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;
28 HANDLE iocomplete_event;
29 CANCELIOEX pCancelIoEx;
30 BOOL iothread_running;
35 struct windowsaio_data *wd;
38 static int fio_windowsaio_cancel(struct thread_data *td,
40 static BOOL timeout_expired(DWORD start_count, DWORD end_count);
41 static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
42 unsigned int max, struct timespec *t);
43 static struct io_u *fio_windowsaio_event(struct thread_data *td, int event);
44 static int fio_windowsaio_queue(struct thread_data *td,
46 static void fio_windowsaio_cleanup(struct thread_data *td);
47 static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
48 static int fio_windowsaio_init(struct thread_data *td);
49 static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f);
50 static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f);
51 static int win_to_poxix_error(DWORD winerr);
53 static int win_to_poxix_error(DWORD winerr)
57 case ERROR_FILE_NOT_FOUND: return ENOENT;
58 case ERROR_PATH_NOT_FOUND: return ENOENT;
59 case ERROR_ACCESS_DENIED: return EACCES;
60 case ERROR_INVALID_HANDLE: return EBADF;
61 case ERROR_NOT_ENOUGH_MEMORY: return ENOMEM;
62 case ERROR_INVALID_DATA: return EINVAL;
63 case ERROR_OUTOFMEMORY: return ENOMEM;
64 case ERROR_INVALID_DRIVE: return ENODEV;
65 case ERROR_NOT_SAME_DEVICE: return EXDEV;
66 case ERROR_WRITE_PROTECT: return EROFS;
67 case ERROR_BAD_UNIT: return ENODEV;
68 case ERROR_SHARING_VIOLATION: return EACCES;
69 case ERROR_LOCK_VIOLATION: return EACCES;
70 case ERROR_SHARING_BUFFER_EXCEEDED: return ENOLCK;
71 case ERROR_HANDLE_DISK_FULL: return ENOSPC;
72 case ERROR_NOT_SUPPORTED: return ENOSYS;
73 case ERROR_FILE_EXISTS: return EEXIST;
74 case ERROR_CANNOT_MAKE: return EPERM;
75 case ERROR_INVALID_PARAMETER: return EINVAL;
76 case ERROR_NO_PROC_SLOTS: return EAGAIN;
77 case ERROR_BROKEN_PIPE: return EPIPE;
78 case ERROR_OPEN_FAILED: return EIO;
79 case ERROR_NO_MORE_SEARCH_HANDLES: return ENFILE;
80 case ERROR_CALL_NOT_IMPLEMENTED: return ENOSYS;
81 case ERROR_INVALID_NAME: return ENOENT;
82 case ERROR_WAIT_NO_CHILDREN: return ECHILD;
83 case ERROR_CHILD_NOT_COMPLETE: return EBUSY;
84 case ERROR_DIR_NOT_EMPTY: return ENOTEMPTY;
85 case ERROR_SIGNAL_REFUSED: return EIO;
86 case ERROR_BAD_PATHNAME: return ENOENT;
87 case ERROR_SIGNAL_PENDING: return EBUSY;
88 case ERROR_MAX_THRDS_REACHED: return EAGAIN;
89 case ERROR_BUSY: return EBUSY;
90 case ERROR_ALREADY_EXISTS: return EEXIST;
91 case ERROR_NO_SIGNAL_SENT: return EIO;
92 case ERROR_FILENAME_EXCED_RANGE: return EINVAL;
93 case ERROR_META_EXPANSION_TOO_LONG: return EINVAL;
94 case ERROR_INVALID_SIGNAL_NUMBER: return EINVAL;
95 case ERROR_THREAD_1_INACTIVE: return EINVAL;
96 case ERROR_BAD_PIPE: return EINVAL;
97 case ERROR_PIPE_BUSY: return EBUSY;
98 case ERROR_NO_DATA: return EPIPE;
99 case ERROR_MORE_DATA: return EAGAIN;
100 case ERROR_DIRECTORY: return ENOTDIR;
101 case ERROR_PIPE_CONNECTED: return EBUSY;
102 case ERROR_NO_TOKEN: return EINVAL;
103 case ERROR_PROCESS_ABORTED: return EFAULT;
104 case ERROR_BAD_DEVICE: return ENODEV;
105 case ERROR_BAD_USERNAME: return EINVAL;
106 case ERROR_OPEN_FILES: return EAGAIN;
107 case ERROR_ACTIVE_CONNECTIONS: return EAGAIN;
108 case ERROR_DEVICE_IN_USE: return EAGAIN;
109 case ERROR_INVALID_AT_INTERRUPT_TIME: return EINTR;
110 case ERROR_IO_DEVICE: return EIO;
111 case ERROR_NOT_OWNER: return EPERM;
112 case ERROR_END_OF_MEDIA: return ENOSPC;
113 case ERROR_EOM_OVERFLOW: return ENOSPC;
114 case ERROR_BEGINNING_OF_MEDIA: return ESPIPE;
115 case ERROR_SETMARK_DETECTED: return ESPIPE;
116 case ERROR_NO_DATA_DETECTED: return ENOSPC;
117 case ERROR_POSSIBLE_DEADLOCK: return EDEADLOCK;
118 case ERROR_CRC: return EIO;
119 case ERROR_NEGATIVE_SEEK: return EINVAL;
120 case ERROR_DISK_FULL: return ENOSPC;
121 case ERROR_NOACCESS: return EFAULT;
122 case ERROR_FILE_INVALID: return ENXIO;
128 int sync_file_range(int fd, off64_t offset, off64_t nbytes,
135 static int fio_windowsaio_init(struct thread_data *td)
137 struct windowsaio_data *wd;
141 wd = malloc(sizeof(struct windowsaio_data));
143 ZeroMemory(wd, sizeof(struct windowsaio_data));
148 wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
149 if (wd->aio_events == NULL)
154 /* Create an auto-reset event */
155 wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
156 if (wd->iocomplete_event == NULL)
162 if (wd->aio_events != NULL)
163 free(wd->aio_events);
169 hKernel32Dll = GetModuleHandle("kernel32.dll");
170 wd->pCancelIoEx = (CANCELIOEX)GetProcAddress(hKernel32Dll, "CancelIoEx");
171 td->io_ops->data = wd;
176 static void fio_windowsaio_cleanup(struct thread_data *td)
178 struct windowsaio_data *wd;
180 wd = td->io_ops->data;
183 wd->iothread_running = FALSE;
184 WaitForSingleObject(wd->iothread, INFINITE);
186 CloseHandle(wd->iothread);
187 CloseHandle(wd->iocomplete_event);
189 free(wd->aio_events);
192 td->io_ops->data = NULL;
197 static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
201 DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED;
202 DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE;
203 DWORD openmode = OPEN_ALWAYS;
206 dprint(FD_FILE, "fd open %s\n", f->file_name);
208 if (f->filetype == FIO_TYPE_PIPE) {
209 log_err("fio: windowsaio doesn't support pipes\n");
213 if (!strcmp(f->file_name, "-")) {
214 log_err("fio: can't read/write to stdin/out\n");
219 flags |= FILE_FLAG_NO_BUFFERING;
221 flags |= FILE_FLAG_WRITE_THROUGH;
224 * Inform Windows whether we're going to be doing sequential or
225 * random io so it can tune the Cache Manager
227 if (td->o.td_ddir == TD_DDIR_READ ||
228 td->o.td_ddir == TD_DDIR_WRITE)
229 flags |= FILE_FLAG_SEQUENTIAL_SCAN;
231 flags |= FILE_FLAG_RANDOM_ACCESS;
233 if (!td_write(td) || read_only)
234 access = GENERIC_READ;
236 access = (GENERIC_READ | GENERIC_WRITE);
238 if (td->o.create_on_open)
239 openmode = OPEN_ALWAYS;
241 openmode = OPEN_EXISTING;
243 f->hFile = CreateFile(f->file_name, access, sharemode,
244 NULL, openmode, flags, NULL);
246 if (f->hFile == INVALID_HANDLE_VALUE)
249 /* Only set up the completion port and thread if we're not just
250 * querying the device size */
251 if (!rc && td->io_ops->data != NULL) {
252 struct thread_ctx *ctx;
253 struct windowsaio_data *wd;
255 hFile = CreateIoCompletionPort(f->hFile, NULL, 0, 0);
257 wd = td->io_ops->data;
258 wd->iothread_running = TRUE;
261 ctx = malloc(sizeof(struct thread_ctx));
265 wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
268 if (rc || wd->iothread == NULL)
275 static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f)
279 dprint(FD_FILE, "fd close %s\n", f->file_name);
281 if (f->hFile != INVALID_HANDLE_VALUE) {
282 if (!CloseHandle(f->hFile))
286 f->hFile = INVALID_HANDLE_VALUE;
290 static BOOL timeout_expired(DWORD start_count, DWORD end_count)
292 BOOL expired = FALSE;
295 current_time = GetTickCount();
297 if ((end_count > start_count) && current_time >= end_count)
299 else if (current_time < start_count && current_time > end_count)
305 static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
307 struct windowsaio_data *wd = td->io_ops->data;
308 return wd->aio_events[event];
311 static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
312 unsigned int max, struct timespec *t)
314 struct windowsaio_data *wd = td->io_ops->data;
315 struct flist_head *entry;
316 unsigned int dequeued = 0;
318 struct fio_overlapped *fov;
319 DWORD start_count = 0;
325 mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
326 start_count = GetTickCount();
327 end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
331 flist_for_each(entry, &td->io_u_busylist) {
332 io_u = flist_entry(entry, struct io_u, list);
333 fov = (struct fio_overlapped*)io_u->engine_data;
335 if (fov->io_complete) {
336 fov->io_complete = FALSE;
337 ResetEvent(fov->o.hEvent);
338 wd->aio_events[dequeued] = io_u;
346 if (dequeued < min) {
347 status = WaitForSingleObject(wd->iocomplete_event, mswait);
348 if (status != WAIT_OBJECT_0 && dequeued >= min)
352 if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count)))
359 static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u)
361 struct fio_overlapped *o = io_u->engine_data;
362 LPOVERLAPPED lpOvl = &o->o;
364 BOOL success = FALSE;
365 int rc = FIO_Q_COMPLETED;
367 fio_ro_check(td, io_u);
369 lpOvl->Internal = STATUS_PENDING;
370 lpOvl->InternalHigh = 0;
371 lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
372 lpOvl->OffsetHigh = io_u->offset >> 32;
374 switch (io_u->ddir) {
376 success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
379 success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
383 case DDIR_SYNC_FILE_RANGE:
384 success = FlushFileBuffers(io_u->file->hFile);
386 io_u->error = win_to_poxix_error(GetLastError());
388 return FIO_Q_COMPLETED;
391 log_err("manual TRIM isn't supported on Windows");
393 io_u->resid = io_u->xfer_buflen;
394 return FIO_Q_COMPLETED;
401 if (success || GetLastError() == ERROR_IO_PENDING)
404 io_u->error = win_to_poxix_error(GetLastError());
405 io_u->resid = io_u->xfer_buflen;
411 /* Runs as a thread and waits for queued IO to complete */
412 static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
415 struct fio_overlapped *fov;
417 struct windowsaio_data *wd;
418 struct thread_ctx *ctx;
422 ctx = (struct thread_ctx*)lpParameter;
426 if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
429 fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
432 if (ovl->Internal == ERROR_SUCCESS) {
433 io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
436 io_u->resid = io_u->xfer_buflen;
437 io_u->error = win_to_poxix_error(GetLastError());
440 fov->io_complete = TRUE;
441 SetEvent(wd->iocomplete_event);
442 } while (ctx->wd->iothread_running);
444 CloseHandle(ctx->iocp);
449 static int fio_windowsaio_cancel(struct thread_data *td,
454 struct windowsaio_data *wd = td->io_ops->data;
456 /* If we're running on Vista or newer, we can cancel individual IO requests */
457 if (wd->pCancelIoEx != NULL) {
458 struct fio_overlapped *ovl = io_u->engine_data;
460 if (!wd->pCancelIoEx(io_u->file->hFile, &ovl->o))
468 static void fio_windowsaio_io_u_free(struct thread_data *td, struct io_u *io_u)
470 struct fio_overlapped *o = io_u->engine_data;
473 CloseHandle(o->o.hEvent);
474 io_u->engine_data = NULL;
479 static int fio_windowsaio_io_u_init(struct thread_data *td, struct io_u *io_u)
481 struct fio_overlapped *o;
483 o = malloc(sizeof(*o));
484 o->io_complete = FALSE:
486 o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
492 io_u->engine_data = o;
496 static struct ioengine_ops ioengine = {
497 .name = "windowsaio",
498 .version = FIO_IOOPS_VERSION,
499 .init = fio_windowsaio_init,
500 .queue = fio_windowsaio_queue,
501 .cancel = fio_windowsaio_cancel,
502 .getevents = fio_windowsaio_getevents,
503 .event = fio_windowsaio_event,
504 .cleanup = fio_windowsaio_cleanup,
505 .open_file = fio_windowsaio_open_file,
506 .close_file = fio_windowsaio_close_file,
507 .get_file_size = generic_get_file_size,
508 .io_u_init = fio_windowsaio_io_u_init,
509 .io_u_free = fio_windowsaio_io_u_free,
512 static void fio_init fio_posixaio_register(void)
514 register_ioengine(&ioengine);
517 static void fio_exit fio_posixaio_unregister(void)
519 unregister_ioengine(&ioengine);