2 * Native Windows async IO engine
3 * Copyright (C) 2012 Bruce Cran <bruce@cran.org.uk>
15 typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);
17 int geterrno_from_win_error (DWORD code, int deferrno);
19 struct fio_overlapped {
25 struct windowsaio_data {
26 struct io_u **aio_events;
29 HANDLE iocomplete_event;
30 CANCELIOEX pCancelIoEx;
31 BOOL iothread_running;
36 struct windowsaio_data *wd;
39 static int fio_windowsaio_cancel(struct thread_data *td,
41 static BOOL timeout_expired(DWORD start_count, DWORD end_count);
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,
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);
52 static int win_to_posix_error(DWORD winerr);
54 static int win_to_posix_error(DWORD winerr)
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;
129 int sync_file_range(int fd, off64_t offset, off64_t nbytes,
136 static int fio_windowsaio_init(struct thread_data *td)
138 struct windowsaio_data *wd;
142 wd = malloc(sizeof(struct windowsaio_data));
144 ZeroMemory(wd, sizeof(struct windowsaio_data));
149 wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
150 if (wd->aio_events == NULL)
155 /* Create an auto-reset event */
156 wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
157 if (wd->iocomplete_event == NULL)
163 if (wd->aio_events != NULL)
164 free(wd->aio_events);
170 hKernel32Dll = GetModuleHandle("kernel32.dll");
171 wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
172 td->io_ops->data = wd;
176 struct thread_ctx *ctx;
177 struct windowsaio_data *wd;
180 hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0);
181 if (hFile == INVALID_HANDLE_VALUE)
184 wd = td->io_ops->data;
185 wd->iothread_running = TRUE;
189 ctx = malloc(sizeof(struct thread_ctx));
191 if (!rc && ctx == NULL)
193 log_err("fio: out of memory in windowsaio\n");
202 wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
205 if (rc || wd->iothread == NULL)
212 static void fio_windowsaio_cleanup(struct thread_data *td)
214 struct windowsaio_data *wd;
216 wd = td->io_ops->data;
219 wd->iothread_running = FALSE;
220 WaitForSingleObject(wd->iothread, INFINITE);
222 CloseHandle(wd->iothread);
223 CloseHandle(wd->iocomplete_event);
225 free(wd->aio_events);
228 td->io_ops->data = NULL;
233 static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
236 DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED;
237 DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE;
238 DWORD openmode = OPEN_ALWAYS;
241 dprint(FD_FILE, "fd open %s\n", f->file_name);
243 if (f->filetype == FIO_TYPE_PIPE) {
244 log_err("fio: windowsaio doesn't support pipes\n");
248 if (!strcmp(f->file_name, "-")) {
249 log_err("fio: can't read/write to stdin/out\n");
254 flags |= FILE_FLAG_NO_BUFFERING;
256 flags |= FILE_FLAG_WRITE_THROUGH;
259 * Inform Windows whether we're going to be doing sequential or
260 * random io so it can tune the Cache Manager
262 if (td->o.td_ddir == TD_DDIR_READ ||
263 td->o.td_ddir == TD_DDIR_WRITE)
264 flags |= FILE_FLAG_SEQUENTIAL_SCAN;
266 flags |= FILE_FLAG_RANDOM_ACCESS;
268 if (!td_write(td) || read_only)
269 access = GENERIC_READ;
271 access = (GENERIC_READ | GENERIC_WRITE);
273 if (td->o.create_on_open)
274 openmode = OPEN_ALWAYS;
276 openmode = OPEN_EXISTING;
278 f->hFile = CreateFile(f->file_name, access, sharemode,
279 NULL, openmode, flags, NULL);
281 if (f->hFile == INVALID_HANDLE_VALUE)
284 /* Only set up the completion port and thread if we're not just
285 * querying the device size */
286 if (!rc && td->io_ops->data != NULL) {
287 struct windowsaio_data *wd;
289 wd = td->io_ops->data;
291 if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL)
298 static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f)
302 dprint(FD_FILE, "fd close %s\n", f->file_name);
304 if (f->hFile != INVALID_HANDLE_VALUE) {
305 if (!CloseHandle(f->hFile))
309 f->hFile = INVALID_HANDLE_VALUE;
313 static BOOL timeout_expired(DWORD start_count, DWORD end_count)
315 BOOL expired = FALSE;
318 current_time = GetTickCount();
320 if ((end_count > start_count) && current_time >= end_count)
322 else if (current_time < start_count && current_time > end_count)
328 static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
330 struct windowsaio_data *wd = td->io_ops->data;
331 return wd->aio_events[event];
334 static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
335 unsigned int max, struct timespec *t)
337 struct windowsaio_data *wd = td->io_ops->data;
338 struct flist_head *entry;
339 unsigned int dequeued = 0;
341 struct fio_overlapped *fov;
342 DWORD start_count = 0;
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);
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;
358 if (fov->io_complete) {
359 fov->io_complete = FALSE;
360 ResetEvent(fov->o.hEvent);
361 wd->aio_events[dequeued] = io_u;
369 if (dequeued < min) {
370 status = WaitForSingleObject(wd->iocomplete_event, mswait);
371 if (status != WAIT_OBJECT_0 && dequeued >= min)
375 if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count)))
382 static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u)
384 struct fio_overlapped *o = io_u->engine_data;
385 LPOVERLAPPED lpOvl = &o->o;
387 BOOL success = FALSE;
388 int rc = FIO_Q_COMPLETED;
390 fio_ro_check(td, io_u);
392 lpOvl->Internal = STATUS_PENDING;
393 lpOvl->InternalHigh = 0;
394 lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
395 lpOvl->OffsetHigh = io_u->offset >> 32;
397 switch (io_u->ddir) {
399 success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
402 success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
406 case DDIR_SYNC_FILE_RANGE:
407 success = FlushFileBuffers(io_u->file->hFile);
409 io_u->error = win_to_posix_error(GetLastError());
411 return FIO_Q_COMPLETED;
414 log_err("manual TRIM isn't supported on Windows");
416 io_u->resid = io_u->xfer_buflen;
417 return FIO_Q_COMPLETED;
424 if (success || GetLastError() == ERROR_IO_PENDING)
427 io_u->error = win_to_posix_error(GetLastError());
428 io_u->resid = io_u->xfer_buflen;
434 /* Runs as a thread and waits for queued IO to complete */
435 static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
438 struct fio_overlapped *fov;
440 struct windowsaio_data *wd;
441 struct thread_ctx *ctx;
445 ctx = (struct thread_ctx*)lpParameter;
449 if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
452 fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
455 if (ovl->Internal == ERROR_SUCCESS) {
456 io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
459 io_u->resid = io_u->xfer_buflen;
460 io_u->error = win_to_posix_error(GetLastError());
463 fov->io_complete = TRUE;
464 SetEvent(wd->iocomplete_event);
465 } while (ctx->wd->iothread_running);
467 CloseHandle(ctx->iocp);
472 static int fio_windowsaio_cancel(struct thread_data *td,
477 struct windowsaio_data *wd = td->io_ops->data;
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;
483 if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
491 static void fio_windowsaio_io_u_free(struct thread_data *td, struct io_u *io_u)
493 struct fio_overlapped *o = io_u->engine_data;
496 CloseHandle(o->o.hEvent);
497 io_u->engine_data = NULL;
502 static int fio_windowsaio_io_u_init(struct thread_data *td, struct io_u *io_u)
504 struct fio_overlapped *o;
506 o = malloc(sizeof(*o));
507 o->io_complete = FALSE:
509 o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
515 io_u->engine_data = o;
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,
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,
535 static void fio_init fio_posixaio_register(void)
537 register_ioengine(&ioengine);
540 static void fio_exit fio_posixaio_unregister(void)
542 unregister_ioengine(&ioengine);