4 * IO engine using the Linux native aio interface.
13 #include "../lib/pow2.h"
14 #include "../optgroup.h"
15 #include "../lib/memalign.h"
17 #ifndef IOCB_FLAG_HIPRI
18 #define IOCB_FLAG_HIPRI (1 << 2)
21 #ifndef IOCTX_FLAG_USERIOCB
22 #define IOCTX_FLAG_USERIOCB (1 << 0)
24 #ifndef IOCTX_FLAG_IOPOLL
25 #define IOCTX_FLAG_IOPOLL (1 << 1)
27 #ifndef IOCTX_FLAG_FIXEDBUFS
28 #define IOCTX_FLAG_FIXEDBUFS (1 << 2)
31 static int fio_libaio_commit(struct thread_data *td);
35 struct io_event *aio_events;
39 struct iocb *user_iocbs;
40 struct io_u **io_u_index;
43 * Basic ring buffer. 'head' is incremented in _queue(), and
44 * 'tail' is incremented in _commit(). We keep 'queued' so
45 * that we know if the ring is full or empty, when
46 * 'head' == 'tail'. 'entries' is the ring size, and
47 * 'is_pow2' is just an optimization to use AND instead of
48 * modulus to get the remainder on ring increment.
57 struct libaio_options {
59 unsigned int userspace_reap;
61 unsigned int useriocb;
62 unsigned int fixedbufs;
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 = "High Priority",
78 .type = FIO_OPT_STR_SET,
79 .off1 = offsetof(struct libaio_options, hipri),
80 .help = "Use polled IO completions",
81 .category = FIO_OPT_C_ENGINE,
82 .group = FIO_OPT_G_LIBAIO,
86 .lname = "User IOCBs",
87 .type = FIO_OPT_STR_SET,
88 .off1 = offsetof(struct libaio_options, useriocb),
89 .help = "Use user mapped IOCBs",
90 .category = FIO_OPT_C_ENGINE,
91 .group = FIO_OPT_G_LIBAIO,
95 .lname = "Fixed (pre-mapped) IO buffers",
96 .type = FIO_OPT_STR_SET,
97 .off1 = offsetof(struct libaio_options, fixedbufs),
98 .help = "Pre map IO buffers",
99 .category = FIO_OPT_C_ENGINE,
100 .group = FIO_OPT_G_LIBAIO,
107 static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
111 *val = (*val + add) & (ld->entries - 1);
113 *val = (*val + add) % ld->entries;
116 static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
118 struct libaio_data *ld = td->io_ops_data;
119 struct fio_file *f = io_u->file;
120 struct libaio_options *o = td->eo;
124 iocb = &ld->user_iocbs[io_u->index];
128 if (io_u->ddir == DDIR_READ) {
130 iocb->aio_fildes = f->fd;
131 iocb->aio_lio_opcode = IO_CMD_PREAD;
132 iocb->u.c.offset = io_u->offset;
134 io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
135 } else if (io_u->ddir == DDIR_WRITE) {
137 iocb->aio_fildes = f->fd;
138 iocb->aio_lio_opcode = IO_CMD_PWRITE;
139 iocb->u.c.offset = io_u->offset;
141 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
142 } else if (ddir_sync(io_u->ddir))
143 io_prep_fsync(iocb, f->fd);
148 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
150 struct libaio_data *ld = td->io_ops_data;
151 struct libaio_options *o = td->eo;
155 ev = ld->aio_events + event;
157 int index = (int) (uintptr_t) ev->obj;
158 io_u = ld->io_u_index[index];
160 io_u = container_of(ev->obj, struct io_u, iocb);
162 if (ev->res != io_u->xfer_buflen) {
163 if (ev->res > io_u->xfer_buflen)
164 io_u->error = -ev->res;
166 io_u->resid = io_u->xfer_buflen - ev->res;
174 unsigned id; /** kernel internal index number */
175 unsigned nr; /** number of io_events */
180 unsigned compat_features;
181 unsigned incompat_features;
182 unsigned header_length; /** size of aio_ring */
184 struct io_event events[0];
187 #define AIO_RING_MAGIC 0xa10a10a1
189 static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
190 struct io_event *events)
194 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
199 if (head == ring->tail) {
200 /* There are no more completions */
203 /* There is another completion to reap */
204 events[i] = ring->events[head];
206 ring->head = (head + 1) % ring->nr;
214 static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
215 unsigned int max, const struct timespec *t)
217 struct libaio_data *ld = td->io_ops_data;
218 struct libaio_options *o = td->eo;
219 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
220 struct timespec __lt, *lt = NULL;
229 if (o->userspace_reap == 1
231 && ((struct aio_ring *)(ld->aio_ctx))->magic
233 r = user_io_getevents(ld->aio_ctx, max,
234 ld->aio_events + events);
236 r = io_getevents(ld->aio_ctx, actual_min,
237 max, ld->aio_events + events, lt);
241 else if ((min && r == 0) || r == -EAGAIN) {
242 fio_libaio_commit(td);
245 } else if (r != -EINTR)
247 } while (events < min);
249 return r < 0 ? r : events;
252 static enum fio_q_status fio_libaio_queue(struct thread_data *td,
255 struct libaio_data *ld = td->io_ops_data;
256 struct libaio_options *o = td->eo;
258 fio_ro_check(td, io_u);
260 if (ld->queued == td->o.iodepth)
264 * fsync is tricky, since it can fail and we need to do it
265 * serialized with other io. the reason is that linux doesn't
266 * support aio fsync yet. So return busy for the case where we
267 * have pending io, to let fio complete those first.
269 if (ddir_sync(io_u->ddir)) {
273 do_io_u_sync(td, io_u);
274 return FIO_Q_COMPLETED;
277 if (io_u->ddir == DDIR_TRIM) {
281 do_io_u_trim(td, io_u);
282 io_u_mark_submit(td, 1);
283 io_u_mark_complete(td, 1);
284 return FIO_Q_COMPLETED;
288 ld->iocbs[ld->head] = (struct iocb *) (uintptr_t) io_u->index;
290 ld->iocbs[ld->head] = &io_u->iocb;
292 ld->io_us[ld->head] = io_u;
293 ring_inc(ld, &ld->head, 1);
298 static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
304 if (!fio_fill_issue_time(td))
307 fio_gettime(&now, NULL);
309 for (i = 0; i < nr; i++) {
310 struct io_u *io_u = io_us[i];
312 memcpy(&io_u->issue_time, &now, sizeof(now));
313 io_u_queued(td, io_u);
317 static int fio_libaio_commit(struct thread_data *td)
319 struct libaio_data *ld = td->io_ops_data;
323 int ret, wait_start = 0;
329 long nr = ld->queued;
331 nr = min((unsigned int) nr, ld->entries - ld->tail);
332 io_us = ld->io_us + ld->tail;
333 iocbs = ld->iocbs + ld->tail;
335 ret = io_submit(ld->aio_ctx, nr, iocbs);
337 fio_libaio_queued(td, io_us, ret);
338 io_u_mark_submit(td, ret);
341 ring_inc(ld, &ld->tail, ret);
344 } else if (ret == -EINTR || !ret) {
346 io_u_mark_submit(td, ret);
349 } else if (ret == -EAGAIN) {
351 * If we get EAGAIN, we should break out without
352 * error and let the upper layer reap some
353 * events for us. If we have no queued IO, we
354 * must loop here. If we loop for more than 30s,
355 * just error out, something must be buggy in the
363 fio_gettime(&ts, NULL);
365 } else if (mtime_since_now(&ts) > 30000) {
366 log_err("fio: aio appears to be stalled, giving up\n");
371 } else if (ret == -ENOMEM) {
373 * If we get -ENOMEM, reap events if we can. If
374 * we cannot, treat it as a fatal event since there's
375 * nothing we can do about it.
382 } while (ld->queued);
387 static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
389 struct libaio_data *ld = td->io_ops_data;
391 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
394 static void fio_libaio_cleanup(struct thread_data *td)
396 struct libaio_data *ld = td->io_ops_data;
400 * Work-around to avoid huge RCU stalls at exit time. If we
401 * don't do this here, then it'll be torn down by exit_aio().
402 * But for that case we can parallellize the freeing, thus
403 * speeding it up a lot.
405 if (!(td->flags & TD_F_CHILD))
406 io_destroy(ld->aio_ctx);
407 free(ld->aio_events);
410 if (ld->user_iocbs) {
411 size_t size = td->o.iodepth * sizeof(struct iocb);
412 fio_memfree(ld->user_iocbs, size, false);
418 static int fio_libaio_old_queue_init(struct libaio_data *ld, unsigned int depth,
419 bool hipri, bool useriocb, bool fixedbufs)
422 log_err("fio: polled aio not available on your platform\n");
426 log_err("fio: user mapped iocbs not available on your platform\n");
430 log_err("fio: fixed buffers not available on your platform\n");
434 return io_queue_init(depth, &ld->aio_ctx);
437 static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth,
438 bool hipri, bool useriocb, bool fixedbufs)
440 #ifdef __NR_sys_io_setup2
444 flags |= IOCTX_FLAG_IOPOLL;
446 flags |= IOCTX_FLAG_USERIOCB;
448 flags |= IOCTX_FLAG_FIXEDBUFS;
450 ret = syscall(__NR_sys_io_setup2, depth, flags, ld->user_iocbs,
454 /* fall through to old syscall */
456 return fio_libaio_old_queue_init(ld, depth, hipri, useriocb, fixedbufs);
459 static int fio_libaio_post_init(struct thread_data *td)
461 struct libaio_data *ld = td->io_ops_data;
462 struct libaio_options *o = td->eo;
470 for (i = 0; i < td->o.iodepth; i++) {
471 io_u = ld->io_u_index[i];
472 iocb = &ld->user_iocbs[i];
473 iocb->u.c.buf = io_u->buf;
474 iocb->u.c.nbytes = td_max_bs(td);
478 iocb->u.c.flags |= IOCB_FLAG_HIPRI;
482 err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri, o->useriocb,
485 td_verror(td, -err, "io_queue_init");
492 static int fio_libaio_init(struct thread_data *td)
494 struct libaio_options *o = td->eo;
495 struct libaio_data *ld;
497 ld = calloc(1, sizeof(*ld));
502 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
503 size = td->o.iodepth * sizeof(struct iocb);
504 ld->user_iocbs = fio_memalign(page_size, size, false);
505 memset(ld->user_iocbs, 0, size);
508 ld->entries = td->o.iodepth;
509 ld->is_pow2 = is_power_of_2(ld->entries);
510 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
511 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
512 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
514 td->io_ops_data = ld;
518 static int fio_libaio_io_u_init(struct thread_data *td, struct io_u *io_u)
520 struct libaio_options *o = td->eo;
523 struct libaio_data *ld = td->io_ops_data;
525 ld->io_u_index[io_u->index] = io_u;
531 static struct ioengine_ops ioengine = {
533 .version = FIO_IOOPS_VERSION,
534 .init = fio_libaio_init,
535 .post_init = fio_libaio_post_init,
536 .io_u_init = fio_libaio_io_u_init,
537 .prep = fio_libaio_prep,
538 .queue = fio_libaio_queue,
539 .commit = fio_libaio_commit,
540 .cancel = fio_libaio_cancel,
541 .getevents = fio_libaio_getevents,
542 .event = fio_libaio_event,
543 .cleanup = fio_libaio_cleanup,
544 .open_file = generic_open_file,
545 .close_file = generic_close_file,
546 .get_file_size = generic_get_file_size,
548 .option_struct_size = sizeof(struct libaio_options),
551 static void fio_init fio_libaio_register(void)
553 register_ioengine(&ioengine);
556 static void fio_exit fio_libaio_unregister(void)
558 unregister_ioengine(&ioengine);