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"
20 /* Should be defined in newest aio_abi.h */
21 #ifndef IOCB_FLAG_IOPRIO
22 #define IOCB_FLAG_IOPRIO (1 << 1)
25 /* Hack for libaio < 0.3.111 */
26 #ifndef CONFIG_LIBAIO_RW_FLAGS
27 #define aio_rw_flags __pad2
30 static int fio_libaio_commit(struct thread_data *td);
31 static int fio_libaio_init(struct thread_data *td);
35 struct io_event *aio_events;
39 struct io_u **io_u_index;
42 * Basic ring buffer. 'head' is incremented in _queue(), and
43 * 'tail' is incremented in _commit(). We keep 'queued' so
44 * that we know if the ring is full or empty, when
45 * 'head' == 'tail'. 'entries' is the ring size, and
46 * 'is_pow2' is just an optimization to use AND instead of
47 * modulus to get the remainder on ring increment.
58 struct libaio_options {
60 unsigned int userspace_reap;
61 struct cmdprio cmdprio;
65 static struct fio_option options[] = {
67 .name = "userspace_reap",
68 .lname = "Libaio userspace reaping",
69 .type = FIO_OPT_STR_SET,
70 .off1 = offsetof(struct libaio_options, userspace_reap),
71 .help = "Use alternative user-space reap implementation",
72 .category = FIO_OPT_C_ENGINE,
73 .group = FIO_OPT_G_LIBAIO,
75 #ifdef FIO_HAVE_IOPRIO_CLASS
77 .name = "cmdprio_percentage",
78 .lname = "high priority percentage",
80 .off1 = offsetof(struct libaio_options,
81 cmdprio.percentage[DDIR_READ]),
82 .off2 = offsetof(struct libaio_options,
83 cmdprio.percentage[DDIR_WRITE]),
86 .help = "Send high priority I/O this percentage of the time",
87 .category = FIO_OPT_C_ENGINE,
88 .group = FIO_OPT_G_LIBAIO,
91 .name = "cmdprio_class",
92 .lname = "Asynchronous I/O priority class",
94 .off1 = offsetof(struct libaio_options,
95 cmdprio.class[DDIR_READ]),
96 .off2 = offsetof(struct libaio_options,
97 cmdprio.class[DDIR_WRITE]),
98 .help = "Set asynchronous IO priority class",
99 .minval = IOPRIO_MIN_PRIO_CLASS + 1,
100 .maxval = IOPRIO_MAX_PRIO_CLASS,
102 .category = FIO_OPT_C_ENGINE,
103 .group = FIO_OPT_G_LIBAIO,
107 .lname = "Asynchronous I/O priority level",
109 .off1 = offsetof(struct libaio_options,
110 cmdprio.level[DDIR_READ]),
111 .off2 = offsetof(struct libaio_options,
112 cmdprio.level[DDIR_WRITE]),
113 .help = "Set asynchronous IO priority level",
114 .minval = IOPRIO_MIN_PRIO,
115 .maxval = IOPRIO_MAX_PRIO,
117 .category = FIO_OPT_C_ENGINE,
118 .group = FIO_OPT_G_LIBAIO,
122 .name = "cmdprio_percentage",
123 .lname = "high priority percentage",
124 .type = FIO_OPT_UNSUPPORTED,
125 .help = "Your platform does not support I/O priority classes",
128 .name = "cmdprio_class",
129 .lname = "Asynchronous I/O priority class",
130 .type = FIO_OPT_UNSUPPORTED,
131 .help = "Your platform does not support I/O priority classes",
135 .lname = "Asynchronous I/O priority level",
136 .type = FIO_OPT_UNSUPPORTED,
137 .help = "Your platform does not support I/O priority classes",
142 .lname = "RWF_NOWAIT",
143 .type = FIO_OPT_BOOL,
144 .off1 = offsetof(struct libaio_options, nowait),
145 .help = "Set RWF_NOWAIT for reads/writes",
146 .category = FIO_OPT_C_ENGINE,
147 .group = FIO_OPT_G_LIBAIO,
154 static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
158 *val = (*val + add) & (ld->entries - 1);
160 *val = (*val + add) % ld->entries;
163 static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u)
165 struct libaio_options *o = td->eo;
166 struct fio_file *f = io_u->file;
167 struct iocb *iocb = &io_u->iocb;
169 if (io_u->ddir == DDIR_READ) {
170 io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
172 iocb->aio_rw_flags |= RWF_NOWAIT;
173 } else if (io_u->ddir == DDIR_WRITE) {
174 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
176 iocb->aio_rw_flags |= RWF_NOWAIT;
177 } else if (ddir_sync(io_u->ddir))
178 io_prep_fsync(iocb, f->fd);
183 static void fio_libaio_prio_prep(struct thread_data *td, struct io_u *io_u)
185 struct libaio_options *o = td->eo;
186 struct cmdprio *cmdprio = &o->cmdprio;
187 enum fio_ddir ddir = io_u->ddir;
188 unsigned int p = cmdprio->percentage[ddir];
190 if (p && rand_between(&td->prio_state, 0, 99) < p) {
191 io_u->iocb.aio_reqprio =
192 ioprio_value(cmdprio->class[ddir], cmdprio->level[ddir]);
193 io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
194 io_u->flags |= IO_U_F_PRIORITY;
198 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
200 struct libaio_data *ld = td->io_ops_data;
204 ev = ld->aio_events + event;
205 io_u = container_of(ev->obj, struct io_u, iocb);
207 if (ev->res != io_u->xfer_buflen) {
208 if (ev->res > io_u->xfer_buflen)
209 io_u->error = -ev->res;
211 io_u->resid = io_u->xfer_buflen - ev->res;
219 unsigned id; /** kernel internal index number */
220 unsigned nr; /** number of io_events */
225 unsigned compat_features;
226 unsigned incompat_features;
227 unsigned header_length; /** size of aio_ring */
229 struct io_event events[0];
232 #define AIO_RING_MAGIC 0xa10a10a1
234 static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
235 struct io_event *events)
239 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
244 if (head == ring->tail) {
245 /* There are no more completions */
248 /* There is another completion to reap */
249 events[i] = ring->events[head];
250 atomic_store_release(&ring->head,
251 (head + 1) % ring->nr);
259 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
260 unsigned int max, const struct timespec *t)
262 struct libaio_data *ld = td->io_ops_data;
263 struct libaio_options *o = td->eo;
264 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
265 struct timespec __lt, *lt = NULL;
274 if (o->userspace_reap == 1
276 && ((struct aio_ring *)(ld->aio_ctx))->magic
278 r = user_io_getevents(ld->aio_ctx, max,
279 ld->aio_events + events);
281 r = io_getevents(ld->aio_ctx, actual_min,
282 max, ld->aio_events + events, lt);
286 else if ((min && r == 0) || r == -EAGAIN) {
287 fio_libaio_commit(td);
290 } else if (r != -EINTR)
292 } while (events < min);
294 return r < 0 ? r : events;
297 static enum fio_q_status fio_libaio_queue(struct thread_data *td,
300 struct libaio_data *ld = td->io_ops_data;
302 fio_ro_check(td, io_u);
304 if (ld->queued == td->o.iodepth)
308 * fsync is tricky, since it can fail and we need to do it
309 * serialized with other io. the reason is that linux doesn't
310 * support aio fsync yet. So return busy for the case where we
311 * have pending io, to let fio complete those first.
313 if (ddir_sync(io_u->ddir)) {
317 do_io_u_sync(td, io_u);
318 return FIO_Q_COMPLETED;
321 if (io_u->ddir == DDIR_TRIM) {
325 do_io_u_trim(td, io_u);
326 io_u_mark_submit(td, 1);
327 io_u_mark_complete(td, 1);
328 return FIO_Q_COMPLETED;
332 fio_libaio_prio_prep(td, io_u);
334 ld->iocbs[ld->head] = &io_u->iocb;
335 ld->io_us[ld->head] = io_u;
336 ring_inc(ld, &ld->head, 1);
341 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
347 if (!fio_fill_issue_time(td))
350 fio_gettime(&now, NULL);
352 for (i = 0; i < nr; i++) {
353 struct io_u *io_u = io_us[i];
355 memcpy(&io_u->issue_time, &now, sizeof(now));
356 io_u_queued(td, io_u);
360 static int fio_libaio_commit(struct thread_data *td)
362 struct libaio_data *ld = td->io_ops_data;
366 int ret, wait_start = 0;
372 long nr = ld->queued;
374 nr = min((unsigned int) nr, ld->entries - ld->tail);
375 io_us = ld->io_us + ld->tail;
376 iocbs = ld->iocbs + ld->tail;
378 ret = io_submit(ld->aio_ctx, nr, iocbs);
380 fio_libaio_queued(td, io_us, ret);
381 io_u_mark_submit(td, ret);
384 ring_inc(ld, &ld->tail, ret);
387 } else if (ret == -EINTR || !ret) {
389 io_u_mark_submit(td, ret);
392 } else if (ret == -EAGAIN) {
394 * If we get EAGAIN, we should break out without
395 * error and let the upper layer reap some
396 * events for us. If we have no queued IO, we
397 * must loop here. If we loop for more than 30s,
398 * just error out, something must be buggy in the
406 fio_gettime(&ts, NULL);
408 } else if (mtime_since_now(&ts) > 30000) {
409 log_err("fio: aio appears to be stalled, giving up\n");
414 } else if (ret == -ENOMEM) {
416 * If we get -ENOMEM, reap events if we can. If
417 * we cannot, treat it as a fatal event since there's
418 * nothing we can do about it.
425 } while (ld->queued);
430 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
432 struct libaio_data *ld = td->io_ops_data;
434 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
437 static void fio_libaio_cleanup(struct thread_data *td)
439 struct libaio_data *ld = td->io_ops_data;
443 * Work-around to avoid huge RCU stalls at exit time. If we
444 * don't do this here, then it'll be torn down by exit_aio().
445 * But for that case we can parallellize the freeing, thus
446 * speeding it up a lot.
448 if (!(td->flags & TD_F_CHILD))
449 io_destroy(ld->aio_ctx);
450 free(ld->aio_events);
457 static int fio_libaio_post_init(struct thread_data *td)
459 struct libaio_data *ld = td->io_ops_data;
462 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
464 td_verror(td, -err, "io_queue_init");
471 static int fio_libaio_init(struct thread_data *td)
473 struct libaio_data *ld;
474 struct libaio_options *o = td->eo;
475 struct cmdprio *cmdprio = &o->cmdprio;
478 ld = calloc(1, sizeof(*ld));
480 ld->entries = td->o.iodepth;
481 ld->is_pow2 = is_power_of_2(ld->entries);
482 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
483 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
484 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
486 td->io_ops_data = ld;
488 ret = fio_cmdprio_init(td, cmdprio, &ld->use_cmdprio);
490 td_verror(td, EINVAL, "fio_libaio_init");
497 FIO_STATIC struct ioengine_ops ioengine = {
499 .version = FIO_IOOPS_VERSION,
500 .flags = FIO_ASYNCIO_SYNC_TRIM,
501 .init = fio_libaio_init,
502 .post_init = fio_libaio_post_init,
503 .prep = fio_libaio_prep,
504 .queue = fio_libaio_queue,
505 .commit = fio_libaio_commit,
506 .cancel = fio_libaio_cancel,
507 .getevents = fio_libaio_getevents,
508 .event = fio_libaio_event,
509 .cleanup = fio_libaio_cleanup,
510 .open_file = generic_open_file,
511 .close_file = generic_close_file,
512 .get_file_size = generic_get_file_size,
514 .option_struct_size = sizeof(struct libaio_options),
517 static void fio_init fio_libaio_register(void)
519 register_ioengine(&ioengine);
522 static void fio_exit fio_libaio_unregister(void)
524 unregister_ioengine(&ioengine);