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.
55 struct cmdprio cmdprio;
58 struct libaio_options {
59 struct thread_data *td;
60 unsigned int userspace_reap;
61 struct cmdprio_options cmdprio_options;
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,
77 .lname = "RWF_NOWAIT",
79 .off1 = offsetof(struct libaio_options, nowait),
80 .help = "Set RWF_NOWAIT for reads/writes",
81 .category = FIO_OPT_C_ENGINE,
82 .group = FIO_OPT_G_LIBAIO,
84 CMDPRIO_OPTIONS(struct libaio_options, FIO_OPT_G_LIBAIO),
90 static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
94 *val = (*val + add) & (ld->entries - 1);
96 *val = (*val + add) % ld->entries;
99 static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u)
101 struct libaio_options *o = td->eo;
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);
108 iocb->aio_rw_flags |= RWF_NOWAIT;
109 } else if (io_u->ddir == DDIR_WRITE) {
110 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
112 iocb->aio_rw_flags |= RWF_NOWAIT;
113 } else if (ddir_sync(io_u->ddir))
114 io_prep_fsync(iocb, f->fd);
119 static inline void fio_libaio_cmdprio_prep(struct thread_data *td,
122 struct libaio_data *ld = td->io_ops_data;
123 struct cmdprio *cmdprio = &ld->cmdprio;
125 if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) {
126 io_u->iocb.aio_reqprio = io_u->ioprio;
127 io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
131 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
133 struct libaio_data *ld = td->io_ops_data;
137 ev = ld->aio_events + event;
138 io_u = container_of(ev->obj, struct io_u, iocb);
140 if (ev->res != io_u->xfer_buflen) {
141 if (ev->res > io_u->xfer_buflen)
142 io_u->error = -ev->res;
144 io_u->resid = io_u->xfer_buflen - ev->res;
152 unsigned id; /** kernel internal index number */
153 unsigned nr; /** number of io_events */
158 unsigned compat_features;
159 unsigned incompat_features;
160 unsigned header_length; /** size of aio_ring */
162 struct io_event events[0];
165 #define AIO_RING_MAGIC 0xa10a10a1
167 static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
168 struct io_event *events)
172 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
177 if (head == ring->tail) {
178 /* There are no more completions */
181 /* There is another completion to reap */
182 events[i] = ring->events[head];
183 atomic_store_release(&ring->head,
184 (head + 1) % ring->nr);
192 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
193 unsigned int max, const struct timespec *t)
195 struct libaio_data *ld = td->io_ops_data;
196 struct libaio_options *o = td->eo;
197 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
198 struct timespec __lt, *lt = NULL;
207 if (o->userspace_reap == 1
209 && ((struct aio_ring *)(ld->aio_ctx))->magic
211 r = user_io_getevents(ld->aio_ctx, max - events,
212 ld->aio_events + events);
214 r = io_getevents(ld->aio_ctx, actual_min,
215 max - events, ld->aio_events + events, lt);
219 actual_min -= min((unsigned int)events, actual_min);
221 else if ((min && r == 0) || r == -EAGAIN) {
222 fio_libaio_commit(td);
225 } else if (r != -EINTR)
227 } while (events < min);
229 return r < 0 ? r : events;
232 static enum fio_q_status fio_libaio_queue(struct thread_data *td,
235 struct libaio_data *ld = td->io_ops_data;
237 fio_ro_check(td, io_u);
239 if (ld->queued == td->o.iodepth)
243 * fsync is tricky, since it can fail and we need to do it
244 * serialized with other io. the reason is that linux doesn't
245 * support aio fsync yet. So return busy for the case where we
246 * have pending io, to let fio complete those first.
248 if (ddir_sync(io_u->ddir)) {
252 do_io_u_sync(td, io_u);
253 return FIO_Q_COMPLETED;
256 if (io_u->ddir == DDIR_TRIM) {
260 do_io_u_trim(td, io_u);
261 io_u_mark_submit(td, 1);
262 io_u_mark_complete(td, 1);
263 return FIO_Q_COMPLETED;
266 if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
267 fio_libaio_cmdprio_prep(td, io_u);
269 ld->iocbs[ld->head] = &io_u->iocb;
270 ld->io_us[ld->head] = io_u;
271 ring_inc(ld, &ld->head, 1);
276 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
282 if (!fio_fill_issue_time(td))
285 fio_gettime(&now, NULL);
287 for (i = 0; i < nr; i++) {
288 struct io_u *io_u = io_us[i];
290 memcpy(&io_u->issue_time, &now, sizeof(now));
291 io_u_queued(td, io_u);
295 * only used for iolog
297 if (td->o.read_iolog_file)
298 memcpy(&td->last_issue, &now, sizeof(now));
301 static int fio_libaio_commit(struct thread_data *td)
303 struct libaio_data *ld = td->io_ops_data;
307 int ret, wait_start = 0;
313 long nr = ld->queued;
315 nr = min((unsigned int) nr, ld->entries - ld->tail);
316 io_us = ld->io_us + ld->tail;
317 iocbs = ld->iocbs + ld->tail;
319 ret = io_submit(ld->aio_ctx, nr, iocbs);
321 fio_libaio_queued(td, io_us, ret);
322 io_u_mark_submit(td, ret);
325 ring_inc(ld, &ld->tail, ret);
328 } else if (ret == -EINTR || !ret) {
330 io_u_mark_submit(td, ret);
333 } else if (ret == -EAGAIN) {
335 * If we get EAGAIN, we should break out without
336 * error and let the upper layer reap some
337 * events for us. If we have no queued IO, we
338 * must loop here. If we loop for more than 30s,
339 * just error out, something must be buggy in the
347 fio_gettime(&ts, NULL);
349 } else if (mtime_since_now(&ts) > 30000) {
350 log_err("fio: aio appears to be stalled, giving up\n");
355 } else if (ret == -ENOMEM) {
357 * If we get -ENOMEM, reap events if we can. If
358 * we cannot, treat it as a fatal event since there's
359 * nothing we can do about it.
366 } while (ld->queued);
371 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
373 struct libaio_data *ld = td->io_ops_data;
375 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
378 static void fio_libaio_cleanup(struct thread_data *td)
380 struct libaio_data *ld = td->io_ops_data;
384 * Work-around to avoid huge RCU stalls at exit time. If we
385 * don't do this here, then it'll be torn down by exit_aio().
386 * But for that case we can parallellize the freeing, thus
387 * speeding it up a lot.
389 if (!(td->flags & TD_F_CHILD))
390 io_destroy(ld->aio_ctx);
392 fio_cmdprio_cleanup(&ld->cmdprio);
393 free(ld->aio_events);
400 static int fio_libaio_post_init(struct thread_data *td)
402 struct libaio_data *ld = td->io_ops_data;
405 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
407 td_verror(td, -err, "io_queue_init");
414 static int fio_libaio_init(struct thread_data *td)
416 struct libaio_data *ld;
417 struct libaio_options *o = td->eo;
420 ld = calloc(1, sizeof(*ld));
422 ld->entries = td->o.iodepth;
423 ld->is_pow2 = is_power_of_2(ld->entries);
424 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
425 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
426 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
428 td->io_ops_data = ld;
430 ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
432 td_verror(td, EINVAL, "fio_libaio_init");
439 FIO_STATIC struct ioengine_ops ioengine = {
441 .version = FIO_IOOPS_VERSION,
442 .flags = FIO_ASYNCIO_SYNC_TRIM |
443 FIO_ASYNCIO_SETS_ISSUE_TIME,
444 .init = fio_libaio_init,
445 .post_init = fio_libaio_post_init,
446 .prep = fio_libaio_prep,
447 .queue = fio_libaio_queue,
448 .commit = fio_libaio_commit,
449 .cancel = fio_libaio_cancel,
450 .getevents = fio_libaio_getevents,
451 .event = fio_libaio_event,
452 .cleanup = fio_libaio_cleanup,
453 .open_file = generic_open_file,
454 .close_file = generic_close_file,
455 .get_file_size = generic_get_file_size,
457 .option_struct_size = sizeof(struct libaio_options),
460 static void fio_init fio_libaio_register(void)
462 register_ioengine(&ioengine);
465 static void fio_exit fio_libaio_unregister(void)
467 unregister_ioengine(&ioengine);