/*
* Native Windows async IO engine
- * Copyright (C) 2011 Bruce Cran <bruce@cran.org.uk>
+ * Copyright (C) 2012 Bruce Cran <bruce@cran.org.uk>
*/
#include <stdio.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 iocp;
HANDLE iothread;
HANDLE iocomplete_event;
CANCELIOEX pCancelIoEx;
BOOL iothread_running;
- BOOL use_iocp;
};
struct thread_ctx {
struct windowsaio_data *wd;
};
-static void PrintError(LPCSTR lpszFunction);
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_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);
-int sync_file_range(int fd, off64_t offset, off64_t nbytes,
- unsigned int flags)
+static int win_to_posix_error(DWORD winerr)
{
- errno = ENOSYS;
- return -1;
-}
+ 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;
+ }
-static void PrintError(LPCSTR lpszFunction)
-{
- // Retrieve the system error message for the last-error code
-
- LPSTR lpMsgBuf;
- DWORD dw = GetLastError();
-
- FormatMessage(
- FORMAT_MESSAGE_ALLOCATE_BUFFER |
- FORMAT_MESSAGE_FROM_SYSTEM |
- FORMAT_MESSAGE_IGNORE_INSERTS,
- NULL,
- dw,
- MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
- (LPTSTR)&lpMsgBuf,
- 0, NULL );
-
- log_err("%s - %s", lpszFunction, lpMsgBuf);
- LocalFree(lpMsgBuf);
+ return winerr;
}
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)
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 (rc) {
- PrintError(__func__);
if (wd != NULL) {
- if (wd->ovls != NULL)
- free(wd->ovls);
if (wd->aio_events != NULL)
free(wd->aio_events);
}
hKernel32Dll = GetModuleHandle("kernel32.dll");
- wd->pCancelIoEx = GetProcAddress(hKernel32Dll, "CancelIoEx");
-
+ wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
td->io_ops->data = wd;
- return 0;
+
+
+ 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)
{
- int i;
struct windowsaio_data *wd;
wd = td->io_ops->data;
if (wd != NULL) {
- wd->iothread_running = FALSE;
- WaitForSingleObject(wd->iothread, INFINITE);
+ 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;
+ DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED;
DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE;
DWORD openmode = OPEN_ALWAYS;
DWORD access;
return 1;
}
- if (!td->o.odirect && !td->o.sync_io && td->io_ops->data != NULL)
- flags |= FILE_FLAG_OVERLAPPED;
-
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
access = (GENERIC_READ | GENERIC_WRITE);
- if (td->o.create_on_open > 0)
+ 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) {
- PrintError(__func__);
+ if (f->hFile == INVALID_HANDLE_VALUE)
rc = 1;
- }
- /* Only set up the competion port and thread if we're not just
+ /* Only set up the completion port and thread if we're not just
* querying the device size */
- if (!rc && td->io_ops->data != NULL && !td->o.odirect && !td->o.sync_io) {
- struct thread_ctx *ctx;
- struct windowsaio_data *wd;
- hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
-
- wd = td->io_ops->data;
-
- if (!td->o.odirect && !td->o.sync_io)
- wd->use_iocp = 1;
- else
- wd->use_iocp = 0;
-
- wd->iothread_running = TRUE;
-
- if (!rc) {
- ctx = malloc(sizeof(struct thread_ctx));
- ctx->iocp = hFile;
- ctx->wd = wd;
+ if (!rc && td->io_ops->data != NULL) {
+ struct windowsaio_data *wd;
- wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
- }
+ wd = td->io_ops->data;
- if (rc || wd->iothread == NULL) {
- PrintError(__func__);
+ 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))
- PrintError(__func__);
+ rc = 1;
}
f->hFile = INVALID_HANDLE_VALUE;
- return 0;
+ return rc;
}
static BOOL timeout_expired(DWORD start_count, DWORD end_count)
fov = (struct fio_overlapped*)io_u->engine_data;
if (fov->io_complete) {
- fov->io_complete = FALSE;
- fov->io_free = TRUE;
+ fov->io_complete = FALSE;
+ ResetEvent(fov->o.hEvent);
wd->aio_events[dequeued] = io_u;
dequeued++;
}
break;
}
- if (dequeued < min) {
+ if (dequeued < min) {
status = WaitForSingleObject(wd->iocomplete_event, mswait);
- if (status != WAIT_OBJECT_0 && dequeued > 0)
+ if (status != WAIT_OBJECT_0 && dequeued >= min)
break;
}
return dequeued;
}
-static int fio_windowsaio_queue(struct thread_data *td,
- struct io_u *io_u)
+static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u)
{
- LPOVERLAPPED lpOvl = NULL;
- struct windowsaio_data *wd;
+ struct fio_overlapped *o = io_u->engine_data;
+ LPOVERLAPPED lpOvl = &o->o;
DWORD iobytes;
- BOOL success;
- int index;
+ BOOL success = FALSE;
int rc = FIO_Q_COMPLETED;
fio_ro_check(td, io_u);
- wd = td->io_ops->data;
-
- if (wd->use_iocp) {
- for (index = 0; index < td->o.iodepth; index++) {
- if (wd->ovls[index].io_free) {
- wd->ovls[index].io_free = FALSE;
- ResetEvent(wd->ovls[index].o.hEvent);
- break;
- }
- }
-
- assert(index < td->o.iodepth);
-
- lpOvl = &wd->ovls[index].o;
- wd->ovls[index].io_u = io_u;
- lpOvl->Internal = STATUS_PENDING;
- lpOvl->InternalHigh = 0;
- lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
- lpOvl->OffsetHigh = io_u->offset >> 32;
- lpOvl->Pointer = NULL;
- io_u->engine_data = &wd->ovls[index];
- }
+ 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:
+ case DDIR_WRITE:
success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
break;
case DDIR_READ:
case DDIR_SYNC_FILE_RANGE:
success = FlushFileBuffers(io_u->file->hFile);
if (!success)
- io_u->error = GetLastError();
+ io_u->error = win_to_posix_error(GetLastError());
return FIO_Q_COMPLETED;
break;
break;
default:
assert(0);
+ break;
}
- if (wd->use_iocp && (success || GetLastError() == ERROR_IO_PENDING)) {
+ if (success || GetLastError() == ERROR_IO_PENDING)
rc = FIO_Q_QUEUED;
- } else if (success && !wd->use_iocp) {
- io_u->resid = io_u->xfer_buflen - iobytes;
- io_u->error = 0;
- } else {
- PrintError(__func__);
- io_u->error = GetLastError();
+ else {
+ io_u->error = win_to_posix_error(GetLastError());
io_u->resid = io_u->xfer_buflen;
}
wd = ctx->wd;
do {
- if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250))
+ if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
continue;
fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
io_u->error = 0;
} else {
io_u->resid = io_u->xfer_buflen;
- io_u->error = ovl->Internal;
+ io_u->error = win_to_posix_error(GetLastError());
}
- fov->io_complete = TRUE;
+ fov->io_complete = TRUE;
SetEvent(wd->iocomplete_event);
} while (ctx->wd->iothread_running);
/* 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
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,
.cleanup = fio_windowsaio_cleanup,
.open_file = fio_windowsaio_open_file,
.close_file = fio_windowsaio_close_file,
- .get_file_size = generic_get_file_size
+ .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)