4 * IO engine using the Linux native aio interface.
15 #include "../lib/pow2.h"
17 static int fio_libaio_commit(struct thread_data *td);
21 struct io_event *aio_events;
26 * Basic ring buffer. 'head' is incremented in _queue(), and
27 * 'tail' is incremented in _commit(). We keep 'queued' so
28 * that we know if the ring is full or empty, when
29 * 'head' == 'tail'. 'entries' is the ring size, and
30 * 'is_pow2' is just an optimization to use AND instead of
31 * modulus to get the remainder on ring increment.
40 struct libaio_options {
42 unsigned int userspace_reap;
45 static struct fio_option options[] = {
47 .name = "userspace_reap",
48 .lname = "Libaio userspace reaping",
49 .type = FIO_OPT_STR_SET,
50 .off1 = offsetof(struct libaio_options, userspace_reap),
51 .help = "Use alternative user-space reap implementation",
52 .category = FIO_OPT_C_ENGINE,
53 .group = FIO_OPT_G_LIBAIO,
60 static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
64 *val = (*val + add) & (ld->entries - 1);
66 *val = (*val + add) % ld->entries;
69 static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
71 struct fio_file *f = io_u->file;
73 if (io_u->ddir == DDIR_READ)
74 io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
75 else if (io_u->ddir == DDIR_WRITE)
76 io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
77 else if (ddir_sync(io_u->ddir))
78 io_prep_fsync(&io_u->iocb, f->fd);
83 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
85 struct libaio_data *ld = td->io_ops->data;
89 ev = ld->aio_events + event;
90 io_u = container_of(ev->obj, struct io_u, iocb);
92 if (ev->res != io_u->xfer_buflen) {
93 if (ev->res > io_u->xfer_buflen)
94 io_u->error = -ev->res;
96 io_u->resid = io_u->xfer_buflen - ev->res;
104 unsigned id; /** kernel internal index number */
105 unsigned nr; /** number of io_events */
110 unsigned compat_features;
111 unsigned incompat_features;
112 unsigned header_length; /** size of aio_ring */
114 struct io_event events[0];
117 #define AIO_RING_MAGIC 0xa10a10a1
119 static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
120 struct io_event *events)
124 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
129 if (head == ring->tail) {
130 /* There are no more completions */
133 /* There is another completion to reap */
134 events[i] = ring->events[head];
136 ring->head = (head + 1) % ring->nr;
144 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
145 unsigned int max, const struct timespec *t)
147 struct libaio_data *ld = td->io_ops->data;
148 struct libaio_options *o = td->eo;
149 unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
150 struct timespec __lt, *lt = NULL;
159 if (o->userspace_reap == 1
161 && ((struct aio_ring *)(ld->aio_ctx))->magic
163 r = user_io_getevents(ld->aio_ctx, max,
164 ld->aio_events + events);
166 r = io_getevents(ld->aio_ctx, actual_min,
167 max, ld->aio_events + events, lt);
171 else if ((min && r == 0) || r == -EAGAIN) {
172 fio_libaio_commit(td);
174 } else if (r != -EINTR)
176 } while (events < min);
178 return r < 0 ? r : events;
181 static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
183 struct libaio_data *ld = td->io_ops->data;
185 fio_ro_check(td, io_u);
187 if (ld->queued == td->o.iodepth)
191 * fsync is tricky, since it can fail and we need to do it
192 * serialized with other io. the reason is that linux doesn't
193 * support aio fsync yet. So return busy for the case where we
194 * have pending io, to let fio complete those first.
196 if (ddir_sync(io_u->ddir)) {
200 do_io_u_sync(td, io_u);
201 return FIO_Q_COMPLETED;
204 if (io_u->ddir == DDIR_TRIM) {
208 do_io_u_trim(td, io_u);
209 return FIO_Q_COMPLETED;
212 ld->iocbs[ld->head] = &io_u->iocb;
213 ld->io_us[ld->head] = io_u;
214 ring_inc(ld, &ld->head, 1);
219 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
225 if (!fio_fill_issue_time(td))
228 fio_gettime(&now, NULL);
230 for (i = 0; i < nr; i++) {
231 struct io_u *io_u = io_us[i];
233 memcpy(&io_u->issue_time, &now, sizeof(now));
234 io_u_queued(td, io_u);
238 static int fio_libaio_commit(struct thread_data *td)
240 struct libaio_data *ld = td->io_ops->data;
244 int ret, wait_start = 0;
250 long nr = ld->queued;
252 nr = min((unsigned int) nr, ld->entries - ld->tail);
253 io_us = ld->io_us + ld->tail;
254 iocbs = ld->iocbs + ld->tail;
256 ret = io_submit(ld->aio_ctx, nr, iocbs);
258 fio_libaio_queued(td, io_us, ret);
259 io_u_mark_submit(td, ret);
262 ring_inc(ld, &ld->tail, ret);
265 } else if (ret == -EINTR || !ret) {
267 io_u_mark_submit(td, ret);
270 } else if (ret == -EAGAIN) {
272 * If we get EAGAIN, we should break out without
273 * error and let the upper layer reap some
274 * events for us. If we have no queued IO, we
275 * must loop here. If we loop for more than 30s,
276 * just error out, something must be buggy in the
284 fio_gettime(&tv, NULL);
286 } else if (mtime_since_now(&tv) > 30000) {
287 log_err("fio: aio appears to be stalled, giving up\n");
292 } else if (ret == -ENOMEM) {
294 * If we get -ENOMEM, reap events if we can. If
295 * we cannot, treat it as a fatal event since there's
296 * nothing we can do about it.
303 } while (ld->queued);
308 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
310 struct libaio_data *ld = td->io_ops->data;
312 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
315 static void fio_libaio_cleanup(struct thread_data *td)
317 struct libaio_data *ld = td->io_ops->data;
321 * Work-around to avoid huge RCU stalls at exit time. If we
322 * don't do this here, then it'll be torn down by exit_aio().
323 * But for that case we can parallellize the freeing, thus
324 * speeding it up a lot.
326 if (!(td->flags & TD_F_CHILD))
327 io_destroy(ld->aio_ctx);
328 free(ld->aio_events);
335 static int fio_libaio_init(struct thread_data *td)
337 struct libaio_options *o = td->eo;
338 struct libaio_data *ld;
341 ld = calloc(1, sizeof(*ld));
344 * First try passing in 0 for queue depth, since we don't
345 * care about the user ring. If that fails, the kernel is too old
346 * and we need the right depth.
348 if (!o->userspace_reap)
349 err = io_queue_init(INT_MAX, &ld->aio_ctx);
350 if (o->userspace_reap || err == -EINVAL)
351 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
353 td_verror(td, -err, "io_queue_init");
354 log_err("fio: check /proc/sys/fs/aio-max-nr\n");
359 ld->entries = td->o.iodepth;
360 ld->is_pow2 = is_power_of_2(ld->entries);
361 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
362 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
363 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
365 td->io_ops->data = ld;
369 static struct ioengine_ops ioengine = {
371 .version = FIO_IOOPS_VERSION,
372 .init = fio_libaio_init,
373 .prep = fio_libaio_prep,
374 .queue = fio_libaio_queue,
375 .commit = fio_libaio_commit,
376 .cancel = fio_libaio_cancel,
377 .getevents = fio_libaio_getevents,
378 .event = fio_libaio_event,
379 .cleanup = fio_libaio_cleanup,
380 .open_file = generic_open_file,
381 .close_file = generic_close_file,
382 .get_file_size = generic_get_file_size,
384 .option_struct_size = sizeof(struct libaio_options),
387 static void fio_init fio_libaio_register(void)
389 register_ioengine(&ioengine);
392 static void fio_exit fio_libaio_unregister(void)
394 unregister_ioengine(&ioengine);