181	}
5a90bb5f BC	182
	183	wd = td->io_ops->data;
	184	wd->iothread_running = TRUE;
	185	wd->iocp = hFile;
	186
	187	if (!rc)
	188	ctx = malloc(sizeof(struct thread_ctx));
	189
	190	if (!rc && ctx == NULL)
	191	{
03244c19	192	log_err("windowsaio: failed to allocate memory for thread context structure\n");
5a90bb5f BC	193	CloseHandle(hFile);
	194	rc = 1;
	195	}
	196
	197	if (!rc)
	198	{
	199	ctx->iocp = hFile;
	200	ctx->wd = wd;
	201	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
03244c19 BC	202	if (wd->iothread == NULL)
03244c19 BC	203	log_err("windowsaio: failed to create io completion thread\n");
5a90bb5f BC	204	}
	205
	206	if (rc \|\| wd->iothread == NULL)
	207	rc = 1;
	208	}
	209
4e79098f	210	return rc;
ecc314ba BC	211	}
ecc314ba BC	212
67897036 BC	213	static void fio_windowsaio_cleanup(struct thread_data *td)
67897036 BC	214	{
67897036 BC	215	struct windowsaio_data *wd;
	216
	217	wd = td->io_ops->data;
	218
	219	if (wd != NULL) {
40c5db35 JA	220	wd->iothread_running = FALSE;
40c5db35 JA	221	WaitForSingleObject(wd->iothread, INFINITE);
67897036 BC	222
	223	CloseHandle(wd->iothread);
	224	CloseHandle(wd->iocomplete_event);
	225
67897036	226	free(wd->aio_events);
67897036 BC	227	free(wd);
	228
	229	td->io_ops->data = NULL;
	230	}
	231	}
	232
	233
ecc314ba BC	234	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	235	{
	236	int rc = 0;
4e79098f	237	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
ecc314ba BC	238	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	239	DWORD openmode = OPEN_ALWAYS;
	240	DWORD access;
	241
	242	dprint(FD_FILE, "fd open %s\n", f->file_name);
	243
ecc314ba	244	if (f->filetype == FIO_TYPE_PIPE) {
03244c19	245	log_err("windowsaio: pipes are not supported\n");
ecc314ba BC	246	return 1;
	247	}
	248
	249	if (!strcmp(f->file_name, "-")) {
03244c19	250	log_err("windowsaio: can't read/write to stdin/out\n");
ecc314ba BC	251	return 1;
	252	}
	253
	254	if (td->o.odirect)
	255	flags \|= FILE_FLAG_NO_BUFFERING;
	256	if (td->o.sync_io)
	257	flags \|= FILE_FLAG_WRITE_THROUGH;
	258
93bcfd20 BC	259	/*
	260	* Inform Windows whether we're going to be doing sequential or
	261	* random io so it can tune the Cache Manager
	262	*/
ecc314ba	263	if (td->o.td_ddir == TD_DDIR_READ \|\|
ea4500d8	264	td->o.td_ddir == TD_DDIR_WRITE)
ecc314ba	265	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
ecc314ba	266	else
ecc314ba	267	flags \|= FILE_FLAG_RANDOM_ACCESS;
ecc314ba	268
ea4500d8	269	if (!td_write(td) \|\| read_only)
ecc314ba BC	270	access = GENERIC_READ;
	271	else
	272	access = (GENERIC_READ \| GENERIC_WRITE);
	273
93bcfd20	274	if (td->o.create_on_open)
ecc314ba BC	275	openmode = OPEN_ALWAYS;
	276	else
	277	openmode = OPEN_EXISTING;
	278
	279	f->hFile = CreateFile(f->file_name, access, sharemode,
	280	NULL, openmode, flags, NULL);
	281
03244c19 BC	282	if (f->hFile == INVALID_HANDLE_VALUE) {
03244c19 BC	283	log_err("windowsaio: failed to open file \"%s\"\n", f->file_name);
ecc314ba	284	rc = 1;
03244c19	285	}
ecc314ba	286
93bcfd20	287	/* Only set up the completion port and thread if we're not just
ecc314ba	288	* querying the device size */
40c5db35	289	if (!rc && td->io_ops->data != NULL) {
40c5db35	290	struct windowsaio_data *wd;
ecc314ba	291
40c5db35	292	wd = td->io_ops->data;
ecc314ba	293
03244c19 BC	294	if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL) {
03244c19 BC	295	log_err("windowsaio: failed to create io completion port\n");
ecc314ba	296	rc = 1;
03244c19	297	}
ecc314ba BC	298	}
ecc314ba BC	299
ecc314ba BC	300	return rc;
	301	}
	302
	303	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	304	{
4e79098f	305	int rc = 0;
1e42cc3d	306
9b836561 BC	307	dprint(FD_FILE, "fd close %s\n", f->file_name);
9b836561 BC	308
ecc314ba	309	if (f->hFile != INVALID_HANDLE_VALUE) {
03244c19 BC	310	if (!CloseHandle(f->hFile)) {
03244c19 BC	311	log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name);
4e79098f	312	rc = 1;
03244c19	313	}
ecc314ba BC	314	}
	315
	316	f->hFile = INVALID_HANDLE_VALUE;
4e79098f	317	return rc;
ecc314ba BC	318	}
ecc314ba BC	319
67897036 BC	320	static BOOL timeout_expired(DWORD start_count, DWORD end_count)
	321	{
	322	BOOL expired = FALSE;
	323	DWORD current_time;
	324
	325	current_time = GetTickCount();
	326
	327	if ((end_count > start_count) && current_time >= end_count)
	328	expired = TRUE;
	329	else if (current_time < start_count && current_time > end_count)
	330	expired = TRUE;
	331
	332	return expired;
	333	}
	334
	335	static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
	336	{
	337	struct windowsaio_data *wd = td->io_ops->data;
	338	return wd->aio_events[event];
	339	}
	340
	341	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
03244c19	342	unsigned int max, struct timespec *t)
67897036 BC	343	{
	344	struct windowsaio_data *wd = td->io_ops->data;
	345	struct flist_head *entry;
	346	unsigned int dequeued = 0;
	347	struct io_u *io_u;
	348	struct fio_overlapped *fov;
	349	DWORD start_count = 0;
	350	DWORD end_count = 0;
	351	DWORD status;
	352	DWORD mswait = 250;
	353
	354	if (t != NULL) {
	355	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	356	start_count = GetTickCount();
	357	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	358	}
	359
	360	do {
	361	flist_for_each(entry, &td->io_u_busylist) {
	362	io_u = flist_entry(entry, struct io_u, list);
	363	fov = (struct fio_overlapped*)io_u->engine_data;
	364
	365	if (fov->io_complete) {
40c5db35	366	fov->io_complete = FALSE;
f9a58c2a	367	ResetEvent(fov->o.hEvent);
67897036 BC	368	wd->aio_events[dequeued] = io_u;
	369	dequeued++;
	370	}
	371
	372	if (dequeued >= min)
	373	break;
	374	}
	375
40c5db35	376	if (dequeued < min) {
67897036	377	status = WaitForSingleObject(wd->iocomplete_event, mswait);
f9a58c2a	378	if (status != WAIT_OBJECT_0 && dequeued >= min)
03244c19	379	break;
67897036 BC	380	}
	381
	382	if (dequeued >= min \|\| (t != NULL && timeout_expired(start_count, end_count)))
	383	break;
	384	} while (1);
	385
	386	return dequeued;
	387	}
	388
40c5db35	389	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
67897036	390	{
c73ed246 JA	391	struct fio_overlapped *o = io_u->engine_data;
c73ed246 JA	392	LPOVERLAPPED lpOvl = &o->o;
67897036	393	DWORD iobytes;
93bcfd20	394	BOOL success = FALSE;
67897036 BC	395	int rc = FIO_Q_COMPLETED;
	396
	397	fio_ro_check(td, io_u);
	398
4e79098f BC	399	lpOvl->Internal = STATUS_PENDING;
	400	lpOvl->InternalHigh = 0;
	401	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	402	lpOvl->OffsetHigh = io_u->offset >> 32;
67897036 BC	403
67897036 BC	404	switch (io_u->ddir) {
40c5db35	405	case DDIR_WRITE:
67897036 BC	406	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	407	break;
	408	case DDIR_READ:
	409	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	410	break;
	411	case DDIR_SYNC:
	412	case DDIR_DATASYNC:
	413	case DDIR_SYNC_FILE_RANGE:
	414	success = FlushFileBuffers(io_u->file->hFile);
03244c19 BC	415	if (!success) {
	416	log_err("windowsaio: failed to flush file buffers\n");
	417	io_u->error = win_to_posix_error(GetLastError());
	418	}
67897036 BC	419
	420	return FIO_Q_COMPLETED;
	421	break;
	422	case DDIR_TRIM:
03244c19	423	log_err("windowsaio: manual TRIM isn't supported on Windows\n");
67897036 BC	424	io_u->error = 1;
	425	io_u->resid = io_u->xfer_buflen;
	426	return FIO_Q_COMPLETED;
	427	break;
	428	default:
	429	assert(0);
93bcfd20	430	break;
67897036 BC	431	}
67897036 BC	432
40c5db35	433	if (success \|\| GetLastError() == ERROR_IO_PENDING)
67897036	434	rc = FIO_Q_QUEUED;
40c5db35	435	else {
2277d5d5	436	io_u->error = win_to_posix_error(GetLastError());
67897036 BC	437	io_u->resid = io_u->xfer_buflen;
	438	}
	439
	440	return rc;
	441	}
	442
	443	/* Runs as a thread and waits for queued IO to complete */
	444	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	445	{
	446	OVERLAPPED *ovl;
	447	struct fio_overlapped *fov;
	448	struct io_u *io_u;
	449	struct windowsaio_data *wd;
	450	struct thread_ctx *ctx;
	451	ULONG_PTR ulKey = 0;
	452	DWORD bytes;
	453
	454	ctx = (struct thread_ctx*)lpParameter;
	455	wd = ctx->wd;
	456
	457	do {
66c098b8	458	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
67897036 BC	459	continue;
	460
	461	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	462	io_u = fov->io_u;
	463
	464	if (ovl->Internal == ERROR_SUCCESS) {
	465	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	466	io_u->error = 0;
	467	} else {
	468	io_u->resid = io_u->xfer_buflen;
2277d5d5	469	io_u->error = win_to_posix_error(GetLastError());
67897036 BC	470	}
67897036 BC	471
40c5db35	472	fov->io_complete = TRUE;
67897036 BC	473	SetEvent(wd->iocomplete_event);
	474	} while (ctx->wd->iothread_running);
	475
	476	CloseHandle(ctx->iocp);
	477	free(ctx);
	478	return 0;
	479	}
	480
	481	static int fio_windowsaio_cancel(struct thread_data *td,
03244c19	482	struct io_u *io_u)
67897036 BC	483	{
	484	int rc = 0;
	485
	486	struct windowsaio_data *wd = td->io_ops->data;
	487
	488	/* If we're running on Vista or newer, we can cancel individual IO requests */
	489	if (wd->pCancelIoEx != NULL) {
	490	struct fio_overlapped *ovl = io_u->engine_data;
40c5db35	491
03244c19 BC	492	if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o)) {
03244c19 BC	493	log_err("windowsaio: failed to cancel io\n");
67897036	494	rc = 1;
03244c19	495	}
67897036 BC	496	} else
	497	rc = 1;
	498
	499	return rc;
	500	}
	501
c73ed246 JA	502	static void fio_windowsaio_io_u_free(struct thread_data td, struct io_u io_u)
	503	{
	504	struct fio_overlapped *o = io_u->engine_data;
	505
	506	if (o) {
	507	CloseHandle(o->o.hEvent);
	508	io_u->engine_data = NULL;
	509	free(o);
	510	}
	511	}
	512
	513	static int fio_windowsaio_io_u_init(struct thread_data td, struct io_u io_u)
	514	{
	515	struct fio_overlapped *o;
	516
	517	o = malloc(sizeof(*o));
b0106419	518	o->io_complete = FALSE;
c73ed246 JA	519	o->io_u = io_u;
c73ed246 JA	520	o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
03244c19 BC	521	if (o->o.hEvent == NULL) {
03244c19 BC	522	log_err("windowsaio: failed to create event handle\n");
c73ed246 JA	523	free(o);
	524	return 1;
	525	}
	526
	527	io_u->engine_data = o;
	528	return 0;
	529	}
	530
ecc314ba BC	531	static struct ioengine_ops ioengine = {
	532	.name = "windowsaio",
	533	.version = FIO_IOOPS_VERSION,
	534	.init = fio_windowsaio_init,
	535	.queue = fio_windowsaio_queue,
	536	.cancel = fio_windowsaio_cancel,
	537	.getevents = fio_windowsaio_getevents,
	538	.event = fio_windowsaio_event,
	539	.cleanup = fio_windowsaio_cleanup,
	540	.open_file = fio_windowsaio_open_file,
	541	.close_file = fio_windowsaio_close_file,
c73ed246 JA	542	.get_file_size = generic_get_file_size,
	543	.io_u_init = fio_windowsaio_io_u_init,
	544	.io_u_free = fio_windowsaio_io_u_free,
ecc314ba BC	545	};
ecc314ba BC	546
c874d188	547	static void fio_init fio_windowsaio_register(void)
ecc314ba BC	548	{
	549	register_ioengine(&ioengine);
	550	}
	551
c874d188	552	static void fio_exit fio_windowsaio_unregister(void)
ecc314ba BC	553	{
	554	unregister_ioengine(&ioengine);
	555	}