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.
56 struct libaio_options {
57 struct thread_data *td;
58 unsigned int userspace_reap;
59 struct cmdprio cmdprio;
63 static int str_cmdprio_bssplit_cb(void *data, const char *input)
65 struct libaio_options *o = data;
66 struct thread_data *td = o->td;
67 struct cmdprio *cmdprio = &o->cmdprio;
69 return fio_cmdprio_bssplit_parse(td, input, cmdprio);
72 static struct fio_option options[] = {
74 .name = "userspace_reap",
75 .lname = "Libaio userspace reaping",
76 .type = FIO_OPT_STR_SET,
77 .off1 = offsetof(struct libaio_options, userspace_reap),
78 .help = "Use alternative user-space reap implementation",
79 .category = FIO_OPT_C_ENGINE,
80 .group = FIO_OPT_G_LIBAIO,
82 #ifdef FIO_HAVE_IOPRIO_CLASS
84 .name = "cmdprio_percentage",
85 .lname = "high priority percentage",
87 .off1 = offsetof(struct libaio_options,
88 cmdprio.percentage[DDIR_READ]),
89 .off2 = offsetof(struct libaio_options,
90 cmdprio.percentage[DDIR_WRITE]),
93 .help = "Send high priority I/O this percentage of the time",
94 .category = FIO_OPT_C_ENGINE,
95 .group = FIO_OPT_G_LIBAIO,
98 .name = "cmdprio_class",
99 .lname = "Asynchronous I/O priority class",
101 .off1 = offsetof(struct libaio_options,
102 cmdprio.class[DDIR_READ]),
103 .off2 = offsetof(struct libaio_options,
104 cmdprio.class[DDIR_WRITE]),
105 .help = "Set asynchronous IO priority class",
106 .minval = IOPRIO_MIN_PRIO_CLASS + 1,
107 .maxval = IOPRIO_MAX_PRIO_CLASS,
109 .category = FIO_OPT_C_ENGINE,
110 .group = FIO_OPT_G_LIBAIO,
114 .lname = "Asynchronous I/O priority level",
116 .off1 = offsetof(struct libaio_options,
117 cmdprio.level[DDIR_READ]),
118 .off2 = offsetof(struct libaio_options,
119 cmdprio.level[DDIR_WRITE]),
120 .help = "Set asynchronous IO priority level",
121 .minval = IOPRIO_MIN_PRIO,
122 .maxval = IOPRIO_MAX_PRIO,
124 .category = FIO_OPT_C_ENGINE,
125 .group = FIO_OPT_G_LIBAIO,
128 .name = "cmdprio_bssplit",
129 .lname = "Priority percentage block size split",
130 .type = FIO_OPT_STR_ULL,
131 .cb = str_cmdprio_bssplit_cb,
132 .off1 = offsetof(struct libaio_options, cmdprio.bssplit),
133 .help = "Set priority percentages for different block sizes",
134 .category = FIO_OPT_C_ENGINE,
135 .group = FIO_OPT_G_LIBAIO,
139 .name = "cmdprio_percentage",
140 .lname = "high priority percentage",
141 .type = FIO_OPT_UNSUPPORTED,
142 .help = "Your platform does not support I/O priority classes",
145 .name = "cmdprio_class",
146 .lname = "Asynchronous I/O priority class",
147 .type = FIO_OPT_UNSUPPORTED,
148 .help = "Your platform does not support I/O priority classes",
152 .lname = "Asynchronous I/O priority level",
153 .type = FIO_OPT_UNSUPPORTED,
154 .help = "Your platform does not support I/O priority classes",
157 .name = "cmdprio_bssplit",
158 .lname = "Priority percentage block size split",
159 .type = FIO_OPT_UNSUPPORTED,
160 .help = "Your platform does not support I/O priority classes",
165 .lname = "RWF_NOWAIT",
166 .type = FIO_OPT_BOOL,
167 .off1 = offsetof(struct libaio_options, nowait),
168 .help = "Set RWF_NOWAIT for reads/writes",
169 .category = FIO_OPT_C_ENGINE,
170 .group = FIO_OPT_G_LIBAIO,
177 static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
181 *val = (*val + add) & (ld->entries - 1);
183 *val = (*val + add) % ld->entries;
186 static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u)
188 struct libaio_options *o = td->eo;
189 struct fio_file *f = io_u->file;
190 struct iocb *iocb = &io_u->iocb;
192 if (io_u->ddir == DDIR_READ) {
193 io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
195 iocb->aio_rw_flags |= RWF_NOWAIT;
196 } else if (io_u->ddir == DDIR_WRITE) {
197 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
199 iocb->aio_rw_flags |= RWF_NOWAIT;
200 } else if (ddir_sync(io_u->ddir))
201 io_prep_fsync(iocb, f->fd);
206 static inline void fio_libaio_cmdprio_prep(struct thread_data *td,
209 struct libaio_options *o = td->eo;
210 struct cmdprio *cmdprio = &o->cmdprio;
212 if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) {
213 io_u->iocb.aio_reqprio = io_u->ioprio;
214 io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
218 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
220 struct libaio_data *ld = td->io_ops_data;
224 ev = ld->aio_events + event;
225 io_u = container_of(ev->obj, struct io_u, iocb);
227 if (ev->res != io_u->xfer_buflen) {
228 if (ev->res > io_u->xfer_buflen)
229 io_u->error = -ev->res;
231 io_u->resid = io_u->xfer_buflen - ev->res;
239 unsigned id; /** kernel internal index number */
240 unsigned nr; /** number of io_events */
245 unsigned compat_features;
246 unsigned incompat_features;
247 unsigned header_length; /** size of aio_ring */
249 struct io_event events[0];
252 #define AIO_RING_MAGIC 0xa10a10a1
254 static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
255 struct io_event *events)
259 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
264 if (head == ring->tail) {
265 /* There are no more completions */
268 /* There is another completion to reap */
269 events[i] = ring->events[head];
270 atomic_store_release(&ring->head,
271 (head + 1) % ring->nr);
279 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
280 unsigned int max, const struct timespec *t)
282 struct libaio_data *ld = td->io_ops_data;
283 struct libaio_options *o = td->eo;
284 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
285 struct timespec __lt, *lt = NULL;
294 if (o->userspace_reap == 1
296 && ((struct aio_ring *)(ld->aio_ctx))->magic
298 r = user_io_getevents(ld->aio_ctx, max,
299 ld->aio_events + events);
301 r = io_getevents(ld->aio_ctx, actual_min,
302 max, ld->aio_events + events, lt);
306 else if ((min && r == 0) || r == -EAGAIN) {
307 fio_libaio_commit(td);
310 } else if (r != -EINTR)
312 } while (events < min);
314 return r < 0 ? r : events;
317 static enum fio_q_status fio_libaio_queue(struct thread_data *td,
320 struct libaio_data *ld = td->io_ops_data;
321 struct libaio_options *o = td->eo;
323 fio_ro_check(td, io_u);
325 if (ld->queued == td->o.iodepth)
329 * fsync is tricky, since it can fail and we need to do it
330 * serialized with other io. the reason is that linux doesn't
331 * support aio fsync yet. So return busy for the case where we
332 * have pending io, to let fio complete those first.
334 if (ddir_sync(io_u->ddir)) {
338 do_io_u_sync(td, io_u);
339 return FIO_Q_COMPLETED;
342 if (io_u->ddir == DDIR_TRIM) {
346 do_io_u_trim(td, io_u);
347 io_u_mark_submit(td, 1);
348 io_u_mark_complete(td, 1);
349 return FIO_Q_COMPLETED;
352 if (o->cmdprio.mode != CMDPRIO_MODE_NONE)
353 fio_libaio_cmdprio_prep(td, io_u);
355 ld->iocbs[ld->head] = &io_u->iocb;
356 ld->io_us[ld->head] = io_u;
357 ring_inc(ld, &ld->head, 1);
362 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
368 if (!fio_fill_issue_time(td))
371 fio_gettime(&now, NULL);
373 for (i = 0; i < nr; i++) {
374 struct io_u *io_u = io_us[i];
376 memcpy(&io_u->issue_time, &now, sizeof(now));
377 io_u_queued(td, io_u);
381 static int fio_libaio_commit(struct thread_data *td)
383 struct libaio_data *ld = td->io_ops_data;
387 int ret, wait_start = 0;
393 long nr = ld->queued;
395 nr = min((unsigned int) nr, ld->entries - ld->tail);
396 io_us = ld->io_us + ld->tail;
397 iocbs = ld->iocbs + ld->tail;
399 ret = io_submit(ld->aio_ctx, nr, iocbs);
401 fio_libaio_queued(td, io_us, ret);
402 io_u_mark_submit(td, ret);
405 ring_inc(ld, &ld->tail, ret);
408 } else if (ret == -EINTR || !ret) {
410 io_u_mark_submit(td, ret);
413 } else if (ret == -EAGAIN) {
415 * If we get EAGAIN, we should break out without
416 * error and let the upper layer reap some
417 * events for us. If we have no queued IO, we
418 * must loop here. If we loop for more than 30s,
419 * just error out, something must be buggy in the
427 fio_gettime(&ts, NULL);
429 } else if (mtime_since_now(&ts) > 30000) {
430 log_err("fio: aio appears to be stalled, giving up\n");
435 } else if (ret == -ENOMEM) {
437 * If we get -ENOMEM, reap events if we can. If
438 * we cannot, treat it as a fatal event since there's
439 * nothing we can do about it.
446 } while (ld->queued);
451 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
453 struct libaio_data *ld = td->io_ops_data;
455 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
458 static void fio_libaio_cleanup(struct thread_data *td)
460 struct libaio_data *ld = td->io_ops_data;
464 * Work-around to avoid huge RCU stalls at exit time. If we
465 * don't do this here, then it'll be torn down by exit_aio().
466 * But for that case we can parallellize the freeing, thus
467 * speeding it up a lot.
469 if (!(td->flags & TD_F_CHILD))
470 io_destroy(ld->aio_ctx);
471 free(ld->aio_events);
478 static int fio_libaio_post_init(struct thread_data *td)
480 struct libaio_data *ld = td->io_ops_data;
483 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
485 td_verror(td, -err, "io_queue_init");
492 static int fio_libaio_init(struct thread_data *td)
494 struct libaio_data *ld;
495 struct libaio_options *o = td->eo;
496 struct cmdprio *cmdprio = &o->cmdprio;
499 ld = calloc(1, sizeof(*ld));
501 ld->entries = td->o.iodepth;
502 ld->is_pow2 = is_power_of_2(ld->entries);
503 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
504 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
505 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
507 td->io_ops_data = ld;
509 ret = fio_cmdprio_init(td, cmdprio);
511 td_verror(td, EINVAL, "fio_libaio_init");
518 FIO_STATIC struct ioengine_ops ioengine = {
520 .version = FIO_IOOPS_VERSION,
521 .flags = FIO_ASYNCIO_SYNC_TRIM,
522 .init = fio_libaio_init,
523 .post_init = fio_libaio_post_init,
524 .prep = fio_libaio_prep,
525 .queue = fio_libaio_queue,
526 .commit = fio_libaio_commit,
527 .cancel = fio_libaio_cancel,
528 .getevents = fio_libaio_getevents,
529 .event = fio_libaio_event,
530 .cleanup = fio_libaio_cleanup,
531 .open_file = generic_open_file,
532 .close_file = generic_close_file,
533 .get_file_size = generic_get_file_size,
535 .option_struct_size = sizeof(struct libaio_options),
538 static void fio_init fio_libaio_register(void)
540 register_ioengine(&ioengine);
543 static void fio_exit fio_libaio_unregister(void)
545 unregister_ioengine(&ioengine);