4 * IO engine using the Linux native aio interface.
12 #include <sys/resource.h>
15 #include "../lib/pow2.h"
16 #include "../optgroup.h"
17 #include "../lib/memalign.h"
19 /* Should be defined in newest aio_abi.h */
20 #ifndef IOCB_FLAG_IOPRIO
21 #define IOCB_FLAG_IOPRIO (1 << 1)
24 static int fio_libaio_commit(struct thread_data *td);
25 static int fio_libaio_init(struct thread_data *td);
29 struct io_event *aio_events;
33 struct io_u **io_u_index;
36 * Basic ring buffer. 'head' is incremented in _queue(), and
37 * 'tail' is incremented in _commit(). We keep 'queued' so
38 * that we know if the ring is full or empty, when
39 * 'head' == 'tail'. 'entries' is the ring size, and
40 * 'is_pow2' is just an optimization to use AND instead of
41 * modulus to get the remainder on ring increment.
50 struct libaio_options {
52 unsigned int userspace_reap;
53 unsigned int cmdprio_percentage;
56 static struct fio_option options[] = {
58 .name = "userspace_reap",
59 .lname = "Libaio userspace reaping",
60 .type = FIO_OPT_STR_SET,
61 .off1 = offsetof(struct libaio_options, userspace_reap),
62 .help = "Use alternative user-space reap implementation",
63 .category = FIO_OPT_C_ENGINE,
64 .group = FIO_OPT_G_LIBAIO,
66 #ifdef FIO_HAVE_IOPRIO_CLASS
68 .name = "cmdprio_percentage",
69 .lname = "high priority percentage",
71 .off1 = offsetof(struct libaio_options, cmdprio_percentage),
74 .help = "Send high priority I/O this percentage of the time",
75 .category = FIO_OPT_C_ENGINE,
76 .group = FIO_OPT_G_LIBAIO,
80 .name = "cmdprio_percentage",
81 .lname = "high priority percentage",
82 .type = FIO_OPT_UNSUPPORTED,
83 .help = "Your platform does not support I/O priority classes",
91 static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
95 *val = (*val + add) & (ld->entries - 1);
97 *val = (*val + add) % ld->entries;
100 static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
102 struct fio_file *f = io_u->file;
103 struct iocb *iocb = &io_u->iocb;
105 if (io_u->ddir == DDIR_READ) {
106 io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
107 } else if (io_u->ddir == DDIR_WRITE) {
108 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
109 } else if (ddir_sync(io_u->ddir))
110 io_prep_fsync(iocb, f->fd);
115 static void fio_libaio_prio_prep(struct thread_data *td, struct io_u *io_u)
117 struct libaio_options *o = td->eo;
118 if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
119 io_u->iocb.aio_reqprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
120 io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
121 io_u->flags |= IO_U_F_PRIORITY;
126 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
128 struct libaio_data *ld = td->io_ops_data;
132 ev = ld->aio_events + event;
133 io_u = container_of(ev->obj, struct io_u, iocb);
135 if (ev->res != io_u->xfer_buflen) {
136 if (ev->res > io_u->xfer_buflen)
137 io_u->error = -ev->res;
139 io_u->resid = io_u->xfer_buflen - ev->res;
147 unsigned id; /** kernel internal index number */
148 unsigned nr; /** number of io_events */
153 unsigned compat_features;
154 unsigned incompat_features;
155 unsigned header_length; /** size of aio_ring */
157 struct io_event events[0];
160 #define AIO_RING_MAGIC 0xa10a10a1
162 static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
163 struct io_event *events)
167 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
172 if (head == ring->tail) {
173 /* There are no more completions */
176 /* There is another completion to reap */
177 events[i] = ring->events[head];
179 ring->head = (head + 1) % ring->nr;
187 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
188 unsigned int max, const struct timespec *t)
190 struct libaio_data *ld = td->io_ops_data;
191 struct libaio_options *o = td->eo;
192 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
193 struct timespec __lt, *lt = NULL;
202 if (o->userspace_reap == 1
204 && ((struct aio_ring *)(ld->aio_ctx))->magic
206 r = user_io_getevents(ld->aio_ctx, max,
207 ld->aio_events + events);
209 r = io_getevents(ld->aio_ctx, actual_min,
210 max, ld->aio_events + events, lt);
214 else if ((min && r == 0) || r == -EAGAIN) {
215 fio_libaio_commit(td);
218 } else if (r != -EINTR)
220 } while (events < min);
222 return r < 0 ? r : events;
225 static enum fio_q_status fio_libaio_queue(struct thread_data *td,
228 struct libaio_data *ld = td->io_ops_data;
229 struct libaio_options *o = td->eo;
231 fio_ro_check(td, io_u);
233 if (ld->queued == td->o.iodepth)
237 * fsync is tricky, since it can fail and we need to do it
238 * serialized with other io. the reason is that linux doesn't
239 * support aio fsync yet. So return busy for the case where we
240 * have pending io, to let fio complete those first.
242 if (ddir_sync(io_u->ddir)) {
246 do_io_u_sync(td, io_u);
247 return FIO_Q_COMPLETED;
250 if (io_u->ddir == DDIR_TRIM) {
254 do_io_u_trim(td, io_u);
255 io_u_mark_submit(td, 1);
256 io_u_mark_complete(td, 1);
257 return FIO_Q_COMPLETED;
260 if (o->cmdprio_percentage)
261 fio_libaio_prio_prep(td, io_u);
263 ld->iocbs[ld->head] = &io_u->iocb;
264 ld->io_us[ld->head] = io_u;
265 ring_inc(ld, &ld->head, 1);
270 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
276 if (!fio_fill_issue_time(td))
279 fio_gettime(&now, NULL);
281 for (i = 0; i < nr; i++) {
282 struct io_u *io_u = io_us[i];
284 memcpy(&io_u->issue_time, &now, sizeof(now));
285 io_u_queued(td, io_u);
289 static int fio_libaio_commit(struct thread_data *td)
291 struct libaio_data *ld = td->io_ops_data;
295 int ret, wait_start = 0;
301 long nr = ld->queued;
303 nr = min((unsigned int) nr, ld->entries - ld->tail);
304 io_us = ld->io_us + ld->tail;
305 iocbs = ld->iocbs + ld->tail;
307 ret = io_submit(ld->aio_ctx, nr, iocbs);
309 fio_libaio_queued(td, io_us, ret);
310 io_u_mark_submit(td, ret);
313 ring_inc(ld, &ld->tail, ret);
316 } else if (ret == -EINTR || !ret) {
318 io_u_mark_submit(td, ret);
321 } else if (ret == -EAGAIN) {
323 * If we get EAGAIN, we should break out without
324 * error and let the upper layer reap some
325 * events for us. If we have no queued IO, we
326 * must loop here. If we loop for more than 30s,
327 * just error out, something must be buggy in the
335 fio_gettime(&ts, NULL);
337 } else if (mtime_since_now(&ts) > 30000) {
338 log_err("fio: aio appears to be stalled, giving up\n");
343 } else if (ret == -ENOMEM) {
345 * If we get -ENOMEM, reap events if we can. If
346 * we cannot, treat it as a fatal event since there's
347 * nothing we can do about it.
354 } while (ld->queued);
359 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
361 struct libaio_data *ld = td->io_ops_data;
363 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
366 static void fio_libaio_cleanup(struct thread_data *td)
368 struct libaio_data *ld = td->io_ops_data;
372 * Work-around to avoid huge RCU stalls at exit time. If we
373 * don't do this here, then it'll be torn down by exit_aio().
374 * But for that case we can parallellize the freeing, thus
375 * speeding it up a lot.
377 if (!(td->flags & TD_F_CHILD))
378 io_destroy(ld->aio_ctx);
379 free(ld->aio_events);
386 static int fio_libaio_post_init(struct thread_data *td)
388 struct libaio_data *ld = td->io_ops_data;
391 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
393 td_verror(td, -err, "io_queue_init");
400 static int fio_libaio_init(struct thread_data *td)
402 struct libaio_data *ld;
403 struct thread_options *to = &td->o;
404 struct libaio_options *o = td->eo;
406 ld = calloc(1, sizeof(*ld));
408 ld->entries = td->o.iodepth;
409 ld->is_pow2 = is_power_of_2(ld->entries);
410 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
411 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
412 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
414 td->io_ops_data = ld;
416 * Check for option conflicts
418 if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) &&
419 o->cmdprio_percentage != 0) {
420 log_err("%s: cmdprio_percentage option and mutually exclusive "
421 "prio or prioclass option is set, exiting\n", to->name);
422 td_verror(td, EINVAL, "fio_libaio_init");
428 static struct ioengine_ops ioengine = {
430 .version = FIO_IOOPS_VERSION,
431 .flags = FIO_ASYNCIO_SYNC_TRIM,
432 .init = fio_libaio_init,
433 .post_init = fio_libaio_post_init,
434 .prep = fio_libaio_prep,
435 .queue = fio_libaio_queue,
436 .commit = fio_libaio_commit,
437 .cancel = fio_libaio_cancel,
438 .getevents = fio_libaio_getevents,
439 .event = fio_libaio_event,
440 .cleanup = fio_libaio_cleanup,
441 .open_file = generic_open_file,
442 .close_file = generic_close_file,
443 .get_file_size = generic_get_file_size,
445 .option_struct_size = sizeof(struct libaio_options),
448 static void fio_init fio_libaio_register(void)
450 register_ioengine(&ioengine);
453 static void fio_exit fio_libaio_unregister(void)
455 unregister_ioengine(&ioengine);