1/*
2 * libaio engine
3 *
4 * IO engine using the Linux native aio interface.
5 *
6 */
7#include <stdio.h>
8#include <stdlib.h>
9#include <unistd.h>
10#include <errno.h>
11#include <assert.h>
12#include <libaio.h>
13
14#include "../fio.h"
15
16static int fio_libaio_commit(struct thread_data *td);
17
18struct libaio_data {
19	io_context_t aio_ctx;
20	struct io_event *aio_events;
21	struct iocb **iocbs;
22	struct io_u **io_us;
23
24	/*
25	 * Basic ring buffer. 'head' is incremented in _queue(), and
26	 * 'tail' is incremented in _commit(). We keep 'queued' so
27	 * that we know if the ring is full or empty, when
28	 * 'head' == 'tail'. 'entries' is the ring size, and
29	 * 'is_pow2' is just an optimization to use AND instead of
30	 * modulus to get the remainder on ring increment.
31	 */
32	int is_pow2;
33	unsigned int entries;
34	unsigned int queued;
35	unsigned int head;
36	unsigned int tail;
37};
38
39struct libaio_options {
40	void *pad;
41	unsigned int userspace_reap;
42};
43
44static struct fio_option options[] = {
45	{
46		.name	= "userspace_reap",
47		.lname	= "Libaio userspace reaping",
48		.type	= FIO_OPT_STR_SET,
49		.off1	= offsetof(struct libaio_options, userspace_reap),
50		.help	= "Use alternative user-space reap implementation",
51		.category = FIO_OPT_C_ENGINE,
52		.group	= FIO_OPT_G_LIBAIO,
53	},
54	{
55		.name	= NULL,
56	},
57};
58
59static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
60			    unsigned int add)
61{
62	if (ld->is_pow2)
63		*val = (*val + add) & (ld->entries - 1);
64	else
65		*val = (*val + add) % ld->entries;
66}
67
68static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
69{
70	struct fio_file *f = io_u->file;
71
72	if (io_u->ddir == DDIR_READ)
73		io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
74	else if (io_u->ddir == DDIR_WRITE)
75		io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
76	else if (ddir_sync(io_u->ddir))
77		io_prep_fsync(&io_u->iocb, f->fd);
78
79	return 0;
80}
81
82static struct io_u *fio_libaio_event(struct thread_data *td, int event)
83{
84	struct libaio_data *ld = td->io_ops->data;
85	struct io_event *ev;
86	struct io_u *io_u;
87
88	ev = ld->aio_events + event;
89	io_u = container_of(ev->obj, struct io_u, iocb);
90
91	if (ev->res != io_u->xfer_buflen) {
92		if (ev->res > io_u->xfer_buflen)
93			io_u->error = -ev->res;
94		else
95			io_u->resid = io_u->xfer_buflen - ev->res;
96	} else
97		io_u->error = 0;
98
99	return io_u;
100}
101
102struct aio_ring {
103	unsigned id;		 /** kernel internal index number */
104	unsigned nr;		 /** number of io_events */
105	unsigned head;
106	unsigned tail;
107
108	unsigned magic;
109	unsigned compat_features;
110	unsigned incompat_features;
111	unsigned header_length;	/** size of aio_ring */
112
113	struct io_event events[0];
114};
115
116#define AIO_RING_MAGIC	0xa10a10a1
117
118static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
119			     struct io_event *events)
120{
121	long i = 0;
122	unsigned head;
123	struct aio_ring *ring = (struct aio_ring*) aio_ctx;
124
125	while (i < max) {
126		head = ring->head;
127
128		if (head == ring->tail) {
129			/* There are no more completions */
130			break;
131		} else {
132			/* There is another completion to reap */
133			events[i] = ring->events[head];
134			read_barrier();
135			ring->head = (head + 1) % ring->nr;
136			i++;
137		}
138	}
139
140	return i;
141}
142
143static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
144				unsigned int max, const struct timespec *t)
145{
146	struct libaio_data *ld = td->io_ops->data;
147	struct libaio_options *o = td->eo;
148	unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
149	struct timespec __lt, *lt = NULL;
150	int r, events = 0;
151
152	if (t) {
153		__lt = *t;
154		lt = &__lt;
155	}
156
157	do {
158		if (o->userspace_reap == 1
159		    && actual_min == 0
160		    && ((struct aio_ring *)(ld->aio_ctx))->magic
161				== AIO_RING_MAGIC) {
162			r = user_io_getevents(ld->aio_ctx, max,
163				ld->aio_events + events);
164		} else {
165			r = io_getevents(ld->aio_ctx, actual_min,
166				max, ld->aio_events + events, lt);
167		}
168		if (r > 0)
169			events += r;
170		else if ((min && r == 0) || r == -EAGAIN) {
171			fio_libaio_commit(td);
172			usleep(100);
173		} else if (r != -EINTR)
174			break;
175	} while (events < min);
176
177	return r < 0 ? r : events;
178}
179
180static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
181{
182	struct libaio_data *ld = td->io_ops->data;
183
184	fio_ro_check(td, io_u);
185
186	if (ld->queued == td->o.iodepth)
187		return FIO_Q_BUSY;
188
189	/*
190	 * fsync is tricky, since it can fail and we need to do it
191	 * serialized with other io. the reason is that linux doesn't
192	 * support aio fsync yet. So return busy for the case where we
193	 * have pending io, to let fio complete those first.
194	 */
195	if (ddir_sync(io_u->ddir)) {
196		if (ld->queued)
197			return FIO_Q_BUSY;
198
199		do_io_u_sync(td, io_u);
200		return FIO_Q_COMPLETED;
201	}
202
203	if (io_u->ddir == DDIR_TRIM) {
204		if (ld->queued)
205			return FIO_Q_BUSY;
206
207		do_io_u_trim(td, io_u);
208		return FIO_Q_COMPLETED;
209	}
210
211	ld->iocbs[ld->head] = &io_u->iocb;
212	ld->io_us[ld->head] = io_u;
213	ring_inc(ld, &ld->head, 1);
214	ld->queued++;
215	return FIO_Q_QUEUED;
216}
217
218static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
219			      unsigned int nr)
220{
221	struct timeval now;
222	unsigned int i;
223
224	if (!fio_fill_issue_time(td))
225		return;
226
227	fio_gettime(&now, NULL);
228
229	for (i = 0; i < nr; i++) {
230		struct io_u *io_u = io_us[i];
231
232		memcpy(&io_u->issue_time, &now, sizeof(now));
233		io_u_queued(td, io_u);
234	}
235}
236
237static int fio_libaio_commit(struct thread_data *td)
238{
239	struct libaio_data *ld = td->io_ops->data;
240	struct iocb **iocbs;
241	struct io_u **io_us;
242	struct timeval tv;
243	int ret, wait_start = 0;
244
245	if (!ld->queued)
246		return 0;
247
248	do {
249		long nr = ld->queued;
250
251		nr = min((unsigned int) nr, ld->entries - ld->tail);
252		io_us = ld->io_us + ld->tail;
253		iocbs = ld->iocbs + ld->tail;
254
255		ret = io_submit(ld->aio_ctx, nr, iocbs);
256		if (ret > 0) {
257			fio_libaio_queued(td, io_us, ret);
258			io_u_mark_submit(td, ret);
259
260			ld->queued -= ret;
261			ring_inc(ld, &ld->tail, ret);
262			ret = 0;
263			wait_start = 0;
264		} else if (ret == -EINTR || !ret) {
265			if (!ret)
266				io_u_mark_submit(td, ret);
267			wait_start = 0;
268			continue;
269		} else if (ret == -EAGAIN) {
270			/*
271			 * If we get EAGAIN, we should break out without
272			 * error and let the upper layer reap some
273			 * events for us. If we have no queued IO, we
274			 * must loop here. If we loop for more than 30s,
275			 * just error out, something must be buggy in the
276			 * IO path.
277			 */
278			if (ld->queued) {
279				ret = 0;
280				break;
281			}
282			if (!wait_start) {
283				fio_gettime(&tv, NULL);
284				wait_start = 1;
285			} else if (mtime_since_now(&tv) > 30000) {
286				log_err("fio: aio appears to be stalled, giving up\n");
287				break;
288			}
289			usleep(1);
290			continue;
291		} else if (ret == -ENOMEM) {
292			/*
293			 * If we get -ENOMEM, reap events if we can. If
294			 * we cannot, treat it as a fatal event since there's
295			 * nothing we can do about it.
296			 */
297			if (ld->queued)
298				ret = 0;
299			break;
300		} else
301			break;
302	} while (ld->queued);
303
304	return ret;
305}
306
307static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
308{
309	struct libaio_data *ld = td->io_ops->data;
310
311	return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
312}
313
314static void fio_libaio_cleanup(struct thread_data *td)
315{
316	struct libaio_data *ld = td->io_ops->data;
317
318	if (ld) {
319		io_destroy(ld->aio_ctx);
320		free(ld->aio_events);
321		free(ld->iocbs);
322		free(ld->io_us);
323		free(ld);
324	}
325}
326
327static int fio_libaio_init(struct thread_data *td)
328{
329	struct libaio_options *o = td->eo;
330	struct libaio_data *ld;
331	int err = 0;
332
333	ld = calloc(1, sizeof(*ld));
334
335	/*
336	 * First try passing in 0 for queue depth, since we don't
337	 * care about the user ring. If that fails, the kernel is too old
338	 * and we need the right depth.
339	 */
340	if (!o->userspace_reap)
341		err = io_queue_init(INT_MAX, &ld->aio_ctx);
342	if (o->userspace_reap || err == -EINVAL)
343		err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
344	if (err) {
345		td_verror(td, -err, "io_queue_init");
346		log_err("fio: check /proc/sys/fs/aio-max-nr\n");
347		free(ld);
348		return 1;
349	}
350
351	ld->entries = td->o.iodepth;
352	ld->is_pow2 = is_power_of_2(ld->entries);
353	ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
354	ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
355	ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
356
357	td->io_ops->data = ld;
358	return 0;
359}
360
361static struct ioengine_ops ioengine = {
362	.name			= "libaio",
363	.version		= FIO_IOOPS_VERSION,
364	.init			= fio_libaio_init,
365	.prep			= fio_libaio_prep,
366	.queue			= fio_libaio_queue,
367	.commit			= fio_libaio_commit,
368	.cancel			= fio_libaio_cancel,
369	.getevents		= fio_libaio_getevents,
370	.event			= fio_libaio_event,
371	.cleanup		= fio_libaio_cleanup,
372	.open_file		= generic_open_file,
373	.close_file		= generic_close_file,
374	.get_file_size		= generic_get_file_size,
375	.options		= options,
376	.option_struct_size	= sizeof(struct libaio_options),
377};
378
379static void fio_init fio_libaio_register(void)
380{
381	register_ioengine(&ioengine);
382}
383
384static void fio_exit fio_libaio_unregister(void)
385{
386	unregister_ioengine(&ioengine);
387}
388