[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;
}

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;


	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
9b836561 BC	129	int sync_file_range(int fd, off64_t offset, off64_t nbytes,
	130	unsigned int flags)
	131	{
	132	errno = ENOSYS;
	133	return -1;
	134	}
	135
ecc314ba BC	136	static int fio_windowsaio_init(struct thread_data *td)
ecc314ba BC	137	{
ecc314ba	138	struct windowsaio_data *wd;
ea4500d8	139	HANDLE hKernel32Dll;
9b836561	140	int rc = 0;
ecc314ba	141
ecc314ba	142	wd = malloc(sizeof(struct windowsaio_data));
9b836561 BC	143	if (wd != NULL)
	144	ZeroMemory(wd, sizeof(struct windowsaio_data));
	145	else
	146	rc = 1;
ecc314ba	147
9b836561 BC	148	if (!rc) {
	149	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	150	if (wd->aio_events == NULL)
	151	rc = 1;
e4db9fec BC	152	}
e4db9fec BC	153
9b836561 BC	154	if (!rc) {
	155	/* Create an auto-reset event */
	156	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	157	if (wd->iocomplete_event == NULL)
	158	rc = 1;
	159	}
	160
9b836561	161	if (rc) {
9b836561	162	if (wd != NULL) {
9b836561 BC	163	if (wd->aio_events != NULL)
	164	free(wd->aio_events);
	165
	166	free(wd);
	167	}
	168	}
ecc314ba	169
ea4500d8	170	hKernel32Dll = GetModuleHandle("kernel32.dll");
93bcfd20	171	wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
ecc314ba	172	td->io_ops->data = wd;
93bcfd20	173
5a90bb5f BC	174
	175	if (!rc) {
	176	struct thread_ctx *ctx;
	177	struct windowsaio_data *wd;
	178	HANDLE hFile;
	179
	180	hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0);
	181	if (hFile == INVALID_HANDLE_VALUE)
	182	rc = 1;
	183
	184	wd = td->io_ops->data;
	185	wd->iothread_running = TRUE;
	186	wd->iocp = hFile;
	187
	188	if (!rc)
	189	ctx = malloc(sizeof(struct thread_ctx));
	190
	191	if (!rc && ctx == NULL)
	192	{
	193	log_err("fio: out of memory in windowsaio\n");
	194	CloseHandle(hFile);
	195	rc = 1;
	196	}
	197
	198	if (!rc)
	199	{
	200	ctx->iocp = hFile;
	201	ctx->wd = wd;
	202	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	203	}
	204
	205	if (rc \|\| wd->iothread == NULL)
	206	rc = 1;
	207	}
	208
4e79098f	209	return rc;
ecc314ba BC	210	}
ecc314ba BC	211
67897036 BC	212	static void fio_windowsaio_cleanup(struct thread_data *td)
67897036 BC	213	{
67897036 BC	214	struct windowsaio_data *wd;
	215
	216	wd = td->io_ops->data;
	217
	218	if (wd != NULL) {
40c5db35 JA	219	wd->iothread_running = FALSE;
40c5db35 JA	220	WaitForSingleObject(wd->iothread, INFINITE);
67897036 BC	221
	222	CloseHandle(wd->iothread);
	223	CloseHandle(wd->iocomplete_event);
	224
67897036	225	free(wd->aio_events);
67897036 BC	226	free(wd);
	227
	228	td->io_ops->data = NULL;
	229	}
	230	}
	231
	232
ecc314ba BC	233	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	234	{
	235	int rc = 0;
4e79098f	236	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
ecc314ba BC	237	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	238	DWORD openmode = OPEN_ALWAYS;
	239	DWORD access;
	240
	241	dprint(FD_FILE, "fd open %s\n", f->file_name);
	242
ecc314ba BC	243	if (f->filetype == FIO_TYPE_PIPE) {
	244	log_err("fio: windowsaio doesn't support pipes\n");
	245	return 1;
	246	}
	247
	248	if (!strcmp(f->file_name, "-")) {
	249	log_err("fio: can't read/write to stdin/out\n");
	250	return 1;
	251	}
	252
	253	if (td->o.odirect)
	254	flags \|= FILE_FLAG_NO_BUFFERING;
	255	if (td->o.sync_io)
	256	flags \|= FILE_FLAG_WRITE_THROUGH;
	257
93bcfd20 BC	258	/*
	259	* Inform Windows whether we're going to be doing sequential or
	260	* random io so it can tune the Cache Manager
	261	*/
ecc314ba	262	if (td->o.td_ddir == TD_DDIR_READ \|\|
ea4500d8	263	td->o.td_ddir == TD_DDIR_WRITE)
ecc314ba	264	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
ecc314ba	265	else
ecc314ba	266	flags \|= FILE_FLAG_RANDOM_ACCESS;
ecc314ba	267
ea4500d8	268	if (!td_write(td) \|\| read_only)
ecc314ba BC	269	access = GENERIC_READ;
	270	else
	271	access = (GENERIC_READ \| GENERIC_WRITE);
	272
93bcfd20	273	if (td->o.create_on_open)
ecc314ba BC	274	openmode = OPEN_ALWAYS;
	275	else
	276	openmode = OPEN_EXISTING;
	277
	278	f->hFile = CreateFile(f->file_name, access, sharemode,
	279	NULL, openmode, flags, NULL);
	280
4e79098f	281	if (f->hFile == INVALID_HANDLE_VALUE)
ecc314ba	282	rc = 1;
ecc314ba	283
93bcfd20	284	/* Only set up the completion port and thread if we're not just
ecc314ba	285	* querying the device size */
40c5db35	286	if (!rc && td->io_ops->data != NULL) {
40c5db35	287	struct windowsaio_data *wd;
ecc314ba	288
40c5db35	289	wd = td->io_ops->data;
ecc314ba	290
5a90bb5f	291	if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL)
ecc314ba	292	rc = 1;
ecc314ba BC	293	}
ecc314ba BC	294
ecc314ba BC	295	return rc;
	296	}
	297
	298	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	299	{
4e79098f	300	int rc = 0;
1e42cc3d	301
9b836561 BC	302	dprint(FD_FILE, "fd close %s\n", f->file_name);
9b836561 BC	303
ecc314ba	304	if (f->hFile != INVALID_HANDLE_VALUE) {
9b836561	305	if (!CloseHandle(f->hFile))
4e79098f	306	rc = 1;
ecc314ba BC	307	}
	308
	309	f->hFile = INVALID_HANDLE_VALUE;
4e79098f	310	return rc;
ecc314ba BC	311	}
ecc314ba BC	312
67897036 BC	313	static BOOL timeout_expired(DWORD start_count, DWORD end_count)
	314	{
	315	BOOL expired = FALSE;
	316	DWORD current_time;
	317
	318	current_time = GetTickCount();
	319
	320	if ((end_count > start_count) && current_time >= end_count)
	321	expired = TRUE;
	322	else if (current_time < start_count && current_time > end_count)
	323	expired = TRUE;
	324
	325	return expired;
	326	}
	327
	328	static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
	329	{
	330	struct windowsaio_data *wd = td->io_ops->data;
	331	return wd->aio_events[event];
	332	}
	333
	334	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	335	unsigned int max, struct timespec *t)
	336	{
	337	struct windowsaio_data *wd = td->io_ops->data;
	338	struct flist_head *entry;
	339	unsigned int dequeued = 0;
	340	struct io_u *io_u;
	341	struct fio_overlapped *fov;
	342	DWORD start_count = 0;
	343	DWORD end_count = 0;
	344	DWORD status;
	345	DWORD mswait = 250;
	346
	347	if (t != NULL) {
	348	mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	349	start_count = GetTickCount();
	350	end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	351	}
	352
	353	do {
	354	flist_for_each(entry, &td->io_u_busylist) {
	355	io_u = flist_entry(entry, struct io_u, list);
	356	fov = (struct fio_overlapped*)io_u->engine_data;
	357
	358	if (fov->io_complete) {
40c5db35	359	fov->io_complete = FALSE;
f9a58c2a	360	ResetEvent(fov->o.hEvent);
67897036 BC	361	wd->aio_events[dequeued] = io_u;
	362	dequeued++;
	363	}
	364
	365	if (dequeued >= min)
	366	break;
	367	}
	368
40c5db35	369	if (dequeued < min) {
67897036	370	status = WaitForSingleObject(wd->iocomplete_event, mswait);
f9a58c2a	371	if (status != WAIT_OBJECT_0 && dequeued >= min)
67897036 BC	372	break;
	373	}
	374
	375	if (dequeued >= min \|\| (t != NULL && timeout_expired(start_count, end_count)))
	376	break;
	377	} while (1);
	378
	379	return dequeued;
	380	}
	381
40c5db35	382	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
67897036	383	{
c73ed246 JA	384	struct fio_overlapped *o = io_u->engine_data;
c73ed246 JA	385	LPOVERLAPPED lpOvl = &o->o;
67897036	386	DWORD iobytes;
93bcfd20	387	BOOL success = FALSE;
67897036 BC	388	int rc = FIO_Q_COMPLETED;
	389
	390	fio_ro_check(td, io_u);
	391
4e79098f BC	392	lpOvl->Internal = STATUS_PENDING;
	393	lpOvl->InternalHigh = 0;
	394	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	395	lpOvl->OffsetHigh = io_u->offset >> 32;
67897036 BC	396
67897036 BC	397	switch (io_u->ddir) {
40c5db35	398	case DDIR_WRITE:
67897036 BC	399	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	400	break;
	401	case DDIR_READ:
	402	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	403	break;
	404	case DDIR_SYNC:
	405	case DDIR_DATASYNC:
	406	case DDIR_SYNC_FILE_RANGE:
	407	success = FlushFileBuffers(io_u->file->hFile);
	408	if (!success)
2277d5d5	409	io_u->error = win_to_posix_error(GetLastError());
67897036 BC	410
	411	return FIO_Q_COMPLETED;
	412	break;
	413	case DDIR_TRIM:
	414	log_err("manual TRIM isn't supported on Windows");
	415	io_u->error = 1;
	416	io_u->resid = io_u->xfer_buflen;
	417	return FIO_Q_COMPLETED;
	418	break;
	419	default:
	420	assert(0);
93bcfd20	421	break;
67897036 BC	422	}
67897036 BC	423
40c5db35	424	if (success \|\| GetLastError() == ERROR_IO_PENDING)
67897036	425	rc = FIO_Q_QUEUED;
40c5db35	426	else {
2277d5d5	427	io_u->error = win_to_posix_error(GetLastError());
67897036 BC	428	io_u->resid = io_u->xfer_buflen;
	429	}
	430
	431	return rc;
	432	}
	433
	434	/* Runs as a thread and waits for queued IO to complete */
	435	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	436	{
	437	OVERLAPPED *ovl;
	438	struct fio_overlapped *fov;
	439	struct io_u *io_u;
	440	struct windowsaio_data *wd;
	441	struct thread_ctx *ctx;
	442	ULONG_PTR ulKey = 0;
	443	DWORD bytes;
	444
	445	ctx = (struct thread_ctx*)lpParameter;
	446	wd = ctx->wd;
	447
	448	do {
66c098b8	449	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
67897036 BC	450	continue;
	451
	452	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	453	io_u = fov->io_u;
	454
	455	if (ovl->Internal == ERROR_SUCCESS) {
	456	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	457	io_u->error = 0;
	458	} else {
	459	io_u->resid = io_u->xfer_buflen;
2277d5d5	460	io_u->error = win_to_posix_error(GetLastError());
67897036 BC	461	}
67897036 BC	462
40c5db35	463	fov->io_complete = TRUE;
67897036 BC	464	SetEvent(wd->iocomplete_event);
	465	} while (ctx->wd->iothread_running);
	466
	467	CloseHandle(ctx->iocp);
	468	free(ctx);
	469	return 0;
	470	}
	471
	472	static int fio_windowsaio_cancel(struct thread_data *td,
	473	struct io_u *io_u)
	474	{
	475	int rc = 0;
	476
	477	struct windowsaio_data *wd = td->io_ops->data;
	478
	479	/* If we're running on Vista or newer, we can cancel individual IO requests */
	480	if (wd->pCancelIoEx != NULL) {
	481	struct fio_overlapped *ovl = io_u->engine_data;
40c5db35	482
67897036 BC	483	if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
	484	rc = 1;
	485	} else
	486	rc = 1;
	487
	488	return rc;
	489	}
	490
c73ed246 JA	491	static void fio_windowsaio_io_u_free(struct thread_data td, struct io_u io_u)
	492	{
	493	struct fio_overlapped *o = io_u->engine_data;
	494
	495	if (o) {
	496	CloseHandle(o->o.hEvent);
	497	io_u->engine_data = NULL;
	498	free(o);
	499	}
	500	}
	501
	502	static int fio_windowsaio_io_u_init(struct thread_data td, struct io_u io_u)
	503	{
	504	struct fio_overlapped *o;
	505
	506	o = malloc(sizeof(*o));
b0106419	507	o->io_complete = FALSE;
c73ed246 JA	508	o->io_u = io_u;
	509	o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
	510	if (!o->o.hEvent) {
	511	free(o);
	512	return 1;
	513	}
	514
	515	io_u->engine_data = o;
	516	return 0;
	517	}
	518
ecc314ba BC	519	static struct ioengine_ops ioengine = {
	520	.name = "windowsaio",
	521	.version = FIO_IOOPS_VERSION,
	522	.init = fio_windowsaio_init,
	523	.queue = fio_windowsaio_queue,
	524	.cancel = fio_windowsaio_cancel,
	525	.getevents = fio_windowsaio_getevents,
	526	.event = fio_windowsaio_event,
	527	.cleanup = fio_windowsaio_cleanup,
	528	.open_file = fio_windowsaio_open_file,
	529	.close_file = fio_windowsaio_close_file,
c73ed246 JA	530	.get_file_size = generic_get_file_size,
	531	.io_u_init = fio_windowsaio_io_u_init,
	532	.io_u_free = fio_windowsaio_io_u_free,
ecc314ba BC	533	};
	534
	535	static void fio_init fio_posixaio_register(void)
	536	{
	537	register_ioengine(&ioengine);
	538	}
	539
	540	static void fio_exit fio_posixaio_unregister(void)
	541	{
	542	unregister_ioengine(&ioengine);
	543	}