[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;
	BOOL io_free;
};

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

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

static int fio_windowsaio_cancel(struct thread_data *td,
			       struct io_u *io_u);
static BOOL timeout_expired(DWORD start_count, DWORD end_count);
static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
				    unsigned int max, struct timespec *t);
static struct io_u *fio_windowsaio_event(struct thread_data *td, int event);
static int fio_windowsaio_queue(struct thread_data *td,
			      struct io_u *io_u);
static void fio_windowsaio_cleanup(struct thread_data *td);
static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
static int fio_windowsaio_init(struct thread_data *td);
static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f);
static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f);
static int win_to_poxix_error(DWORD winerr);

static int win_to_poxix_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;
	int i;

	wd = malloc(sizeof(struct windowsaio_data));
	if (wd != NULL)
		ZeroMemory(wd, sizeof(struct windowsaio_data));
	else
		rc = 1;

	if (!rc) {
		wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
		if (wd->aio_events == NULL)
			rc = 1;
	}

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

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

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

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

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

			free(wd);
		}
	}

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

	return rc;
}

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

	wd = td->io_ops->data;

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

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

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

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

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


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

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

	if (f->filetype == FIO_TYPE_PIPE) {
		log_err("fio: windowsaio doesn't support pipes\n");
		return 1;
	}

	if (!strcmp(f->file_name, "-")) {
		log_err("fio: can't read/write to stdin/out\n");
		return 1;
	}

	if (td->o.odirect)
		flags |= FILE_FLAG_NO_BUFFERING;
	if (td->o.sync_io)
		flags |= FILE_FLAG_WRITE_THROUGH;

	/*
	 * Inform Windows whether we're going to be doing sequential or
	 * random io so it can tune the Cache Manager
	 */
	if (td->o.td_ddir == TD_DDIR_READ  ||
		td->o.td_ddir == TD_DDIR_WRITE)
		flags |= FILE_FLAG_SEQUENTIAL_SCAN;
	else
		flags |= FILE_FLAG_RANDOM_ACCESS;

	if (!td_write(td) || read_only)
		access = GENERIC_READ;
	else
		access = (GENERIC_READ | GENERIC_WRITE);

	if (td->o.create_on_open)
		openmode = OPEN_ALWAYS;
	else
		openmode = OPEN_EXISTING;

	f->hFile = CreateFile(f->file_name, access, sharemode,
		NULL, openmode, flags, NULL);

	if (f->hFile == INVALID_HANDLE_VALUE)
		rc = 1;

	/* Only set up the completion port and thread if we're not just
	 * querying the device size */
	if (!rc && td->io_ops->data != NULL) {
		struct thread_ctx *ctx;
		struct windowsaio_data *wd;

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

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

		if (!rc) {
			ctx = malloc(sizeof(struct thread_ctx));
			ctx->iocp = hFile;
			ctx->wd = wd;

			wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
		}

		if (rc || wd->iothread == NULL)
			rc = 1;
	}

	return rc;
}

static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f)
{
	int rc = 0;

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

	if (f->hFile != INVALID_HANDLE_VALUE) {
		if (!CloseHandle(f->hFile))
			rc = 1;
	}

	f->hFile = INVALID_HANDLE_VALUE;
	return rc;
}

static BOOL timeout_expired(DWORD start_count, DWORD end_count)
{
	BOOL expired = FALSE;
	DWORD current_time;

	current_time = GetTickCount();

	if ((end_count > start_count) && current_time >= end_count)
		expired = TRUE;
	else if (current_time < start_count && current_time > end_count)
		expired = TRUE;

	return expired;
}

static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
{
	struct windowsaio_data *wd = td->io_ops->data;
	return wd->aio_events[event];
}

static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
				    unsigned int max, struct timespec *t)
{
	struct windowsaio_data *wd = td->io_ops->data;
	struct flist_head *entry;
	unsigned int dequeued = 0;
	struct io_u *io_u;
	struct fio_overlapped *fov;
	DWORD start_count = 0;
	DWORD end_count = 0;
	DWORD status;
	DWORD mswait = 250;

	if (t != NULL) {
		mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
		start_count = GetTickCount();
		end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
	}

	do {
		flist_for_each(entry, &td->io_u_busylist) {
			io_u = flist_entry(entry, struct io_u, list);
			fov = (struct fio_overlapped*)io_u->engine_data;

			if (fov->io_complete) {
				fov->io_complete = FALSE;
				fov->io_free  = TRUE;
				ResetEvent(fov->o.hEvent);
				wd->aio_events[dequeued] = io_u;
				dequeued++;
			}

			if (dequeued >= min)
				break;
		}

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

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

	return dequeued;
}

static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u)
{
	LPOVERLAPPED lpOvl = NULL;
	struct windowsaio_data *wd;
	DWORD iobytes;
	BOOL success = FALSE;
	int index;
	int rc = FIO_Q_COMPLETED;

	fio_ro_check(td, io_u);

	wd = td->io_ops->data;

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

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

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

	switch (io_u->ddir) {
	case DDIR_WRITE:
		success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
		break;
	case DDIR_READ:
		success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
		break;
	case DDIR_SYNC:
	case DDIR_DATASYNC:
	case DDIR_SYNC_FILE_RANGE:
		success = FlushFileBuffers(io_u->file->hFile);
		if (!success)
		    io_u->error = win_to_poxix_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_poxix_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_poxix_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 struct ioengine_ops ioengine = {
	.name		= "windowsaio",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_windowsaio_init,
	.queue		= fio_windowsaio_queue,
	.cancel		= fio_windowsaio_cancel,
	.getevents	= fio_windowsaio_getevents,
	.event		= fio_windowsaio_event,
	.cleanup	= fio_windowsaio_cleanup,
	.open_file	= fio_windowsaio_open_file,
	.close_file	= fio_windowsaio_close_file,
	.get_file_size	= generic_get_file_size
};

static void fio_init fio_posixaio_register(void)
{
	register_ioengine(&ioengine);
}

static void fio_exit fio_posixaio_unregister(void)
{
	unregister_ioengine(&ioengine);
}
Commit	Line	Data
ecc314ba BC	1	/*
ecc314ba BC	2	* Native Windows async IO engine
93bcfd20	3	* Copyright (C) 2012 Bruce Cran <bruce@cran.org.uk>
ecc314ba BC	4	*/
ecc314ba BC	5
ecc314ba BC	6	#include <stdio.h>
	7	#include <stdlib.h>
	8	#include <unistd.h>
	9	#include <signal.h>
	10	#include <errno.h>
	11	#include <windows.h>
	12
	13	#include "../fio.h"
	14
ea4500d8 BC	15	typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);
ea4500d8 BC	16
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;
40c5db35	23	BOOL io_free;
67897036	24	};
ecc314ba	25
9b836561	26	struct windowsaio_data {
67897036	27	struct fio_overlapped *ovls;
9b836561	28	struct io_u **aio_events;
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 BC	40	static int fio_windowsaio_cancel(struct thread_data *td,
ecc314ba BC	41	struct io_u *io_u);
67897036	42	static BOOL timeout_expired(DWORD start_count, DWORD end_count);
ecc314ba BC	43	static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
	44	unsigned int max, struct timespec *t);
	45	static struct io_u fio_windowsaio_event(struct thread_data td, int event);
	46	static int fio_windowsaio_queue(struct thread_data *td,
	47	struct io_u *io_u);
	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);
66c098b8 BC	53	static int win_to_poxix_error(DWORD winerr);
	54
	55	static int win_to_poxix_error(DWORD winerr)
	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
9b836561 BC	130	int sync_file_range(int fd, off64_t offset, off64_t nbytes,
	131	unsigned int flags)
	132	{
	133	errno = ENOSYS;
	134	return -1;
	135	}
	136
ecc314ba BC	137	static int fio_windowsaio_init(struct thread_data *td)
ecc314ba BC	138	{
ecc314ba	139	struct windowsaio_data *wd;
ea4500d8	140	HANDLE hKernel32Dll;
9b836561	141	int rc = 0;
67897036	142	int i;
ecc314ba	143
ecc314ba	144	wd = malloc(sizeof(struct windowsaio_data));
9b836561 BC	145	if (wd != NULL)
	146	ZeroMemory(wd, sizeof(struct windowsaio_data));
	147	else
	148	rc = 1;
ecc314ba	149
9b836561 BC	150	if (!rc) {
	151	wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
	152	if (wd->aio_events == NULL)
	153	rc = 1;
e4db9fec BC	154	}
e4db9fec BC	155
9b836561	156	if (!rc) {
67897036 BC	157	wd->ovls = malloc(td->o.iodepth * sizeof(struct fio_overlapped));
67897036 BC	158	if (wd->ovls == NULL)
9b836561	159	rc = 1;
e4db9fec BC	160	}
e4db9fec BC	161
9b836561	162	if (!rc) {
f9a58c2a BC	163	for (i = 0; i < td->o.iodepth; i++) {
	164	wd->ovls[i].io_free = TRUE;
	165	wd->ovls[i].io_complete = FALSE;
67897036 BC	166
	167	wd->ovls[i].o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
	168	if (wd->ovls[i].o.hEvent == NULL) {
	169	rc = 1;
	170	break;
	171	}
f9a58c2a	172	}
9b836561	173	}
e4db9fec	174
9b836561 BC	175	if (!rc) {
	176	/* Create an auto-reset event */
	177	wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
	178	if (wd->iocomplete_event == NULL)
	179	rc = 1;
	180	}
	181
9b836561	182	if (rc) {
9b836561 BC	183	if (wd != NULL) {
	184	if (wd->ovls != NULL)
	185	free(wd->ovls);
9b836561 BC	186	if (wd->aio_events != NULL)
	187	free(wd->aio_events);
	188
	189	free(wd);
	190	}
	191	}
ecc314ba	192
ea4500d8	193	hKernel32Dll = GetModuleHandle("kernel32.dll");
93bcfd20	194	wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
ecc314ba	195	td->io_ops->data = wd;
93bcfd20	196
4e79098f	197	return rc;
ecc314ba BC	198	}
ecc314ba BC	199
67897036 BC	200	static void fio_windowsaio_cleanup(struct thread_data *td)
	201	{
	202	int i;
	203	struct windowsaio_data *wd;
	204
	205	wd = td->io_ops->data;
	206
	207	if (wd != NULL) {
40c5db35 JA	208	wd->iothread_running = FALSE;
40c5db35 JA	209	WaitForSingleObject(wd->iothread, INFINITE);
67897036 BC	210
	211	CloseHandle(wd->iothread);
	212	CloseHandle(wd->iocomplete_event);
	213
93bcfd20	214	for (i = 0; i < td->o.iodepth; i++) {
67897036	215	CloseHandle(wd->ovls[i].o.hEvent);
93bcfd20	216	}
67897036 BC	217
	218	free(wd->aio_events);
	219	free(wd->ovls);
	220	free(wd);
	221
	222	td->io_ops->data = NULL;
	223	}
	224	}
	225
	226
ecc314ba BC	227	static int fio_windowsaio_open_file(struct thread_data td, struct fio_file f)
	228	{
	229	int rc = 0;
	230	HANDLE hFile;
4e79098f	231	DWORD flags = FILE_FLAG_POSIX_SEMANTICS \| FILE_FLAG_OVERLAPPED;
ecc314ba BC	232	DWORD sharemode = FILE_SHARE_READ \| FILE_SHARE_WRITE;
	233	DWORD openmode = OPEN_ALWAYS;
	234	DWORD access;
	235
	236	dprint(FD_FILE, "fd open %s\n", f->file_name);
	237
ecc314ba BC	238	if (f->filetype == FIO_TYPE_PIPE) {
	239	log_err("fio: windowsaio doesn't support pipes\n");
	240	return 1;
	241	}
	242
	243	if (!strcmp(f->file_name, "-")) {
	244	log_err("fio: can't read/write to stdin/out\n");
	245	return 1;
	246	}
	247
	248	if (td->o.odirect)
	249	flags \|= FILE_FLAG_NO_BUFFERING;
	250	if (td->o.sync_io)
	251	flags \|= FILE_FLAG_WRITE_THROUGH;
	252
93bcfd20 BC	253	/*
	254	* Inform Windows whether we're going to be doing sequential or
	255	* random io so it can tune the Cache Manager
	256	*/
ecc314ba	257	if (td->o.td_ddir == TD_DDIR_READ \|\|
ea4500d8	258	td->o.td_ddir == TD_DDIR_WRITE)
ecc314ba	259	flags \|= FILE_FLAG_SEQUENTIAL_SCAN;
ecc314ba	260	else
ecc314ba	261	flags \|= FILE_FLAG_RANDOM_ACCESS;
ecc314ba	262
ea4500d8	263	if (!td_write(td) \|\| read_only)
ecc314ba BC	264	access = GENERIC_READ;
	265	else
	266	access = (GENERIC_READ \| GENERIC_WRITE);
	267
93bcfd20	268	if (td->o.create_on_open)
ecc314ba BC	269	openmode = OPEN_ALWAYS;
	270	else
	271	openmode = OPEN_EXISTING;
	272
	273	f->hFile = CreateFile(f->file_name, access, sharemode,
	274	NULL, openmode, flags, NULL);
	275
4e79098f	276	if (f->hFile == INVALID_HANDLE_VALUE)
ecc314ba	277	rc = 1;
ecc314ba	278
93bcfd20	279	/* Only set up the completion port and thread if we're not just
ecc314ba	280	* querying the device size */
40c5db35	281	if (!rc && td->io_ops->data != NULL) {
ecc314ba	282	struct thread_ctx *ctx;
40c5db35	283	struct windowsaio_data *wd;
ecc314ba	284
40c5db35	285	hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
ea4500d8	286
40c5db35	287	wd = td->io_ops->data;
9b836561	288	wd->iothread_running = TRUE;
ecc314ba	289
9b836561 BC	290	if (!rc) {
	291	ctx = malloc(sizeof(struct thread_ctx));
	292	ctx->iocp = hFile;
	293	ctx->wd = wd;
	294
	295	wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
	296	}
ecc314ba	297
4e79098f	298	if (rc \|\| wd->iothread == NULL)
ecc314ba	299	rc = 1;
ecc314ba BC	300	}
ecc314ba BC	301
ecc314ba BC	302	return rc;
	303	}
	304
	305	static int fio_windowsaio_close_file(struct thread_data fio_unused td, struct fio_file f)
	306	{
4e79098f	307	int rc = 0;
1e42cc3d	308
9b836561 BC	309	dprint(FD_FILE, "fd close %s\n", f->file_name);
9b836561 BC	310
ecc314ba	311	if (f->hFile != INVALID_HANDLE_VALUE) {
9b836561	312	if (!CloseHandle(f->hFile))
4e79098f	313	rc = 1;
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,
	342	unsigned int max, struct timespec *t)
	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;
67897036	367	fov->io_free = TRUE;
f9a58c2a	368	ResetEvent(fov->o.hEvent);
67897036 BC	369	wd->aio_events[dequeued] = io_u;
	370	dequeued++;
	371	}
	372
	373	if (dequeued >= min)
	374	break;
	375	}
	376
40c5db35	377	if (dequeued < min) {
67897036	378	status = WaitForSingleObject(wd->iocomplete_event, mswait);
f9a58c2a	379	if (status != WAIT_OBJECT_0 && dequeued >= min)
67897036 BC	380	break;
	381	}
	382
	383	if (dequeued >= min \|\| (t != NULL && timeout_expired(start_count, end_count)))
	384	break;
	385	} while (1);
	386
	387	return dequeued;
	388	}
	389
40c5db35	390	static int fio_windowsaio_queue(struct thread_data td, struct io_u io_u)
67897036	391	{
1e42cc3d	392	LPOVERLAPPED lpOvl = NULL;
67897036 BC	393	struct windowsaio_data *wd;
67897036 BC	394	DWORD iobytes;
93bcfd20	395	BOOL success = FALSE;
67897036 BC	396	int index;
	397	int rc = FIO_Q_COMPLETED;
	398
	399	fio_ro_check(td, io_u);
	400
	401	wd = td->io_ops->data;
	402
1e42cc3d	403	for (index = 0; index < td->o.iodepth; index++) {
f9a58c2a	404	if (wd->ovls[index].io_free)
1e42cc3d	405	break;
1e42cc3d	406	}
67897036	407
1e42cc3d	408	assert(index < td->o.iodepth);
67897036	409
f9a58c2a	410	wd->ovls[index].io_free = FALSE;
1e42cc3d	411	wd->ovls[index].io_u = io_u;
f9a58c2a	412	lpOvl = &wd->ovls[index].o;
4e79098f BC	413	lpOvl->Internal = STATUS_PENDING;
	414	lpOvl->InternalHigh = 0;
	415	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
	416	lpOvl->OffsetHigh = io_u->offset >> 32;
1e42cc3d	417	io_u->engine_data = &wd->ovls[index];
67897036 BC	418
67897036 BC	419	switch (io_u->ddir) {
40c5db35	420	case DDIR_WRITE:
67897036 BC	421	success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	422	break;
	423	case DDIR_READ:
	424	success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
	425	break;
	426	case DDIR_SYNC:
	427	case DDIR_DATASYNC:
	428	case DDIR_SYNC_FILE_RANGE:
	429	success = FlushFileBuffers(io_u->file->hFile);
	430	if (!success)
66c098b8	431	io_u->error = win_to_poxix_error(GetLastError());
67897036 BC	432
	433	return FIO_Q_COMPLETED;
	434	break;
	435	case DDIR_TRIM:
	436	log_err("manual TRIM isn't supported on Windows");
	437	io_u->error = 1;
	438	io_u->resid = io_u->xfer_buflen;
	439	return FIO_Q_COMPLETED;
	440	break;
	441	default:
	442	assert(0);
93bcfd20	443	break;
67897036 BC	444	}
67897036 BC	445
40c5db35	446	if (success \|\| GetLastError() == ERROR_IO_PENDING)
67897036	447	rc = FIO_Q_QUEUED;
40c5db35	448	else {
66c098b8	449	io_u->error = win_to_poxix_error(GetLastError());
67897036 BC	450	io_u->resid = io_u->xfer_buflen;
	451	}
	452
	453	return rc;
	454	}
	455
	456	/* Runs as a thread and waits for queued IO to complete */
	457	static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
	458	{
	459	OVERLAPPED *ovl;
	460	struct fio_overlapped *fov;
	461	struct io_u *io_u;
	462	struct windowsaio_data *wd;
	463	struct thread_ctx *ctx;
	464	ULONG_PTR ulKey = 0;
	465	DWORD bytes;
	466
	467	ctx = (struct thread_ctx*)lpParameter;
	468	wd = ctx->wd;
	469
	470	do {
66c098b8	471	if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
67897036 BC	472	continue;
	473
	474	fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
	475	io_u = fov->io_u;
	476
	477	if (ovl->Internal == ERROR_SUCCESS) {
	478	io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
	479	io_u->error = 0;
	480	} else {
	481	io_u->resid = io_u->xfer_buflen;
66c098b8	482	io_u->error = win_to_poxix_error(GetLastError());
67897036 BC	483	}
67897036 BC	484
40c5db35	485	fov->io_complete = TRUE;
67897036 BC	486	SetEvent(wd->iocomplete_event);
	487	} while (ctx->wd->iothread_running);
	488
	489	CloseHandle(ctx->iocp);
	490	free(ctx);
	491	return 0;
	492	}
	493
	494	static int fio_windowsaio_cancel(struct thread_data *td,
	495	struct io_u *io_u)
	496	{
	497	int rc = 0;
	498
	499	struct windowsaio_data *wd = td->io_ops->data;
	500
	501	/* If we're running on Vista or newer, we can cancel individual IO requests */
	502	if (wd->pCancelIoEx != NULL) {
	503	struct fio_overlapped *ovl = io_u->engine_data;
40c5db35	504
67897036 BC	505	if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
	506	rc = 1;
	507	} else
	508	rc = 1;
	509
	510	return rc;
	511	}
	512
ecc314ba BC	513	static struct ioengine_ops ioengine = {
	514	.name = "windowsaio",
	515	.version = FIO_IOOPS_VERSION,
	516	.init = fio_windowsaio_init,
	517	.queue = fio_windowsaio_queue,
	518	.cancel = fio_windowsaio_cancel,
	519	.getevents = fio_windowsaio_getevents,
	520	.event = fio_windowsaio_event,
	521	.cleanup = fio_windowsaio_cleanup,
	522	.open_file = fio_windowsaio_open_file,
	523	.close_file = fio_windowsaio_close_file,
	524	.get_file_size = generic_get_file_size
	525	};
	526
	527	static void fio_init fio_posixaio_register(void)
	528	{
	529	register_ioengine(&ioengine);
	530	}
	531
	532	static void fio_exit fio_posixaio_unregister(void)
	533	{
	534	unregister_ioengine(&ioengine);
	535	}