4 * IO engine using Windows IO Completion Ports.
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 BOOL iothread_running;
35 struct windowsaio_data *wd;
38 static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
40 static int fio_windowsaio_init(struct thread_data *td)
42 struct windowsaio_data *wd;
45 wd = calloc(1, sizeof(struct windowsaio_data));
47 log_err("windowsaio: failed to allocate memory for engine data\n");
52 wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
53 if (wd->aio_events == NULL) {
54 log_err("windowsaio: failed to allocate memory for aio events list\n");
60 /* Create an auto-reset event */
61 wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
62 if (wd->iocomplete_event == NULL) {
63 log_err("windowsaio: failed to create io complete event handle\n");
70 if (wd->aio_events != NULL)
80 struct thread_ctx *ctx;
81 struct windowsaio_data *wd;
84 hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0);
85 if (hFile == INVALID_HANDLE_VALUE) {
86 log_err("windowsaio: failed to create io completion port\n");
91 wd->iothread_running = TRUE;
95 ctx = malloc(sizeof(struct thread_ctx));
97 if (!rc && ctx == NULL)
99 log_err("windowsaio: failed to allocate memory for thread context structure\n");
110 wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, &threadid);
112 if (wd->iothread != NULL)
113 fio_setaffinity(threadid, td->o.cpumask);
115 log_err("windowsaio: failed to create io completion thread\n");
118 if (rc || wd->iothread == NULL)
125 static void fio_windowsaio_cleanup(struct thread_data *td)
127 struct windowsaio_data *wd;
129 wd = td->io_ops_data;
132 wd->iothread_running = FALSE;
133 WaitForSingleObject(wd->iothread, INFINITE);
135 CloseHandle(wd->iothread);
136 CloseHandle(wd->iocomplete_event);
138 free(wd->aio_events);
141 td->io_ops_data = NULL;
145 static int windowsaio_invalidate_cache(struct fio_file *f)
148 DWORD isharemode = (FILE_SHARE_DELETE | FILE_SHARE_READ |
154 * Encourage Windows to drop cached parts of a file by temporarily
155 * opening it for non-buffered access. Note: this will only work when
156 * the following is the only thing with the file open on the whole
159 dprint(FD_IO, "windowaio: attempt invalidate cache for %s\n",
161 ihFile = CreateFile(f->file_name, 0, isharemode, NULL, OPEN_EXISTING,
162 FILE_FLAG_NO_BUFFERING, NULL);
164 if (ihFile != INVALID_HANDLE_VALUE) {
165 if (!CloseHandle(ihFile)) {
166 error = GetLastError();
167 log_info("windowsaio: invalidation fd close %s "
168 "failed: error %d\n", f->file_name, error);
172 error = GetLastError();
173 if (error != ERROR_FILE_NOT_FOUND) {
174 log_info("windowsaio: cache invalidation of %s failed: "
175 "error %d\n", f->file_name, error);
183 static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
186 DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED;
187 DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE;
188 DWORD openmode = OPEN_ALWAYS;
191 dprint(FD_FILE, "fd open %s\n", f->file_name);
193 if (f->filetype == FIO_TYPE_PIPE) {
194 log_err("windowsaio: pipes are not supported\n");
198 if (!strcmp(f->file_name, "-")) {
199 log_err("windowsaio: can't read/write to stdin/out\n");
204 flags |= FILE_FLAG_NO_BUFFERING;
206 flags |= FILE_FLAG_WRITE_THROUGH;
209 * Inform Windows whether we're going to be doing sequential or
210 * random IO so it can tune the Cache Manager
212 switch (td->o.fadvise_hint) {
215 flags |= FILE_FLAG_RANDOM_ACCESS;
217 flags |= FILE_FLAG_SEQUENTIAL_SCAN;
220 flags |= FILE_FLAG_RANDOM_ACCESS;
222 case F_ADV_SEQUENTIAL:
223 flags |= FILE_FLAG_SEQUENTIAL_SCAN;
228 log_err("fio: unknown fadvise type %d\n", td->o.fadvise_hint);
231 if (!td_write(td) || read_only)
232 access = GENERIC_READ;
234 access = (GENERIC_READ | GENERIC_WRITE);
236 if (td->o.create_on_open)
237 openmode = OPEN_ALWAYS;
239 openmode = OPEN_EXISTING;
241 /* If we're going to use direct I/O, Windows will try and invalidate
242 * its cache at that point so there's no need to do it here */
243 if (td->o.invalidate_cache && !td->o.odirect)
244 windowsaio_invalidate_cache(f);
246 f->hFile = CreateFile(f->file_name, access, sharemode,
247 NULL, openmode, flags, NULL);
249 if (f->hFile == INVALID_HANDLE_VALUE) {
250 log_err("windowsaio: failed to open file \"%s\"\n", f->file_name);
254 /* Only set up the completion port and thread if we're not just
255 * querying the device size */
256 if (!rc && td->io_ops_data != NULL) {
257 struct windowsaio_data *wd;
259 wd = td->io_ops_data;
261 if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL) {
262 log_err("windowsaio: failed to create io completion port\n");
270 static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f)
274 dprint(FD_FILE, "fd close %s\n", f->file_name);
276 if (f->hFile != INVALID_HANDLE_VALUE) {
277 if (!CloseHandle(f->hFile)) {
278 log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name);
283 f->hFile = INVALID_HANDLE_VALUE;
287 static BOOL timeout_expired(DWORD start_count, DWORD end_count)
289 BOOL expired = FALSE;
292 current_time = GetTickCount();
294 if ((end_count > start_count) && current_time >= end_count)
296 else if (current_time < start_count && current_time > end_count)
302 static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
304 struct windowsaio_data *wd = td->io_ops_data;
305 return wd->aio_events[event];
308 static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
310 const struct timespec *t)
312 struct windowsaio_data *wd = td->io_ops_data;
313 unsigned int dequeued = 0;
316 struct fio_overlapped *fov;
317 DWORD start_count = 0;
323 mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
324 start_count = GetTickCount();
325 end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
329 io_u_qiter(&td->io_u_all, io_u, i) {
330 if (!(io_u->flags & IO_U_F_FLIGHT))
333 fov = (struct fio_overlapped*)io_u->engine_data;
335 if (fov->io_complete) {
336 fov->io_complete = FALSE;
337 wd->aio_events[dequeued] = io_u;
345 if (dequeued < min) {
346 status = WaitForSingleObject(wd->iocomplete_event, mswait);
347 if (status != WAIT_OBJECT_0 && dequeued >= min)
351 if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count)))
358 static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u)
360 struct fio_overlapped *o = io_u->engine_data;
361 LPOVERLAPPED lpOvl = &o->o;
362 BOOL success = FALSE;
363 int rc = FIO_Q_COMPLETED;
365 fio_ro_check(td, io_u);
368 lpOvl->InternalHigh = 0;
369 lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
370 lpOvl->OffsetHigh = io_u->offset >> 32;
372 switch (io_u->ddir) {
374 success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, NULL, lpOvl);
377 success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, NULL, lpOvl);
381 case DDIR_SYNC_FILE_RANGE:
382 success = FlushFileBuffers(io_u->file->hFile);
384 log_err("windowsaio: failed to flush file buffers\n");
385 io_u->error = win_to_posix_error(GetLastError());
388 return FIO_Q_COMPLETED;
391 log_err("windowsaio: manual TRIM isn't supported on Windows\n");
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_posix_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_posix_error(GetLastError());
440 fov->io_complete = TRUE;
441 SetEvent(wd->iocomplete_event);
442 } while (ctx->wd->iothread_running);
444 CloseHandle(ctx->iocp);
449 static void fio_windowsaio_io_u_free(struct thread_data *td, struct io_u *io_u)
451 struct fio_overlapped *o = io_u->engine_data;
454 io_u->engine_data = NULL;
459 static int fio_windowsaio_io_u_init(struct thread_data *td, struct io_u *io_u)
461 struct fio_overlapped *o;
463 o = malloc(sizeof(*o));
464 o->io_complete = FALSE;
467 io_u->engine_data = o;
471 static struct ioengine_ops ioengine = {
472 .name = "windowsaio",
473 .version = FIO_IOOPS_VERSION,
474 .init = fio_windowsaio_init,
475 .queue = fio_windowsaio_queue,
476 .getevents = fio_windowsaio_getevents,
477 .event = fio_windowsaio_event,
478 .cleanup = fio_windowsaio_cleanup,
479 .open_file = fio_windowsaio_open_file,
480 .close_file = fio_windowsaio_close_file,
481 .get_file_size = generic_get_file_size,
482 .io_u_init = fio_windowsaio_io_u_init,
483 .io_u_free = fio_windowsaio_io_u_free,
486 static void fio_init fio_windowsaio_register(void)
488 register_ioengine(&ioengine);
491 static void fio_exit fio_windowsaio_unregister(void)
493 unregister_ioengine(&ioengine);