4 * IO engine using Ceph's librbd to test RADOS Block Devices.
8 #include <rbd/librbd.h>
11 #include "../optgroup.h"
12 #ifdef CONFIG_RBD_BLKIN
16 #ifdef CONFIG_RBD_POLL
19 #include <sys/eventfd.h>
24 rbd_completion_t completion;
27 #ifdef CONFIG_RBD_BLKIN
28 struct blkin_trace_info info;
36 struct io_u **aio_events;
37 struct io_u **sort_events;
38 int fd; /* add for poll */
50 static struct fio_option options[] = {
52 .name = "clustername",
53 .lname = "ceph cluster name",
54 .type = FIO_OPT_STR_STORE,
55 .help = "Cluster name for ceph",
56 .off1 = offsetof(struct rbd_options, cluster_name),
57 .category = FIO_OPT_C_ENGINE,
58 .group = FIO_OPT_G_RBD,
62 .lname = "rbd engine rbdname",
63 .type = FIO_OPT_STR_STORE,
64 .help = "RBD name for RBD engine",
65 .off1 = offsetof(struct rbd_options, rbd_name),
66 .category = FIO_OPT_C_ENGINE,
67 .group = FIO_OPT_G_RBD,
71 .lname = "rbd engine pool",
72 .type = FIO_OPT_STR_STORE,
73 .help = "Name of the pool hosting the RBD for the RBD engine",
74 .off1 = offsetof(struct rbd_options, pool_name),
75 .category = FIO_OPT_C_ENGINE,
76 .group = FIO_OPT_G_RBD,
80 .lname = "rbd engine clientname",
81 .type = FIO_OPT_STR_STORE,
82 .help = "Name of the ceph client to access the RBD for the RBD engine",
83 .off1 = offsetof(struct rbd_options, client_name),
84 .category = FIO_OPT_C_ENGINE,
85 .group = FIO_OPT_G_RBD,
91 .help = "Busy poll for completions instead of sleeping",
92 .off1 = offsetof(struct rbd_options, busy_poll),
94 .category = FIO_OPT_C_ENGINE,
95 .group = FIO_OPT_G_RBD,
102 static int _fio_setup_rbd_data(struct thread_data *td,
103 struct rbd_data **rbd_data_ptr)
105 struct rbd_data *rbd;
110 rbd = calloc(1, sizeof(struct rbd_data));
114 /* add for poll, init fd: -1 */
117 rbd->aio_events = calloc(td->o.iodepth, sizeof(struct io_u *));
118 if (!rbd->aio_events)
121 rbd->sort_events = calloc(td->o.iodepth, sizeof(struct io_u *));
122 if (!rbd->sort_events)
131 free(rbd->aio_events);
132 if (rbd->sort_events)
133 free(rbd->sort_events);
140 #ifdef CONFIG_RBD_POLL
141 static bool _fio_rbd_setup_poll(struct rbd_data *rbd)
145 /* add for rbd poll */
146 rbd->fd = eventfd(0, EFD_NONBLOCK);
148 log_err("eventfd failed.\n");
152 r = rbd_set_image_notification(rbd->image, rbd->fd, EVENT_TYPE_EVENTFD);
154 log_err("rbd_set_image_notification failed.\n");
163 static bool _fio_rbd_setup_poll(struct rbd_data *rbd)
169 static int _fio_rbd_connect(struct thread_data *td)
171 struct rbd_data *rbd = td->io_ops_data;
172 struct rbd_options *o = td->eo;
175 if (o->cluster_name) {
176 char *client_name = NULL;
179 * If we specify cluser name, the rados_create2
180 * will not assume 'client.'. name is considered
181 * as a full type.id namestr
183 if (o->client_name) {
184 if (!index(o->client_name, '.')) {
185 client_name = calloc(1, strlen("client.") +
186 strlen(o->client_name) + 1);
187 strcat(client_name, "client.");
188 strcat(client_name, o->client_name);
190 client_name = o->client_name;
194 r = rados_create2(&rbd->cluster, o->cluster_name,
197 if (client_name && !index(o->client_name, '.'))
200 r = rados_create(&rbd->cluster, o->client_name);
203 log_err("rados_create failed.\n");
207 r = rados_conf_read_file(rbd->cluster, NULL);
209 log_err("rados_conf_read_file failed.\n");
213 r = rados_connect(rbd->cluster);
215 log_err("rados_connect failed.\n");
216 goto failed_shutdown;
219 r = rados_ioctx_create(rbd->cluster, o->pool_name, &rbd->io_ctx);
221 log_err("rados_ioctx_create failed.\n");
222 goto failed_shutdown;
225 r = rbd_open(rbd->io_ctx, o->rbd_name, &rbd->image, NULL /*snap */ );
227 log_err("rbd_open failed.\n");
231 if (!_fio_rbd_setup_poll(rbd))
237 rbd_close(rbd->image);
240 rados_ioctx_destroy(rbd->io_ctx);
243 rados_shutdown(rbd->cluster);
249 static void _fio_rbd_disconnect(struct rbd_data *rbd)
260 /* shutdown everything */
262 rbd_close(rbd->image);
267 rados_ioctx_destroy(rbd->io_ctx);
272 rados_shutdown(rbd->cluster);
277 static void _fio_rbd_finish_aiocb(rbd_completion_t comp, void *data)
279 struct fio_rbd_iou *fri = data;
280 struct io_u *io_u = fri->io_u;
284 * Looks like return value is 0 for success, or < 0 for
285 * a specific error. So we have to assume that it can't do
286 * partial completions.
288 ret = rbd_aio_get_return_value(fri->completion);
291 io_u->resid = io_u->xfer_buflen;
295 fri->io_complete = 1;
298 static struct io_u *fio_rbd_event(struct thread_data *td, int event)
300 struct rbd_data *rbd = td->io_ops_data;
302 return rbd->aio_events[event];
305 static inline int fri_check_complete(struct rbd_data *rbd, struct io_u *io_u,
306 unsigned int *events)
308 struct fio_rbd_iou *fri = io_u->engine_data;
310 if (fri->io_complete) {
312 rbd->aio_events[*events] = io_u;
315 rbd_aio_release(fri->completion);
322 static inline int rbd_io_u_seen(struct io_u *io_u)
324 struct fio_rbd_iou *fri = io_u->engine_data;
329 static void rbd_io_u_wait_complete(struct io_u *io_u)
331 struct fio_rbd_iou *fri = io_u->engine_data;
333 rbd_aio_wait_for_complete(fri->completion);
336 static int rbd_io_u_cmp(const void *p1, const void *p2)
338 const struct io_u **a = (const struct io_u **) p1;
339 const struct io_u **b = (const struct io_u **) p2;
342 at = utime_since_now(&(*a)->start_time);
343 bt = utime_since_now(&(*b)->start_time);
353 static int rbd_iter_events(struct thread_data *td, unsigned int *events,
354 unsigned int min_evts, int wait)
356 struct rbd_data *rbd = td->io_ops_data;
357 unsigned int this_events = 0;
361 #ifdef CONFIG_RBD_POLL
364 struct fio_rbd_iou *fri = NULL;
365 rbd_completion_t comps[min_evts];
371 ret = poll(&pfd, 1, -1);
375 assert(pfd.revents & POLLIN);
377 event_num = rbd_poll_io_events(rbd->image, comps, min_evts);
379 for (i = 0; i < event_num; i++) {
380 fri = rbd_aio_get_arg(comps[i]);
383 io_u_qiter(&td->io_u_all, io_u, i) {
385 if (!(io_u->flags & IO_U_F_FLIGHT))
387 if (rbd_io_u_seen(io_u))
390 if (fri_check_complete(rbd, io_u, events))
393 rbd->sort_events[sidx++] = io_u;
400 * Sort events, oldest issue first, then wait on as many as we
401 * need in order of age. If we have enough events, stop waiting,
402 * and just check if any of the older ones are done.
405 qsort(rbd->sort_events, sidx, sizeof(struct io_u *), rbd_io_u_cmp);
407 for (i = 0; i < sidx; i++) {
408 io_u = rbd->sort_events[i];
410 if (fri_check_complete(rbd, io_u, events)) {
416 * Stop waiting when we have enough, but continue checking
417 * all pending IOs if they are complete.
419 if (*events >= min_evts)
422 rbd_io_u_wait_complete(io_u);
424 if (fri_check_complete(rbd, io_u, events))
431 static int fio_rbd_getevents(struct thread_data *td, unsigned int min,
432 unsigned int max, const struct timespec *t)
434 unsigned int this_events, events = 0;
435 struct rbd_options *o = td->eo;
439 this_events = rbd_iter_events(td, &events, min, wait);
455 static int fio_rbd_queue(struct thread_data *td, struct io_u *io_u)
457 struct rbd_data *rbd = td->io_ops_data;
458 struct fio_rbd_iou *fri = io_u->engine_data;
461 fio_ro_check(td, io_u);
464 fri->io_complete = 0;
466 r = rbd_aio_create_completion(fri, _fio_rbd_finish_aiocb,
469 log_err("rbd_aio_create_completion failed.\n");
473 if (io_u->ddir == DDIR_WRITE) {
474 #ifdef CONFIG_RBD_BLKIN
475 blkin_init_trace_info(&fri->info);
476 r = rbd_aio_write_traced(rbd->image, io_u->offset, io_u->xfer_buflen,
477 io_u->xfer_buf, fri->completion, &fri->info);
479 r = rbd_aio_write(rbd->image, io_u->offset, io_u->xfer_buflen,
480 io_u->xfer_buf, fri->completion);
483 log_err("rbd_aio_write failed.\n");
487 } else if (io_u->ddir == DDIR_READ) {
488 #ifdef CONFIG_RBD_BLKIN
489 blkin_init_trace_info(&fri->info);
490 r = rbd_aio_read_traced(rbd->image, io_u->offset, io_u->xfer_buflen,
491 io_u->xfer_buf, fri->completion, &fri->info);
493 r = rbd_aio_read(rbd->image, io_u->offset, io_u->xfer_buflen,
494 io_u->xfer_buf, fri->completion);
498 log_err("rbd_aio_read failed.\n");
501 } else if (io_u->ddir == DDIR_TRIM) {
502 r = rbd_aio_discard(rbd->image, io_u->offset,
503 io_u->xfer_buflen, fri->completion);
505 log_err("rbd_aio_discard failed.\n");
508 } else if (io_u->ddir == DDIR_SYNC) {
509 r = rbd_aio_flush(rbd->image, fri->completion);
511 log_err("rbd_flush failed.\n");
515 dprint(FD_IO, "%s: Warning: unhandled ddir: %d\n", __func__,
522 rbd_aio_release(fri->completion);
525 td_verror(td, io_u->error, "xfer");
526 return FIO_Q_COMPLETED;
529 static int fio_rbd_init(struct thread_data *td)
533 r = _fio_rbd_connect(td);
535 log_err("fio_rbd_connect failed, return code: %d .\n", r);
545 static void fio_rbd_cleanup(struct thread_data *td)
547 struct rbd_data *rbd = td->io_ops_data;
550 _fio_rbd_disconnect(rbd);
551 free(rbd->aio_events);
552 free(rbd->sort_events);
557 static int fio_rbd_setup(struct thread_data *td)
559 rbd_image_info_t info;
561 struct rbd_data *rbd = NULL;
562 int major, minor, extra;
565 /* log version of librbd. No cluster connection required. */
566 rbd_version(&major, &minor, &extra);
567 log_info("rbd engine: RBD version: %d.%d.%d\n", major, minor, extra);
569 /* allocate engine specific structure to deal with librbd. */
570 r = _fio_setup_rbd_data(td, &rbd);
572 log_err("fio_setup_rbd_data failed.\n");
575 td->io_ops_data = rbd;
577 /* librbd does not allow us to run first in the main thread and later
578 * in a fork child. It needs to be the same process context all the
581 td->o.use_thread = 1;
583 /* connect in the main thread to determine to determine
584 * the size of the given RADOS block device. And disconnect
587 r = _fio_rbd_connect(td);
589 log_err("fio_rbd_connect failed.\n");
593 /* get size of the RADOS block device */
594 r = rbd_stat(rbd->image, &info, sizeof(info));
596 log_err("rbd_status failed.\n");
598 } else if (info.size == 0) {
599 log_err("image size should be larger than zero.\n");
604 dprint(FD_IO, "rbd-engine: image size: %lu\n", info.size);
606 /* taken from "net" engine. Pretend we deal with files,
607 * even if we do not have any ideas about files.
608 * The size of the RBD is set instead of a artificial file.
610 if (!td->files_index) {
611 add_file(td, td->o.filename ? : "rbd", 0, 0);
612 td->o.nr_files = td->o.nr_files ? : 1;
616 f->real_file_size = info.size;
618 /* disconnect, then we were only connected to determine
619 * the size of the RBD.
621 _fio_rbd_disconnect(rbd);
625 _fio_rbd_disconnect(rbd);
631 static int fio_rbd_open(struct thread_data *td, struct fio_file *f)
636 static int fio_rbd_invalidate(struct thread_data *td, struct fio_file *f)
638 #if defined(CONFIG_RBD_INVAL)
639 struct rbd_data *rbd = td->io_ops_data;
641 return rbd_invalidate_cache(rbd->image);
647 static void fio_rbd_io_u_free(struct thread_data *td, struct io_u *io_u)
649 struct fio_rbd_iou *fri = io_u->engine_data;
652 io_u->engine_data = NULL;
657 static int fio_rbd_io_u_init(struct thread_data *td, struct io_u *io_u)
659 struct fio_rbd_iou *fri;
661 fri = calloc(1, sizeof(*fri));
663 io_u->engine_data = fri;
667 static struct ioengine_ops ioengine = {
669 .version = FIO_IOOPS_VERSION,
670 .setup = fio_rbd_setup,
671 .init = fio_rbd_init,
672 .queue = fio_rbd_queue,
673 .getevents = fio_rbd_getevents,
674 .event = fio_rbd_event,
675 .cleanup = fio_rbd_cleanup,
676 .open_file = fio_rbd_open,
677 .invalidate = fio_rbd_invalidate,
679 .io_u_init = fio_rbd_io_u_init,
680 .io_u_free = fio_rbd_io_u_free,
681 .option_struct_size = sizeof(struct rbd_options),
684 static void fio_init fio_rbd_register(void)
686 register_ioengine(&ioengine);
689 static void fio_exit fio_rbd_unregister(void)
691 unregister_ioengine(&ioengine);