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 static int fio_libaio_commit(struct thread_data *td);
23 struct io_event *aio_events;
27 struct io_u **io_u_index;
30 * Basic ring buffer. 'head' is incremented in _queue(), and
31 * 'tail' is incremented in _commit(). We keep 'queued' so
32 * that we know if the ring is full or empty, when
33 * 'head' == 'tail'. 'entries' is the ring size, and
34 * 'is_pow2' is just an optimization to use AND instead of
35 * modulus to get the remainder on ring increment.
44 struct libaio_options {
46 unsigned int userspace_reap;
50 static struct fio_option options[] = {
52 .name = "userspace_reap",
53 .lname = "Libaio userspace reaping",
54 .type = FIO_OPT_STR_SET,
55 .off1 = offsetof(struct libaio_options, userspace_reap),
56 .help = "Use alternative user-space reap implementation",
57 .category = FIO_OPT_C_ENGINE,
58 .group = FIO_OPT_G_LIBAIO,
65 static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
69 *val = (*val + add) & (ld->entries - 1);
71 *val = (*val + add) % ld->entries;
74 static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
76 struct fio_file *f = io_u->file;
77 struct iocb *iocb = &io_u->iocb;
79 if (io_u->ddir == DDIR_READ) {
80 io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
81 } else if (io_u->ddir == DDIR_WRITE) {
82 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
83 } else if (ddir_sync(io_u->ddir))
84 io_prep_fsync(iocb, f->fd);
89 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
91 struct libaio_data *ld = td->io_ops_data;
95 ev = ld->aio_events + event;
96 io_u = container_of(ev->obj, struct io_u, iocb);
98 if (ev->res != io_u->xfer_buflen) {
99 if (ev->res > io_u->xfer_buflen)
100 io_u->error = -ev->res;
102 io_u->resid = io_u->xfer_buflen - ev->res;
110 unsigned id; /** kernel internal index number */
111 unsigned nr; /** number of io_events */
116 unsigned compat_features;
117 unsigned incompat_features;
118 unsigned header_length; /** size of aio_ring */
120 struct io_event events[0];
123 #define AIO_RING_MAGIC 0xa10a10a1
125 static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
126 struct io_event *events)
130 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
135 if (head == ring->tail) {
136 /* There are no more completions */
139 /* There is another completion to reap */
140 events[i] = ring->events[head];
142 ring->head = (head + 1) % ring->nr;
150 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
151 unsigned int max, const struct timespec *t)
153 struct libaio_data *ld = td->io_ops_data;
154 struct libaio_options *o = td->eo;
155 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
156 struct timespec __lt, *lt = NULL;
165 if (o->userspace_reap == 1
167 && ((struct aio_ring *)(ld->aio_ctx))->magic
169 r = user_io_getevents(ld->aio_ctx, max,
170 ld->aio_events + events);
172 r = io_getevents(ld->aio_ctx, actual_min,
173 max, ld->aio_events + events, lt);
177 else if ((min && r == 0) || r == -EAGAIN) {
178 fio_libaio_commit(td);
181 } else if (r != -EINTR)
183 } while (events < min);
185 return r < 0 ? r : events;
188 static enum fio_q_status fio_libaio_queue(struct thread_data *td,
191 struct libaio_data *ld = td->io_ops_data;
193 fio_ro_check(td, io_u);
195 if (ld->queued == td->o.iodepth)
199 * fsync is tricky, since it can fail and we need to do it
200 * serialized with other io. the reason is that linux doesn't
201 * support aio fsync yet. So return busy for the case where we
202 * have pending io, to let fio complete those first.
204 if (ddir_sync(io_u->ddir)) {
208 do_io_u_sync(td, io_u);
209 return FIO_Q_COMPLETED;
212 if (io_u->ddir == DDIR_TRIM) {
216 do_io_u_trim(td, io_u);
217 io_u_mark_submit(td, 1);
218 io_u_mark_complete(td, 1);
219 return FIO_Q_COMPLETED;
222 ld->iocbs[ld->head] = &io_u->iocb;
223 ld->io_us[ld->head] = io_u;
224 ring_inc(ld, &ld->head, 1);
229 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
235 if (!fio_fill_issue_time(td))
238 fio_gettime(&now, NULL);
240 for (i = 0; i < nr; i++) {
241 struct io_u *io_u = io_us[i];
243 memcpy(&io_u->issue_time, &now, sizeof(now));
244 io_u_queued(td, io_u);
248 static int fio_libaio_commit(struct thread_data *td)
250 struct libaio_data *ld = td->io_ops_data;
254 int ret, wait_start = 0;
260 long nr = ld->queued;
262 nr = min((unsigned int) nr, ld->entries - ld->tail);
263 io_us = ld->io_us + ld->tail;
264 iocbs = ld->iocbs + ld->tail;
266 ret = io_submit(ld->aio_ctx, nr, iocbs);
268 fio_libaio_queued(td, io_us, ret);
269 io_u_mark_submit(td, ret);
272 ring_inc(ld, &ld->tail, ret);
275 } else if (ret == -EINTR || !ret) {
277 io_u_mark_submit(td, ret);
280 } else if (ret == -EAGAIN) {
282 * If we get EAGAIN, we should break out without
283 * error and let the upper layer reap some
284 * events for us. If we have no queued IO, we
285 * must loop here. If we loop for more than 30s,
286 * just error out, something must be buggy in the
294 fio_gettime(&ts, NULL);
296 } else if (mtime_since_now(&ts) > 30000) {
297 log_err("fio: aio appears to be stalled, giving up\n");
302 } else if (ret == -ENOMEM) {
304 * If we get -ENOMEM, reap events if we can. If
305 * we cannot, treat it as a fatal event since there's
306 * nothing we can do about it.
313 } while (ld->queued);
318 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
320 struct libaio_data *ld = td->io_ops_data;
322 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
325 static void fio_libaio_cleanup(struct thread_data *td)
327 struct libaio_data *ld = td->io_ops_data;
331 * Work-around to avoid huge RCU stalls at exit time. If we
332 * don't do this here, then it'll be torn down by exit_aio().
333 * But for that case we can parallellize the freeing, thus
334 * speeding it up a lot.
336 if (!(td->flags & TD_F_CHILD))
337 io_destroy(ld->aio_ctx);
338 free(ld->aio_events);
345 static int fio_libaio_post_init(struct thread_data *td)
347 struct libaio_data *ld = td->io_ops_data;
350 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
352 td_verror(td, -err, "io_queue_init");
359 static int fio_libaio_init(struct thread_data *td)
361 struct libaio_data *ld;
363 ld = calloc(1, sizeof(*ld));
365 ld->entries = td->o.iodepth;
366 ld->is_pow2 = is_power_of_2(ld->entries);
367 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
368 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
369 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
371 td->io_ops_data = ld;
375 static struct ioengine_ops ioengine = {
377 .version = FIO_IOOPS_VERSION,
378 .flags = FIO_ASYNCIO_SYNC_TRIM,
379 .init = fio_libaio_init,
380 .post_init = fio_libaio_post_init,
381 .prep = fio_libaio_prep,
382 .queue = fio_libaio_queue,
383 .commit = fio_libaio_commit,
384 .cancel = fio_libaio_cancel,
385 .getevents = fio_libaio_getevents,
386 .event = fio_libaio_event,
387 .cleanup = fio_libaio_cleanup,
388 .open_file = generic_open_file,
389 .close_file = generic_close_file,
390 .get_file_size = generic_get_file_size,
392 .option_struct_size = sizeof(struct libaio_options),
395 static void fio_init fio_libaio_register(void)
397 register_ioengine(&ioengine);
400 static void fio_exit fio_libaio_unregister(void)
402 unregister_ioengine(&ioengine);