4 * IO engine using Ceph's librbd to test RADOS Block Devices.
8 #include <rbd/librbd.h>
11 #include "../optgroup.h"
13 #ifdef CONFIG_RBD_POLL
16 #include <sys/eventfd.h>
21 rbd_completion_t completion;
30 struct io_u **aio_events;
31 struct io_u **sort_events;
32 int fd; /* add for poll */
45 static struct fio_option options[] = {
47 .name = "clustername",
48 .lname = "ceph cluster name",
49 .type = FIO_OPT_STR_STORE,
50 .help = "Cluster name for ceph",
51 .off1 = offsetof(struct rbd_options, cluster_name),
52 .category = FIO_OPT_C_ENGINE,
53 .group = FIO_OPT_G_RBD,
57 .lname = "rbd engine rbdname",
58 .type = FIO_OPT_STR_STORE,
59 .help = "RBD name for RBD engine",
60 .off1 = offsetof(struct rbd_options, rbd_name),
61 .category = FIO_OPT_C_ENGINE,
62 .group = FIO_OPT_G_RBD,
66 .lname = "rbd engine pool",
67 .type = FIO_OPT_STR_STORE,
68 .help = "Name of the pool hosting the RBD for the RBD engine",
69 .off1 = offsetof(struct rbd_options, pool_name),
70 .category = FIO_OPT_C_ENGINE,
71 .group = FIO_OPT_G_RBD,
75 .lname = "rbd engine clientname",
76 .type = FIO_OPT_STR_STORE,
77 .help = "Name of the ceph client to access the RBD for the RBD engine",
78 .off1 = offsetof(struct rbd_options, client_name),
79 .category = FIO_OPT_C_ENGINE,
80 .group = FIO_OPT_G_RBD,
86 .help = "Busy poll for completions instead of sleeping",
87 .off1 = offsetof(struct rbd_options, busy_poll),
89 .category = FIO_OPT_C_ENGINE,
90 .group = FIO_OPT_G_RBD,
97 static int _fio_setup_rbd_data(struct thread_data *td,
98 struct rbd_data **rbd_data_ptr)
100 struct rbd_data *rbd;
105 rbd = calloc(1, sizeof(struct rbd_data));
109 rbd->connected = false;
111 /* add for poll, init fd: -1 */
114 rbd->aio_events = calloc(td->o.iodepth, sizeof(struct io_u *));
115 if (!rbd->aio_events)
118 rbd->sort_events = calloc(td->o.iodepth, sizeof(struct io_u *));
119 if (!rbd->sort_events)
128 free(rbd->aio_events);
129 if (rbd->sort_events)
130 free(rbd->sort_events);
137 #ifdef CONFIG_RBD_POLL
138 static bool _fio_rbd_setup_poll(struct rbd_data *rbd)
142 /* add for rbd poll */
143 rbd->fd = eventfd(0, EFD_SEMAPHORE);
145 log_err("eventfd failed.\n");
149 r = rbd_set_image_notification(rbd->image, rbd->fd, EVENT_TYPE_EVENTFD);
151 log_err("rbd_set_image_notification failed.\n");
160 static bool _fio_rbd_setup_poll(struct rbd_data *rbd)
166 static int _fio_rbd_connect(struct thread_data *td)
168 struct rbd_data *rbd = td->io_ops_data;
169 struct rbd_options *o = td->eo;
172 if (o->cluster_name) {
173 char *client_name = NULL;
176 * If we specify cluser name, the rados_create2
177 * will not assume 'client.'. name is considered
178 * as a full type.id namestr
180 if (o->client_name) {
181 if (!index(o->client_name, '.')) {
182 client_name = calloc(1, strlen("client.") +
183 strlen(o->client_name) + 1);
184 strcat(client_name, "client.");
185 strcat(client_name, o->client_name);
187 client_name = o->client_name;
191 r = rados_create2(&rbd->cluster, o->cluster_name,
194 if (client_name && !index(o->client_name, '.'))
197 r = rados_create(&rbd->cluster, o->client_name);
200 log_err("rados_create failed.\n");
204 r = rados_conf_read_file(rbd->cluster, NULL);
206 log_err("rados_conf_read_file failed.\n");
210 r = rados_connect(rbd->cluster);
212 log_err("rados_connect failed.\n");
213 goto failed_shutdown;
216 r = rados_ioctx_create(rbd->cluster, o->pool_name, &rbd->io_ctx);
218 log_err("rados_ioctx_create failed.\n");
219 goto failed_shutdown;
222 r = rbd_open(rbd->io_ctx, o->rbd_name, &rbd->image, NULL /*snap */ );
224 log_err("rbd_open failed.\n");
228 if (!_fio_rbd_setup_poll(rbd))
234 rbd_close(rbd->image);
237 rados_ioctx_destroy(rbd->io_ctx);
240 rados_shutdown(rbd->cluster);
246 static void _fio_rbd_disconnect(struct rbd_data *rbd)
257 /* shutdown everything */
259 rbd_close(rbd->image);
264 rados_ioctx_destroy(rbd->io_ctx);
269 rados_shutdown(rbd->cluster);
274 static void _fio_rbd_finish_aiocb(rbd_completion_t comp, void *data)
276 struct fio_rbd_iou *fri = data;
277 struct io_u *io_u = fri->io_u;
281 * Looks like return value is 0 for success, or < 0 for
282 * a specific error. So we have to assume that it can't do
283 * partial completions.
285 ret = rbd_aio_get_return_value(fri->completion);
288 io_u->resid = io_u->xfer_buflen;
292 fri->io_complete = 1;
295 static struct io_u *fio_rbd_event(struct thread_data *td, int event)
297 struct rbd_data *rbd = td->io_ops_data;
299 return rbd->aio_events[event];
302 static inline int fri_check_complete(struct rbd_data *rbd, struct io_u *io_u,
303 unsigned int *events)
305 struct fio_rbd_iou *fri = io_u->engine_data;
307 if (fri->io_complete) {
309 rbd->aio_events[*events] = io_u;
312 rbd_aio_release(fri->completion);
319 #ifndef CONFIG_RBD_POLL
320 static inline int rbd_io_u_seen(struct io_u *io_u)
322 struct fio_rbd_iou *fri = io_u->engine_data;
328 static void rbd_io_u_wait_complete(struct io_u *io_u)
330 struct fio_rbd_iou *fri = io_u->engine_data;
332 rbd_aio_wait_for_complete(fri->completion);
335 static int rbd_io_u_cmp(const void *p1, const void *p2)
337 const struct io_u **a = (const struct io_u **) p1;
338 const struct io_u **b = (const struct io_u **) p2;
341 at = utime_since_now(&(*a)->start_time);
342 bt = utime_since_now(&(*b)->start_time);
352 static int rbd_iter_events(struct thread_data *td, unsigned int *events,
353 unsigned int min_evts, int wait)
355 struct rbd_data *rbd = td->io_ops_data;
356 unsigned int this_events = 0;
360 #ifdef CONFIG_RBD_POLL
363 struct fio_rbd_iou *fri = NULL;
364 rbd_completion_t comps[min_evts];
372 ret = poll(&pfd, 1, wait ? -1 : 0);
375 if (!(pfd.revents & POLLIN))
378 event_num = rbd_poll_io_events(rbd->image, comps, min_evts);
380 for (i = 0; i < event_num; i++) {
381 fri = rbd_aio_get_arg(comps[i]);
384 /* best effort to decrement the semaphore */
385 ret = read(rbd->fd, &counter, sizeof(counter));
387 log_err("rbd_iter_events failed to decrement semaphore.\n");
389 completed = fri_check_complete(rbd, io_u, events);
395 io_u_qiter(&td->io_u_all, io_u, i) {
396 if (!(io_u->flags & IO_U_F_FLIGHT))
398 if (rbd_io_u_seen(io_u))
401 if (fri_check_complete(rbd, io_u, events))
404 rbd->sort_events[sidx++] = io_u;
412 * Sort events, oldest issue first, then wait on as many as we
413 * need in order of age. If we have enough events, stop waiting,
414 * and just check if any of the older ones are done.
417 qsort(rbd->sort_events, sidx, sizeof(struct io_u *), rbd_io_u_cmp);
419 for (i = 0; i < sidx; i++) {
420 io_u = rbd->sort_events[i];
422 if (fri_check_complete(rbd, io_u, events)) {
428 * Stop waiting when we have enough, but continue checking
429 * all pending IOs if they are complete.
431 if (*events >= min_evts)
434 rbd_io_u_wait_complete(io_u);
436 if (fri_check_complete(rbd, io_u, events))
443 static int fio_rbd_getevents(struct thread_data *td, unsigned int min,
444 unsigned int max, const struct timespec *t)
446 unsigned int this_events, events = 0;
447 struct rbd_options *o = td->eo;
451 this_events = rbd_iter_events(td, &events, min, wait);
467 static enum fio_q_status fio_rbd_queue(struct thread_data *td,
470 struct rbd_data *rbd = td->io_ops_data;
471 struct fio_rbd_iou *fri = io_u->engine_data;
474 fio_ro_check(td, io_u);
477 fri->io_complete = 0;
479 r = rbd_aio_create_completion(fri, _fio_rbd_finish_aiocb,
482 log_err("rbd_aio_create_completion failed.\n");
486 if (io_u->ddir == DDIR_WRITE) {
487 r = rbd_aio_write(rbd->image, io_u->offset, io_u->xfer_buflen,
488 io_u->xfer_buf, fri->completion);
490 log_err("rbd_aio_write failed.\n");
494 } else if (io_u->ddir == DDIR_READ) {
495 r = rbd_aio_read(rbd->image, io_u->offset, io_u->xfer_buflen,
496 io_u->xfer_buf, fri->completion);
499 log_err("rbd_aio_read failed.\n");
502 } else if (io_u->ddir == DDIR_TRIM) {
503 r = rbd_aio_discard(rbd->image, io_u->offset,
504 io_u->xfer_buflen, fri->completion);
506 log_err("rbd_aio_discard failed.\n");
509 } else if (io_u->ddir == DDIR_SYNC) {
510 r = rbd_aio_flush(rbd->image, fri->completion);
512 log_err("rbd_flush failed.\n");
516 dprint(FD_IO, "%s: Warning: unhandled ddir: %d\n", __func__,
524 rbd_aio_release(fri->completion);
527 td_verror(td, io_u->error, "xfer");
528 return FIO_Q_COMPLETED;
531 static int fio_rbd_init(struct thread_data *td)
534 struct rbd_data *rbd = td->io_ops_data;
539 r = _fio_rbd_connect(td);
541 log_err("fio_rbd_connect failed, return code: %d .\n", r);
551 static void fio_rbd_cleanup(struct thread_data *td)
553 struct rbd_data *rbd = td->io_ops_data;
556 _fio_rbd_disconnect(rbd);
557 free(rbd->aio_events);
558 free(rbd->sort_events);
563 static int fio_rbd_setup(struct thread_data *td)
565 rbd_image_info_t info;
567 struct rbd_data *rbd = NULL;
570 /* allocate engine specific structure to deal with librbd. */
571 r = _fio_setup_rbd_data(td, &rbd);
573 log_err("fio_setup_rbd_data failed.\n");
576 td->io_ops_data = rbd;
578 /* librbd does not allow us to run first in the main thread and later
579 * in a fork child. It needs to be the same process context all the
582 td->o.use_thread = 1;
584 /* connect in the main thread to determine to determine
585 * the size of the given RADOS block device. And disconnect
588 r = _fio_rbd_connect(td);
590 log_err("fio_rbd_connect failed.\n");
593 rbd->connected = true;
595 /* get size of the RADOS block device */
596 r = rbd_stat(rbd->image, &info, sizeof(info));
598 log_err("rbd_status failed.\n");
600 } else if (info.size == 0) {
601 log_err("image size should be larger than zero.\n");
606 dprint(FD_IO, "rbd-engine: image size: %" PRIu64 "\n", info.size);
608 /* taken from "net" engine. Pretend we deal with files,
609 * even if we do not have any ideas about files.
610 * The size of the RBD is set instead of a artificial file.
612 if (!td->files_index) {
613 add_file(td, td->o.filename ? : "rbd", 0, 0);
614 td->o.nr_files = td->o.nr_files ? : 1;
618 f->real_file_size = info.size;
627 static int fio_rbd_open(struct thread_data *td, struct fio_file *f)
632 static int fio_rbd_invalidate(struct thread_data *td, struct fio_file *f)
634 #if defined(CONFIG_RBD_INVAL)
635 struct rbd_data *rbd = td->io_ops_data;
637 return rbd_invalidate_cache(rbd->image);
643 static void fio_rbd_io_u_free(struct thread_data *td, struct io_u *io_u)
645 struct fio_rbd_iou *fri = io_u->engine_data;
648 io_u->engine_data = NULL;
653 static int fio_rbd_io_u_init(struct thread_data *td, struct io_u *io_u)
655 struct fio_rbd_iou *fri;
657 fri = calloc(1, sizeof(*fri));
659 io_u->engine_data = fri;
663 static struct ioengine_ops ioengine = {
665 .version = FIO_IOOPS_VERSION,
666 .setup = fio_rbd_setup,
667 .init = fio_rbd_init,
668 .queue = fio_rbd_queue,
669 .getevents = fio_rbd_getevents,
670 .event = fio_rbd_event,
671 .cleanup = fio_rbd_cleanup,
672 .open_file = fio_rbd_open,
673 .invalidate = fio_rbd_invalidate,
675 .io_u_init = fio_rbd_io_u_init,
676 .io_u_free = fio_rbd_io_u_free,
677 .option_struct_size = sizeof(struct rbd_options),
680 static void fio_init fio_rbd_register(void)
682 register_ioengine(&ioengine);
685 static void fio_exit fio_rbd_unregister(void)
687 unregister_ioengine(&ioengine);