1/*
2 * windowsaio engine
3 *
4 * IO engine using Windows IO Completion Ports.
5 */
6
7#include <stdio.h>
8#include <stdlib.h>
9#include <unistd.h>
10#include <signal.h>
11#include <errno.h>
12
13#include "../fio.h"
14
15typedef BOOL (WINAPI *CANCELIOEX)(HANDLE hFile, LPOVERLAPPED lpOverlapped);
16
17int geterrno_from_win_error (DWORD code, int deferrno);
18
19struct fio_overlapped {
20	OVERLAPPED o;
21	struct io_u *io_u;
22	BOOL io_complete;
23};
24
25struct windowsaio_data {
26	struct io_u **aio_events;
27	HANDLE iocp;
28	HANDLE iothread;
29	HANDLE iocomplete_event;
30	BOOL iothread_running;
31};
32
33struct thread_ctx {
34	HANDLE iocp;
35	struct windowsaio_data *wd;
36};
37
38static BOOL timeout_expired(DWORD start_count, DWORD end_count);
39static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
40					unsigned int max, struct timespec *t);
41static struct io_u *fio_windowsaio_event(struct thread_data *td, int event);
42static int fio_windowsaio_queue(struct thread_data *td,
43				  struct io_u *io_u);
44static void fio_windowsaio_cleanup(struct thread_data *td);
45static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter);
46static int fio_windowsaio_init(struct thread_data *td);
47static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f);
48static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f);
49
50static int fio_windowsaio_init(struct thread_data *td)
51{
52	struct windowsaio_data *wd;
53	int rc = 0;
54
55	wd = calloc(1, sizeof(struct windowsaio_data));
56	if (wd == NULL) {
57		 log_err("windowsaio: failed to allocate memory for engine data\n");
58		rc = 1;
59	}
60
61	if (!rc) {
62		wd->aio_events = malloc(td->o.iodepth * sizeof(struct io_u*));
63		if (wd->aio_events == NULL) {
64			log_err("windowsaio: failed to allocate memory for aio events list\n");
65			rc = 1;
66		}
67	}
68
69	if (!rc) {
70		/* Create an auto-reset event */
71		wd->iocomplete_event = CreateEvent(NULL, FALSE, FALSE, NULL);
72		if (wd->iocomplete_event == NULL) {
73			log_err("windowsaio: failed to create io complete event handle\n");
74			rc = 1;
75		}
76	}
77
78	if (rc) {
79		if (wd != NULL) {
80			if (wd->aio_events != NULL)
81				free(wd->aio_events);
82
83			free(wd);
84		}
85	}
86
87	td->io_ops->data = wd;
88
89	if (!rc) {
90		struct thread_ctx *ctx;
91		struct windowsaio_data *wd;
92		HANDLE hFile;
93
94		hFile = CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, 0, 0);
95		if (hFile == INVALID_HANDLE_VALUE) {
96			log_err("windowsaio: failed to create io completion port\n");
97			rc = 1;
98		}
99
100		wd = td->io_ops->data;
101		wd->iothread_running = TRUE;
102		wd->iocp = hFile;
103
104		if (!rc)
105			ctx = malloc(sizeof(struct thread_ctx));
106
107		if (!rc && ctx == NULL)
108		{
109			log_err("windowsaio: failed to allocate memory for thread context structure\n");
110			CloseHandle(hFile);
111			rc = 1;
112		}
113
114		if (!rc)
115		{
116			ctx->iocp = hFile;
117			ctx->wd = wd;
118			wd->iothread = CreateThread(NULL, 0, IoCompletionRoutine, ctx, 0, NULL);
119			if (wd->iothread == NULL)
120				log_err("windowsaio: failed to create io completion thread\n");
121		}
122
123		if (rc || wd->iothread == NULL)
124			rc = 1;
125	}
126
127	return rc;
128}
129
130static void fio_windowsaio_cleanup(struct thread_data *td)
131{
132	struct windowsaio_data *wd;
133
134	wd = td->io_ops->data;
135
136	if (wd != NULL) {
137		wd->iothread_running = FALSE;
138		WaitForSingleObject(wd->iothread, INFINITE);
139
140		CloseHandle(wd->iothread);
141		CloseHandle(wd->iocomplete_event);
142
143		free(wd->aio_events);
144		free(wd);
145
146		td->io_ops->data = NULL;
147	}
148}
149
150
151static int fio_windowsaio_open_file(struct thread_data *td, struct fio_file *f)
152{
153	int rc = 0;
154	DWORD flags = FILE_FLAG_POSIX_SEMANTICS | FILE_FLAG_OVERLAPPED;
155	DWORD sharemode = FILE_SHARE_READ | FILE_SHARE_WRITE;
156	DWORD openmode = OPEN_ALWAYS;
157	DWORD access;
158
159	dprint(FD_FILE, "fd open %s\n", f->file_name);
160
161	if (f->filetype == FIO_TYPE_PIPE) {
162		log_err("windowsaio: pipes are not supported\n");
163		return 1;
164	}
165
166	if (!strcmp(f->file_name, "-")) {
167		log_err("windowsaio: can't read/write to stdin/out\n");
168		return 1;
169	}
170
171	if (td->o.odirect)
172		flags |= FILE_FLAG_NO_BUFFERING;
173	if (td->o.sync_io)
174		flags |= FILE_FLAG_WRITE_THROUGH;
175
176	/*
177	 * Inform Windows whether we're going to be doing sequential or
178	 * random io so it can tune the Cache Manager
179	 */
180	if (td->o.td_ddir == TD_DDIR_READ  ||
181		td->o.td_ddir == TD_DDIR_WRITE)
182		flags |= FILE_FLAG_SEQUENTIAL_SCAN;
183	else
184		flags |= FILE_FLAG_RANDOM_ACCESS;
185
186	if (!td_write(td) || read_only)
187		access = GENERIC_READ;
188	else
189		access = (GENERIC_READ | GENERIC_WRITE);
190
191	if (td->o.create_on_open)
192		openmode = OPEN_ALWAYS;
193	else
194		openmode = OPEN_EXISTING;
195
196	f->hFile = CreateFile(f->file_name, access, sharemode,
197		NULL, openmode, flags, NULL);
198
199	if (f->hFile == INVALID_HANDLE_VALUE) {
200		log_err("windowsaio: failed to open file \"%s\"\n", f->file_name);
201		rc = 1;
202	}
203
204	/* Only set up the completion port and thread if we're not just
205	 * querying the device size */
206	if (!rc && td->io_ops->data != NULL) {
207		struct windowsaio_data *wd;
208
209		wd = td->io_ops->data;
210
211		if (CreateIoCompletionPort(f->hFile, wd->iocp, 0, 0) == NULL) {
212			log_err("windowsaio: failed to create io completion port\n");
213			rc = 1;
214		}
215	}
216
217	return rc;
218}
219
220static int fio_windowsaio_close_file(struct thread_data fio_unused *td, struct fio_file *f)
221{
222	int rc = 0;
223
224	dprint(FD_FILE, "fd close %s\n", f->file_name);
225
226	if (f->hFile != INVALID_HANDLE_VALUE) {
227		if (!CloseHandle(f->hFile)) {
228			log_info("windowsaio: failed to close file handle for \"%s\"\n", f->file_name);
229			rc = 1;
230		}
231	}
232
233	f->hFile = INVALID_HANDLE_VALUE;
234	return rc;
235}
236
237static BOOL timeout_expired(DWORD start_count, DWORD end_count)
238{
239	BOOL expired = FALSE;
240	DWORD current_time;
241
242	current_time = GetTickCount();
243
244	if ((end_count > start_count) && current_time >= end_count)
245		expired = TRUE;
246	else if (current_time < start_count && current_time > end_count)
247		expired = TRUE;
248
249	return expired;
250}
251
252static struct io_u* fio_windowsaio_event(struct thread_data *td, int event)
253{
254	struct windowsaio_data *wd = td->io_ops->data;
255	return wd->aio_events[event];
256}
257
258static int fio_windowsaio_getevents(struct thread_data *td, unsigned int min,
259					unsigned int max, struct timespec *t)
260{
261	struct windowsaio_data *wd = td->io_ops->data;
262	unsigned int dequeued = 0;
263	struct io_u *io_u;
264	int i;
265	struct fio_overlapped *fov;
266	DWORD start_count = 0;
267	DWORD end_count = 0;
268	DWORD status;
269	DWORD mswait = 250;
270
271	if (t != NULL) {
272		mswait = (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
273		start_count = GetTickCount();
274		end_count = start_count + (t->tv_sec * 1000) + (t->tv_nsec / 1000000);
275	}
276
277	do {
278		io_u_qiter(&td->io_u_all, io_u, i) {
279			if (!(io_u->flags & IO_U_F_FLIGHT))
280				continue;
281
282			fov = (struct fio_overlapped*)io_u->engine_data;
283
284			if (fov->io_complete) {
285				fov->io_complete = FALSE;
286				ResetEvent(fov->o.hEvent);
287				wd->aio_events[dequeued] = io_u;
288				dequeued++;
289			}
290
291			if (dequeued >= min)
292				break;
293		}
294
295		if (dequeued < min) {
296			status = WaitForSingleObject(wd->iocomplete_event, mswait);
297			if (status != WAIT_OBJECT_0 && dequeued >= min)
298				break;
299		}
300
301		if (dequeued >= min || (t != NULL && timeout_expired(start_count, end_count)))
302			break;
303	} while (1);
304
305	return dequeued;
306}
307
308static int fio_windowsaio_queue(struct thread_data *td, struct io_u *io_u)
309{
310	struct fio_overlapped *o = io_u->engine_data;
311	LPOVERLAPPED lpOvl = &o->o;
312	DWORD iobytes;
313	BOOL success = FALSE;
314	int rc = FIO_Q_COMPLETED;
315
316	fio_ro_check(td, io_u);
317
318	lpOvl->Internal = STATUS_PENDING;
319	lpOvl->InternalHigh = 0;
320	lpOvl->Offset = io_u->offset & 0xFFFFFFFF;
321	lpOvl->OffsetHigh = io_u->offset >> 32;
322
323	switch (io_u->ddir) {
324	case DDIR_WRITE:
325		success = WriteFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
326		break;
327	case DDIR_READ:
328		success = ReadFile(io_u->file->hFile, io_u->xfer_buf, io_u->xfer_buflen, &iobytes, lpOvl);
329		break;
330	case DDIR_SYNC:
331	case DDIR_DATASYNC:
332	case DDIR_SYNC_FILE_RANGE:
333		success = FlushFileBuffers(io_u->file->hFile);
334		if (!success) {
335			log_err("windowsaio: failed to flush file buffers\n");
336			io_u->error = win_to_posix_error(GetLastError());
337		}
338
339		return FIO_Q_COMPLETED;
340		break;
341	case DDIR_TRIM:
342		log_err("windowsaio: manual TRIM isn't supported on Windows\n");
343		io_u->error = 1;
344		io_u->resid = io_u->xfer_buflen;
345		return FIO_Q_COMPLETED;
346		break;
347	default:
348		assert(0);
349		break;
350	}
351
352	if (success || GetLastError() == ERROR_IO_PENDING)
353		rc = FIO_Q_QUEUED;
354	else {
355		io_u->error = win_to_posix_error(GetLastError());
356		io_u->resid = io_u->xfer_buflen;
357	}
358
359	return rc;
360}
361
362/* Runs as a thread and waits for queued IO to complete */
363static DWORD WINAPI IoCompletionRoutine(LPVOID lpParameter)
364{
365	OVERLAPPED *ovl;
366	struct fio_overlapped *fov;
367	struct io_u *io_u;
368	struct windowsaio_data *wd;
369	struct thread_ctx *ctx;
370	ULONG_PTR ulKey = 0;
371	DWORD bytes;
372
373	ctx = (struct thread_ctx*)lpParameter;
374	wd = ctx->wd;
375
376	do {
377		if (!GetQueuedCompletionStatus(ctx->iocp, &bytes, &ulKey, &ovl, 250) && ovl == NULL)
378			continue;
379
380		fov = CONTAINING_RECORD(ovl, struct fio_overlapped, o);
381		io_u = fov->io_u;
382
383		if (ovl->Internal == ERROR_SUCCESS) {
384			io_u->resid = io_u->xfer_buflen - ovl->InternalHigh;
385			io_u->error = 0;
386		} else {
387			io_u->resid = io_u->xfer_buflen;
388			io_u->error = win_to_posix_error(GetLastError());
389		}
390
391		fov->io_complete = TRUE;
392		SetEvent(wd->iocomplete_event);
393	} while (ctx->wd->iothread_running);
394
395	CloseHandle(ctx->iocp);
396	free(ctx);
397	return 0;
398}
399
400static void fio_windowsaio_io_u_free(struct thread_data *td, struct io_u *io_u)
401{
402	struct fio_overlapped *o = io_u->engine_data;
403
404	if (o) {
405		CloseHandle(o->o.hEvent);
406		io_u->engine_data = NULL;
407		free(o);
408	}
409}
410
411static int fio_windowsaio_io_u_init(struct thread_data *td, struct io_u *io_u)
412{
413	struct fio_overlapped *o;
414
415	o = malloc(sizeof(*o));
416	o->io_complete = FALSE;
417	o->io_u = io_u;
418	o->o.hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
419	if (o->o.hEvent == NULL) {
420		log_err("windowsaio: failed to create event handle\n");
421		free(o);
422		return 1;
423	}
424
425	io_u->engine_data = o;
426	return 0;
427}
428
429static struct ioengine_ops ioengine = {
430	.name		= "windowsaio",
431	.version	= FIO_IOOPS_VERSION,
432	.init		= fio_windowsaio_init,
433	.queue		= fio_windowsaio_queue,
434	.getevents	= fio_windowsaio_getevents,
435	.event		= fio_windowsaio_event,
436	.cleanup	= fio_windowsaio_cleanup,
437	.open_file	= fio_windowsaio_open_file,
438	.close_file	= fio_windowsaio_close_file,
439	.get_file_size	= generic_get_file_size,
440	.io_u_init	= fio_windowsaio_io_u_init,
441	.io_u_free	= fio_windowsaio_io_u_free,
442};
443
444static void fio_init fio_windowsaio_register(void)
445{
446	register_ioengine(&ioengine);
447}
448
449static void fio_exit fio_windowsaio_unregister(void)
450{
451	unregister_ioengine(&ioengine);
452}
453